Mathematical

Finance with
Applications
Edited by
Wing-Keung Wong, Xu Guo and Sergio Ortobelli Lozza
Printed Edition of the Special Issue Published in
Journal of Risk and Financial Management

[Link]/journal/jrfm
Mathematical Finance with
Applications
Mathematical Finance with
Applications

Editors
Wing-Keung Wong
Xu Guo
Sergio Ortobelli Lozza

MDPI • Basel • Beijing • Wuhan • Barcelona • Belgrade • Manchester • Tokyo • Cluj • Tianjin
Editors
Wing-Keung Wong Xu Guo Sergio Ortobelli Lozza
Asia University Beijing Normal University University of Bergamo
Taiwan China Italy

Editorial Office
MDPI
St. Alban-Anlage 66
4052 Basel, Switzerland

This is a reprint of articles from the Special Issue published online in the open access journal
Journal of Risk and Financial Management (ISSN 1911-8074) (available at: [Link]
journal/jrfm/special issues/mathematical finance applications).

For citation purposes, cite each article independently as indicated on the article page online and as
indicated below:

LastName, A.A.; LastName, B.B.; LastName, C.C. Article Title. Journal Name Year, Article Number,
Page Range.

ISBN 978-3-03943-573-9 (Hbk)

ISBN 978-3-03943-574-6 (PDF)


c 2020 by the authors. Articles in this book are Open Access and distributed under the Creative
Commons Attribution (CC BY) license, which allows users to download, copy and build upon
published articles, as long as the author and publisher are properly credited, which ensures maximum
dissemination and a wider impact of our publications.
The book as a whole is distributed by MDPI under the terms and conditions of the Creative Commons
license CC BY-NC-ND.
Contents

About the Editors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Wing-Keung Wong
Editorial Statement for Mathematical Finance
Reprinted from: J. Risk Financial Manag. 2020, 13, 18, doi:10.3390/jrfm13020018 . . . . . . . . . . 1

Imran Yousaf, Shoaib Ali and Wing-Keung Wong

An Empirical Analysis of the Volatility Spillover Effect between World-Leading and the Asian
Stock Markets: Implications for Portfolio Management
Reprinted from: J. Risk Financial Manag. 2020, 13, 226, doi:10.3390/jrfm13100226 . . . . . . . . . . 5

Mohammad Rafiqul Islam and Nguyet Nguyen

Comparison of Financial Models for Stock Price Prediction
Reprinted from: J. Risk Financial Manag. 2020, 13, 181, doi:10.3390/jrfm13080181 . . . . . . . . . . 33

Imran Yousaf, Shoaib Ali and Wing-Keung Wong

Return and Volatility Transmission between World-Leading and Latin American Stock Markets:
Portfolio Implications
Reprinted from: J. Risk Financial Manag. 2020, 13, 148, doi:10.3390/jrfm13070148 . . . . . . . . . . 53

Haim Levy
The Investment Home Bias with Peer Effect
Reprinted from: J. Risk Financial Manag. 2020, 13, 94, doi:10.3390/jrfm13050094 . . . . . . . . . . 73

Zhe Li
Equity Option Pricing with Systematic and Idiosyncratic Volatility and Jump Risks
Reprinted from: J. Risk Financial Manag. 2020, 13, 16, doi:10.3390/jrfm13010016 . . . . . . . . . . 93

Alex Golodnikov, Viktor Kuzmenko and Stan Uryasev

CVaR Regression Based on the Relation between CVaR and Mixed-Quantile Quadrangles
Reprinted from: J. Risk Financial Manag. 2019, 12, 107, doi:10.3390/jrfm12030107 . . . . . . . . . . 111

Sel Ly, Kim-Hung Pho, Sal Ly and Wing-Keung Wong

Determining Distribution for the Quotients of Dependent and Independent Random Variables
by Using Copulas
Reprinted from: J. Risk Financial Manag. 2019, 12, 42, doi:10.3390/jrfm12010042 . . . . . . . . . . 133

László Nagy and Mihály Ormos

Friendship of Stock Market Indices: A Cluster-Based Investigation of Stock Markets
Reprinted from: J. Risk Financial Manag. 2018, 11, 88, doi:10.3390/jrfm11040088 . . . . . . . . . . 161

Rafiuddin Ahmed and Rafiqul Bhuyan

Capital Structure and Firm Performance in Australian Service Sector Firms: A Panel
Data Analysis
Reprinted from: J. Risk Financial Manag. 2020, 13, 214, doi:10.3390/jrfm13090214 . . . . . . . . . . 177

Jian Huang and Huazhang Liu

Examination and Modification of Multi-Factor Model in Explaining Stock Excess Return with
Hybrid Approach in Empirical Study of Chinese Stock Market
Reprinted from: J. Risk Financial Manag. 2019, 12, 91, doi:10.3390/jrfm12020091 . . . . . . . . . . 193

v
About the Editors
Wing-Keung Wong obtained his Ph.D. from the University of Wisconsin-Madison, the USA
with a major in Business Statistics (Statistics and Finance) and obtained his Bachelor degree from
the Chinese University of Hong Kong, Hong Kong, with a major in Mathematics and a double minor
in Economics and Statistics. Currently, he is a Chair Professor at the Department of Finance, Asia
University. He was a Full Professor at the Department of Economics, Hong Kong Baptist University,
and Deputy Director at Risk Management Institute, National University of Singapore.
Professor WONG appears in “Who’s Who in the World” and gets Albert Nelson Marquis
Lifetime Achievement Award. 2017, Marquis Who’s Who. His Erdos number is 3. He is ranked top
1% by Social Science Research Network and in the list of top Taiwan economists and Asian economists
and top economists by RePEc. He has published more than three hundred papers including papers
published in some top journals. He has more than 9400 citations in Google Scholar, more than 6900
citations in Researchgate, and more than 3200 citations in Mendeley. His h-index is 54, (40 since 2015)
and i10-index is 205, (182 since 2015) by Google Scholar citation.
He has been serving international academies, Government, society, and universities, providing
consultancy to several Government departments and corporations, and giving lectures and seminars
to several universities. For example, he has been serving as editor, guest leading editor, advisor,
associate editor for some international journals, appointed as an advisor/member of various
international associations/institutes, serving as a referee for many journals/conferences, supervising
solely or jointly several overseas graduate students, appointed as an external reviewer and external
examiner by other universities, and invited by many universities/institutions to present papers or
conduct seminars.

Xu Guo is an Associate Professor in School of Statistics, Beijing Normal University. He was

an Assistant Research Professor in the Department of Statistics, Pennsylvania State University from
Sep. 2018 to Feb. 2020. He got his Ph.D. degree from the Department of Mathematics, Hong
Kong Baptist University in 2014. His research interests are model specification testing, missing
data, semiparametric regression analysis, risk management and decision making under risk and
uncertainty. Up to now, he has published more than 30 papers in numerous international journals,
such as Journal of the Royal Statistical Society: Series B, Biometrika, Journal of Multivariate Analysis,
Computational Statistics & Data Analysis, Insurance: Mathematics and Economics, Economics
Letters, Economic Modelling, North American Journal of Economics and Finance, etc.

Sergio Ortobelli Lozza is a full professor in Mathematical Finance at the University of Bergamo.
He holds a Ph.D. in Computational Methods for Financial and Economic Forecasting and Decisions
from the University of Bergamo. His research, published in various academic journals in mathematics
and finance, focuses on the application of probability theory and operational research to portfolio
theory, risk management, and option theory. He taught numerous courses at the Universities of
Bergamo, Calabria and Milan, including basic and advanced calculus, measure theory, stochastic
processes, portfolio theory, and advanced mathematical finance. From 1999 till 2020 he wrote
more than 170 refereed scientific works together with several well known Italian and international
professors. Some of these papers have been well recognized by the scientific community.

vii
 Journal of
Risk and Financial
Management

Editorial
Editorial Statement for Mathematical Finance
Wing-Keung Wong 1,2,3
1 Department of Finance, Fintech Center, and Big Data Research Center, Asia University, Wufeng,
Taichung 41354, Taiwan; wong@[Link]
2 Department of Medical Research, China Medical University Hospital, Taichung 40402, Taiwan
3 Department of Economics and Finance, The Hang Seng University of Hong Kong, Hong Kong, China

Received: 7 January 2020; Accepted: 8 January 2020; Published: 21 January 2020

Abstract: Mathematics plays a vital role in many areas of finance and provides the theories and tools
that have been widely used in all areas of finance. In this editorial, we tell authors the ideas on what
types of papers we will accept for publication in the area of mathematical finance. We will discuss
some well-cited papers of mathematical finance.

Keywords: mathematics; probability; statistics; finance; applications

Mathematics plays a vital role in many areas of finance. In particular, it provides the theories
and tools that have been widely used in all areas of finance. Knowledge of mathematics, probability,
statistics, and other analytic approaches is essential to develop methods and theories in finance and
test their validity through the analysis of empirical real-world data. For example, mathematics,
probability, and statistics could help to develop pricing models for financial assets such as equities,
bonds, currencies, and derivative securities, and propose financially optimal strategies coherently to
decision-makers according to their preferences. This section will bring together theory, practice, and
applications of mathematical finance. We discuss some of the most cited papers, as follows:
Ly et al. (2019) develop the theory on both density and distribution functions for the quotient
Y = X1/ X2 and the ratio of one variable over the sum of two variables Z = X1/ (X1 + X2 ) of two dependent
or independent random variables X1 and X2 by using copulas to capture the structures between X1
and X2 , and extend the theory by establishing the density and distribution functions for the quotients
Y = X1/ X2 and Z = X1/ (X1 + X2 ) of two dependent normal random variables X1 and X2 in the case
of Gaussian copulas. Thereafter, they develop the theory on the median for the ratios of both Y and
Z on two normal random variables X1 and X2 and extend the result of the median for Z to a larger
family of symmetric distributions and symmetric copulas of X1 and X2 . In addition, they introduce
the Monte Carlo algorithm, numerical analysis, and graphical approach to efficiently compute the
complicated integrals and study the behaviors of density and distribution and illustrate their proposed
approaches by using a simulation study with ratios of normal random variables on several different
copulas, including Gaussian, Student-t, Clayton, Gumbel, Frank, and Joe Copulas, and discuss the
behaviors via all copulas above with the same Kendall’s coefficient. They find that copulas make big
impacts from different copulas on the behavior of distributions, especially on median, spread, scale,
and skewness effects.
Golodnikov et al. (2019) show that CVaR linear regression can be reduced to minimize the
Rockafellar error function with linear programming. They establish the theoretical basis for the analysis
with the quadrangle theory of risk functions and derive relationships between elements of CVaR
quadrangle and mixed-quantile quadrangle for discrete distributions with equally probable atoms.
They present two equivalent variants of discretization of the integral, which resulted in two sets of
parameters for the mixed-quantile quadrangle. For the first set of parameters, the minimization of
error from the CVaR quadrangle is equivalent to the minimization of the Rockafellar error from the
mixed-quantile quadrangle. Alternatively, a two-stage procedure based on the decomposition theorem

JRFM 2020, 13, 18; doi:10.3390/jrfm13020018 1 [Link]/journal/jrfm

JRFM 2020, 13, 18

can be used for CVaR linear regression with both sets of parameters. They find that this procedure is
valid because the deviation in the mixed-quantile quadrangle (called mixed CVaR deviation) coincides
with the deviation in the CVaR quadrangle for both sets of parameters. In addition, they illustrate
theoretical results with a case study demonstrating the numerical efficiency of the suggested approach.
De Gaetano (2018) investigates the relevance of structural breaks for forecasting the volatility of
daily returns on BRICS countries by using the data from 19 July 1999 to 16 July 2015 to identify structural
breaks in the unconditional variance, a binary segmentation algorithm with a test. He introduces
some forecast combinations that account for the identified structural breaks and evaluate and compare
their performance by using the model confidence set (MCS). He obtains significant evidence of the
relevance of the structural breaks, in particular, in the regimes identified by the structural breaks;
a substantial change in the unconditional variance is quite evident. In addition, He finds that the
combination that averages forecasts obtained using different rolling estimation windows outperforms
all the other combinations.
Van Dijk et al. (2018) propose improved regression models to estimate calibrated parameters
(including the market variables in a real-world simulation), predict out-of-sample implied volatility
surfaces, and evaluate the impact on the solvency capital requirement for different points in time.
Korkmaz et al. (2018) introduce and study a new three-parameter Pareto distribution. They discuss
various mathematical and statistical properties of the new model to perform some estimation methods
of the model parameters, use the peaks-over-threshold method to estimate value-at-risk (VaR) by
means of the proposed distribution, and compare the distribution with a few other models to show its
versatility in modeling data with heavy tails. In addition, they present VaR estimation with the Burr ×
Pareto distribution by using time series data and consider the new model as an alternative VaR model
against the generalized Pareto model for financial institutions.
The popular replication formula to price variance swaps assumes continuity of traded option
strikes. In practice, however, there is only a discrete set of option strikes traded on the market. Le Floc’h
(2018) presents different discrete replication strategies and explains why the continuous replication
price is more relevant.
Ghitany et al. (2018) propose an alternative generalization of the Pareto distribution, study its
properties, and apply their proposed model to analyze earthquake insurance data.
Nagy and Ormos (2018) introduce a spectral clustering-based method to show that stock prices
contain not only firm but also network-level information. Clustering different stock indices and
reconstructing the equity index graph from historical daily closing prices, they show that tail events
have a minor effect on the equity index structure. In addition, they find that covariance and Shannon
entropy do not provide enough information about the network, but Gaussian clusters can explain a
substantial part of the total variance.
Employing a time-varying vector autoregression with stochastic volatility, Feldkircher and Huber
(2018) compare the transmission of a conventional monetary policy shock with that of an unexpected
decrease in the term spread, which mirrors quantitative easing. They find that the spread shock works
mainly through a boost to consumer wealth growth, while a conventional monetary policy shock
affects real output growth via a broad credit/bank lending channel. In addition, they find small output
effects of a conventional monetary policy shock during the period of the global financial crisis and
stronger effects in its aftermath. Their findings imply that when the central bank has left the policy rate
unaltered for an extended period of time, a policy surprise might boost output particularly strongly
while the spread shock has affected output growth most strongly during the period of the global
financial crisis, and less so thereafter.

Funding: This research received no external funding.

Acknowledgments: The author would like to thank Robert B. Miller and Howard E. Thompson for their continuous
guidance and encouragement. For financial and research support, the author acknowledges Asia University,
China Medical University Hospital, The Hang Seng University of Hong Kong, the Research Grants Council

2
JRFM 2020, 13, 18

of Hong Kong (project number 12500915), and Ministry of Science and Technology (MOST, Project Numbers
106-2410-H-468-002 and 107-2410-H-468-002-MY3), Taiwan.
Conflicts of Interest: The author declares no conflict of interest.

References
De Gaetano, Davide. 2018. Forecast Combinations for Structural Breaks in Volatility: Evidence from BRICS
Countries. Journal Risk Financial Management 11: 64. [CrossRef]
Feldkircher, Martin, and Florian Huber. 2018. Unconventional U.S. Monetary Policy: New Tools, Same Channels?
Journal Risk Financial Management 11: 71. [CrossRef]
Ghitany, Mohamed E., Emilio Gómez-Déniz, and Saralees Nadarajah. 2018. A New Generalization of the Pareto
Distribution and Its Application to Insurance Data. Journal Risk Financial Management 11: 10. [CrossRef]
Golodnikov, Alex, Viktor Kuzmenko, and Stan Uryasev. 2019. CVaR Regression Based on the Relation between
CVaR and Mixed-Quantile Quadrangles. Journal Risk Financial Management 12: 107. [CrossRef]
Korkmaz, Mustafa Ç, Emrah Altun, Haitham M. Yousof, Ahmed Z. Afify, and Saralees Nadarajah. 2018. The Burr
X Pareto Distribution: Properties, Applications and VaR Estimation. Journal Risk Financial Management 11: 1.
[CrossRef]
Le Floc’h, Fabien. 2018. Variance Swap Replication: Discrete or Continuous? Journal Risk Financial Management 11: 11.
[CrossRef]
Ly, Sel, Kim-Hung Pho, Sal Ly, and Wing-Keung Wong. 2019. Determining Distribution for the Quotients of
Dependent and Independent Random Variables by Using Copulas. Journal Risk Financial Management 12: 42.
[CrossRef]
Nagy, László, and Mihály Ormos. 2018. Friendship of Stock Market Indices: A Cluster-Based Investigation of
Stock Markets. Journal Risk Financial Management 11: 88. [CrossRef]
Van Dijk, Marcel T. P., Cornelis S. L. De Graaf, and Cornelis W. Oosterlee. 2018. Between λ and μ : The λμ Measure
for Pricing in Asset Liability Management. Journal Risk Financial Management 11: 67. [CrossRef]

© 2020 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

3
 Journal of
Risk and Financial
Management

Article
An Empirical Analysis of the Volatility Spillover
Effect between World-Leading and the Asian Stock
Markets: Implications for Portfolio Management
Imran Yousaf 1, *, Shoaib Ali 1 and Wing-Keung Wong 2,3,4
1 Air University School of Management, Air University, Islamabad 44000, Pakistan; ShoaibAli@[Link]
2 Department of Finance, Fintech Center, and Big Data Research Center, Asia University, Taichung 41354,
Taiwan; wong@[Link]
3 Department of Medical Research, China Medical University Hospital, Taichung 40402, Taiwan
4 Department of Economics and Finance, The Hang Seng University of Hong Kong, Hong Kong 999077, China
* Correspondence: imranyousaf.fin@[Link]

Received: 14 August 2020; Accepted: 23 September 2020; Published: 25 September 2020

Abstract: This study employs the Vector Autoregressive-Generalized Autoregressive Conditional

Heteroskedasticity (VAR-AGARCH) model to examine both return and volatility spillovers from the
USA (developed) and China (Emerging) towards eight emerging Asian stock markets during the
full sample period, the US financial crisis, and the Chinese Stock market crash. We also calculate the
optimal weights and hedge ratios for the stock portfolios. Our results reveal that both return and
volatility transmissions vary across the pairs of stock markets and the financial crises. More specifically,
return spillover was observed from the US and China to the Asian stock markets during the US
financial crisis and the Chinese stock market crash, and the volatility was transmitted from the USA to
the majority of the Asian stock markets during the Chinese stock market crash. Additionally, volatility
was transmitted from China to the majority of the Asian stock markets during the US financial crisis.
The weights of American stocks in the Asia-US portfolios were found to be higher during the Chinese
stock market crash than in the US financial crisis. For the majority of the Asia-China portfolios,
the optimal weights of the Chinese stocks were almost equal during the Chinese stock market crash
and the US financial crisis. Regarding hedge ratios, fewer US stocks were required to minimize the risk
for Asian stock investors during the US financial crisis. In contrast, fewer Chinese stocks were needed
to minimize the risk for Asian stock investors during the Chinese stock market crash. This study
provides useful information to institutional investors, portfolio managers, and policymakers regarding
optimal asset allocation and risk management.

Keywords: return spillover; volatility spillover; shock spillover; US financial crisis; Chinese stock
market crash

JEL Classification: G10; G11; G12; G15

1. Introduction
Information transmissions from both return and volatility across national equity markets are of
greater interest to both investors and policymakers, with increasing financial integration in the stock
markets all over the world (Yousaf et al. 2020). If, for example, asset volatility is transmitted from one
market to another during turmoil or crisis period (Forbes and Rigobon 2002; Diebold and Yilmaz 2009),
then portfolio managers need to adjust their asset allocations (Baele 2005; Engle et al. 2012) and financial
policymakers need to adapt their policies in order to mitigate the contagion risk. Changes in linkages
between national equity markets, especially during a crisis, can also have important implications for
asset allocations, business valuation, risk management, and access to finance.

JRFM 2020, 13, 226; doi:10.3390/jrfm13100226 5 [Link]/journal/jrfm

JRFM 2020, 13, 226

Several studies have examined linkages between the equity markets during the 1997 Asian financial
crisis (In Francis et al. 2001; Wan and Wong 2001; Yang et al. 2003), and the last 2008 global financial crisis
(Yilmaz 2010; Cheung et al. 2007; Kim et al. 2015; Li and Giles 2015; Lean et al. 2015; Vieito et al. 2015;
Zhu et al. 2019) and some studies, see, for example, Fung et al. (2011) and Guo et al. (2017), develop
theories to explain that crisis. However, the linkages between equity markets during the Chinese stock
market crash of 2015have been rarely examined. The Chinese stock market experienced a major crash
in 2015 (Zhu et al. 2017; Yousaf and Hassan 2019; Yousaf et al. 2020; Yousaf and Ali 2020). The CSI 300
index increased before reaching 5178 points in mid-June of 2015. Then, it took a roller-coaster ride and
dropped by up to 34% in just 20 days; Chinese stock market also lost 1000 points within just one week.
Around 50% of Chinese stocks lost more than half of their pre-crash market value. The Chinese stock
market crash affected many other commodities and financial markets, including Asian (Allen 2015)
and the US stock markets (The causes and consequences of China’s market crash 2015).
Despite the importance of the Chinese crash for international portfolio managers, few studies have
examined how it was transmitted to other national financial markets. Xiong et al. (2018) investigate
the time-varying correlation between economic policy uncertainty and Chinese stock market returns
during the Chinese crash of 2015, while Yousaf and Hassan (2019) examine the linkages between
crude oil and emerging Asian stock markets during this crisis. However, research on the linkages
between stock markets has not been investigated yet for the 2015 Chinese crash. Therefore, this
study focuses on providing useful insights about this issue for the Asian region, which has attracted
considerable attention from finance practitioners and academics due to its position as the center of
global economic activity in the 21st century1 . While using the US and Chinese equity markets as the
indicators of global markets, we explore whether global investors can get the maximum benefit of
diversification by adding emerging Asian market stocks in their portfolios. In literature, several studies
have examined the linkages between the global (US and China) and emerging Asian equity markets
during the Asian financial crisis, and the US financial crisis (Yang et al. 2003; Beirne et al. 2013; Jin 2015;
Li and Giles 2015), but not in the Chinese stock market crash.
We address the above-mentioned literature gap by examining the return and volatility spillover
from the US and China to the emerging Asian equity markets during the Chinese stock market crash
by using the VAR-AGARCH model that was developed by Ling and McAleer (2003). Moreover,
we examine the ability of spillovers during the full sample period and the 2008 US financial crisis to
provide comparative insights to investors about whether the impact of the Chinese crash on equity
market spillovers was different from those in the other sample periods. Our findings show that return
spillover was observed from the US and China to the Asian stock market during the US financial crisis
and the Chinese stock market crash. Volatility was also transmitted from the US to the majority of the
Asian stock markets during the Chinese stock market crash. However, volatility was transmitted from
China to the majority of the Asian stock markets during the US financial crisis. Overall, as the return
and volatility transmission vary across pairs of stock markets and financial crises, investors have to
adjust their asset allocations from time to time to improve their profits. Therefore, we also estimate the
optimal weights and hedge ratios during the full sample period, the US financial crisis, and the Chinese
stock market crash. Our findings imply that fewer US stocks were required to minimize the risk for
Asian stock investors during the US financial crisis compared to during the Chinese crash. In contrast,
fewer Chinese stocks were needed to minimize the risk for Asian stock investors during the Chinese
stock market crash as compared to during the US crisis. Overall, our findings draw several important
implications for risk management and portfolio diversification that could be useful for investors and
policymakers related to the US and Asian stock markets.

1 Source: [Link]

6
JRFM 2020, 13, 226

The rest of the paper is organized as follows: Section 2 provides the literature review. Section 3
describes the data and methodology. Section 4 reports the findings, and Section 5 concludes the
whole discussion.

2. Literature Review
The analysis of both return and volatility spillover between stock markets is crucial for investors
in designing optimal portfolios. According to modern portfolio theory, the gains of international
portfolio diversification decrease when the correlation of security returns increases and vice versa.
Michaud et al. (1996) discuss the advantages of a low correlation between the developed and emerging
markets for international portfolio diversification. Due to this trend, investors can benefit by investing
in emerging markets that are weakly interconnected with developed markets. However, this correlation
becomes higher during an economic crisis, suggesting low diversification benefits when diversification
is most required.

2.1. Linkages between US, China, and Asian Stock Markets

Many studies have been conducted to investigate the link between different stock markets during
the last three decades. Liu and Pan (1997) examine the mean and the volatility spillover from the US
and Japan to Singapore, Hong Kong, Thailand, and Taiwan. The results show that the US market is
more dominant than the Japanese stock market in transmitting return and volatility effects to four
Asian stock markets. Huang et al. (2000) investigate the link between the US, Japan, and South China
growth triangle. The US stock market significantly and dominantly affects the south Chinese growth
triangle compared to the impact of Japan on China’s stock market. The return spillover has been also
found to be significant from the US to Hong Kong and Taiwan, and from Hong Kong to the Taiwanese
stock market. Miyakoshi (2003) estimates the return and volatility spillover between the US, Japan,
and seven Asian stock markets (South Korea, Taiwan, Singapore, Thailand, Indonesia, and Hong
Kong). It finds a significant return spillover from the US to Asian markets, whereas no return spillover
is found from Japan to Asian stock markets. Moreover, the volatility spillover from Japan to other
Asian stock markets is observed to be dominant as compared to the volatility spillover from the US to
Asian stock markets.
Johansson and Ljungwall (2009) examine the association between stock markets of China,
Hong Kong, and Thailand. It reports a significant return spillover from Taiwan to China and the Hong
Kong stock market. In contrast, volatility spillover runs from Hong Kong to Taiwan and from Taiwan to
the Chinese stock market. Zhou et al. (2012) estimate the spillover between Chinese and international
(the US, the UK, France, Germany, Japan, India, Hong Kong, Taiwan, South Korea, and Singapore)
stock markets from 1996 to 2009. Before 2005, the Chinese stock market was affected by spillover
from other international markets. After 2005, volatility spillover was significantly transmitted from
China to most of the other international stock markets. Chien et al. (2015) report on the significant
financial integration between China and the ASEAN-5 (Indonesia, Malaysia, the Philippines, Singapore,
and Thailand) stock markets. Huo and Ahmed (2017) provide significant evidence of both return and
volatility effects from China to the Hong Kong equity market.

2.2. Linkages between US, China, and Asian Stock Markets during Crisis
Many studies have examined the linkages between markets during crisis periods. Yang et al. (2003)
investigate the short and long-run relationship between the US, Japan, and ten Asian stock markets,
mainly focusing on the Asian financial crisis of 1997–1998. This study reports a strengthened long-run
co-integration among these stock markets during the Asian financial crises. The degree of integration is
found to change during crises and non-crisis periods. Beirne et al. (2013) look at the volatility spillover
from developed to emerging stock markets during periods of turbulence in mature stock markets.
It finds that volatility in mature markets affects the conditional variances in emerging stock markets.

7
JRFM 2020, 13, 226

Moreover, the spillover effect from developed to emerging markets is also changed during times of
turbulence in mature markets.
Jin (2015) examines the mean and volatility spillover between China, Taiwan, and Hong Kong.
It reveals that financial crises have a substantial and positive effect on expected conditional variances,
but also that the size and dynamics of impacts vary from market to market. Li and Giles (2015)
investigate the volatility spillover across the US, Japan, and four Asian developing economies during
the Asian financial crisis of 1997 and the US financial crises of 2008. The results revealed that there
is a presence of a volatility spillover effect from the USA to Asian developing economies and Japan.
This study also finds a bidirectional volatility spillover between US and Asian markets that occurred
during the Asian financial crisis. Gkillas et al. (2019) explore integration and co-movement between
68 international stock markets (including in the Asian region) during the US financial crisis.
Overall, several studies have examined the return and volatility spillover from the US to Asian
markets during the Asian financial crisis of 1997 and the US financial crisis of 2008. However less has
been done on both return and volatility transmission from China to the emerging Asian stock markets
during the US financial crisis and the Chinese stock market crash. Moreover, no study has examined
return and volatility spillovers from the US to the emerging Asian stock markets during the Chinese
crash. Therefore, this study addresses these above-mentioned literature gaps.

3. Data and Methodology

3.1. Data
We based our empirical investigation on daily data of accepted benchmark stock indices of nine
Asian countries and the US. The Emerging Asian stock markets include China, India, South Korea,
Indonesia, Pakistan, Malaysia, the Philippines, Thailand, and Taiwan. The emerging Asian economies
were selected from the list of countries, including the MSCI (Morgan Stanley Capital International)
emerging market index. The data of stock indices were taken from the Data Stream database. The index
is assumed to be the same on non-trading days (holidays except weekends) as on the previous trading
day, as suggested by Malik and Hammoudeh (2007) and many others.2
This study used the full sample period from 1 January 2000 to 30 June 2018 and studies the
following two sub-samples: the first sub-period from 1 August 2007 to 31 July 2010 presenting the
period with the US financial crisis; and the second sub-period from 1 June 2015 to 30 May 2018
presenting the period with the Chinese Stock market crash. We note that Yousaf and Hassan (2019)
also use similar time frames for the US financial crisis and the Chinese stock market crash. This study
followed He (2001) and many others to use three-year data for each crisis for short-run analysis.
Changes in market correlations take place continuously not only as a result of crises but also due to the
consequences of many financial, economic, and political events. Moreover, Arouri et al. (2015) have
also used the daily data covering periods shorter than three years to estimate the return and volatility
spillover between gold and Chinese stock markets in US financial crisis by applying the VAR-GARCH
model. The difference in the opening time of US and Asian stock markets was adjusted by using lags
where necessary.

2 In time-series data, if there are missing values, there are two ways to deal with the incomplete data: (a) omit the entire
record that contains information, (b) Impute the missing information. We used 10 series in this paper and if we wanted to
omit the missing data for one series then the data of all other nine series needed to be removed as well for that specific day.
So, if we omitted the data for days where values are missing at specific days, then we lost the data for many days, which is
not good for getting realistic results. Therefore, we followed many studies, for example, Malik and Hammoudeh (2007),
and imputed the missing data by using previous day data. Indeed, there are many methods used to impute the missing data
and every method has pros and cons, but we used this imputation method following past literature. Moreover, our missing
observations were less than one percent of overall data, therefore the imputation method should not create a larger effect
than that on results.

8
JRFM 2020, 13, 226

3.2. Methodology
This study estimated the return and volatility transmissions using the Vector Autoregressive-
Generalized Autoregressive Conditional Heteroskedasticity (VAR-AGARCH) model proposed by
McAleer et al. (2009). Several studies have previously used the VAR-GARCH and VAR-AGARCH
model to estimate spillover between different asset classes (Arouri et al. 2011; Arouri et al. 2012;
Jouini 2013; Yousaf and Hassan 2019). This model includes the Constant Conditional Correlation
(CCC-GARCH) model of Bollerslev (1990) as a special case. The selection of the model was based
on three reasons. First, the most commonly used multivariate models are the BEKK (Baba, Engle,
Kraft, and Kroner) model and the DCC (dynamic conditional correlation) model. These models
often suffer from unreasonable parameter estimates and data convergence problems (Bouri 2015).
The VAR-AGARCH model overcomes these problems regarding parameters and data convergence.
Second, it incorporates asymmetry into the model. Third, this model can be used to calculate the
optimal weights and hedge ratios.
Ling and McAleer (2003) propose the multivariate VAR-GARCH Model to estimate the return
and volatility transmission between the different series. For two series, the VAR-GARCH model has
the following specifications for the conditional mean equation3 :

Rt = μ + FRt−1 + et with et = D1/2

t ηt , (1)

in which Rt represents a 2 × 1 vector of daily returns4 on the stocks x and y at time t, μ denotes a 2 × 1
vector of constants, F is a 2 × 2 matrix of parameters measuring the impacts of own lagged and cross
mean transmissions between two series, et is the residual of the mean equation for the two stocks
returns series at time t, ηt indicates a 2 × 1 vector of independently and identically distributed random
 
y y
vectors, and D1/2
t = diag ( hxt , ht ), where hxt and ht representing the conditional variances of the
returns for stocks x and y, respectively, are given as
 2  y 2 y
hxt = Cx + a11 ext−1 + a21 et−1 + b11 hxt−1 + b21 ht−1 , (2)

y
 2  y 2 y
ht = C y + a12 ext−1 + a22 et−1 + b12 hxt−1 + b22 ht−1 . (3)

Equations (2) and (3) reveal how shock and volatility are transmitted over time and across the
markets under investigation. Furthermore, the conditional covariance between returns from two
different stock markets can be estimated as follows:
 
x,y y
ht = p × hxt × ht . (4)

x,y
In the above equation, ht refers to the conditional covariance between the returns of two stock
markets (x, y) at time t. Moreover, p indicates the constant conditional correlation between the returns
of two stock markets (x, y).
The VAR−GARCH model assumes that positive or negative shocks have the same impact on
the conditional variance. To estimate the spillover between different markets, we estimated spillover
between two stock markets by using the VAR−AGARCH Model proposed by the McAleer et al. (2009).

3 Several studies, for example, Hammoudeh et al. (2009), Arouri et al. (2011), and Dutta et al. (2018) have applied the VAR for
the conditional mean equation. 
 Stock Indext
4 Stock Returnst = ln Stock Index .
t−1

9
JRFM 2020, 13, 226

The VAR AGARCH model incorporates asymmetry in the model as well. Specifically, instead of using
Equations (2) and (3), the conditional variance of the VAR AGARCH model was defined as follows:
 2  y 2 y
    
hxt = Cx + a11 A ext−1 + a21 A et−1 + b11 hxt−1 + b21 ht−1 +a11 B ext−1 ( ext−1 < 0)) , (5)

y
 2  y 2 y
 y   y  
ht = C y + a12 A ext−1 + a22 A et−1 + b12 hxt−1 + b22 ht−1 +a22 B et−1 ( et−1 < 0)) . (6)
 2       y 2
In the above equations, A ext−1 and B ext−1 ( ext−1 < 0)) as well as A et−1 and
 y   y  
B et−1 ( et−1 < 0)) reveal the relationships between a market’s volatility and both positive and
negative own lagged returns, respectively (Lin et al. 2014). Equations (5) and (6) show that the
conditional variance of each market depends upon its past shock and past volatility, as well as the past
 2  y 2
shock and past volatility of other markets. In Equation (5), ext−1 and et−1 explain how the past
y
shocks of both x and y affect the current conditional volatility of x. Moreover, hxt−1 and ht−1 measure
how the past volatilities of both x and y affect the current conditional volatility of x. The parameters
of the VAR-AGARCH model can be estimated by using the Quasi-Maximum Likelihood estimation
(QMLE) and using the BFGS algorithm.5
The estimates of the VAR-AGARCH model can be used to calculate optimal portfolio weights.
This study followed Kroner and Ng (1998) to calculate the optimal portfolio weights for the pairs of
the stock market (x, y) as:
h y,t − hxy,t
wxy,t = (7)
hx,y − 2hxy,t + h y,t
⎧
⎪
⎪ 0 i f Wxy,t < 0
⎪
⎪
⎨
wxy,t = ⎪
⎪w i f 0 ≤ wxy,t ≤ 1 ,
⎪ xy,t
⎪
⎩ 1 i f wxy,t > 1
where wxy,t is the weight of stock(x) in a $1 stock(x)-stock(y) portfolio at time t, hxy,t is the conditional
covariance between the two stock markets, hx,t and h y,t are the conditional variance of stock(x) and
stock(y), respectively, and 1-wxy,t is the weight of stock(y) in a $1 stock(x)-stock(y) portfolio.
It is also essential to estimate the risk-minimizing optimal hedge ratios for the portfolio of different
stocks. The estimates of the VAR-AGARCH model can also be used to calculate optimal hedge ratios.
This study followed Kroner and Sultan (1993) to calculate the optimal hedge ratios as:

hxy,t
βxy,t = , (8)
h y,t

where βxy,t represents the hedge ratio. This shows that a short position in the stock (y) market can
hedge a long position in the stock (x). Lastly, RATS 10.0 software is used for estimations.

4. Empirical Results

4.1. Descriptive Analysis

Table 1 reports the summary statistics of the daily returns for the US, China, and eight emerging
Asian stock markets, namely India, Korea, Indonesia, Pakistan, Malaysia, the Philippines, Thailand,
and Taiwan. The average returns of the Pakistani stock market are the highest out of these
markets, whereas the lowest returns are found in the US stock market during the full sample
period. The unconditional volatility is lower in Malaysia and the US market and is highest in the

5 Arouri et al. (2011), Sadorsky (2012), and Allen et al. (2013) use the Quasi-Maximum Likelihood estimation (QMLE) and use
the BFGS algorithm to estimate the parameters in the VAR-GARCH model.

10
JRFM 2020, 13, 226

Chinese stock market. The skewness is negative in all cases, kurtosis is higher than 3 for all stocks,
and Jarque–Bera statistics do not accept the hypothesis of the normality for all stocks. Moreover,
we applied the Ljung–Box Q test for autocorrelation to the standardized residuals and squared
standardized residuals. The coefficients both Q(12) and Q2 (12) were found to be signifcant for all series.
ARCH effects were also statistically significant for all series.6

Table 1. Summary Statistics.

ARCH
Mean Median Max Min Std. Dev. Skewness Kurtosis Jarque-Bera Q-Stat
Test
USA 0.00016 0.00055 0.10958 −0.09470 0.01200 −0.20353 11.57202 14802.7 a 37.24 a 206.42 a
CHN 0.00045 0.00096 0.09401 −0.09256 0.01570 −0.31725 8.21506 5547.4 a 54.64 a 180.10 a
IND 0.00050 0.00094 0.15990 −0.11809 0.01472 −0.22234 10.54239 11474.1 a 84.62 a 283.89 a
INDO 0.00046 0.00113 0.07623 −0.10954 0.01357 −0.85402 10.92376 13206.3 a 154.0 a 457.66 a
KOR 0.00028 0.00080 0.11284 −0.12805 0.01509 −0.57337 9.64860 9149.3 a 24.06 a 210.02 a
MYS 0.00020 0.00041 0.04503 −0.09979 0.00816 −0.85496 13.33067 22038.9 a 226.8 a 267.36 a
PAK 0.00081 0.00109 0.08507 −0.07741 0.01359 −0.34875 6.83764 3058.01 a 165.5 a 594.62 a
PHL 0.00038 0.00055 0.16178 −0.13089 0.01309 0.23024 19.78304 56658.3 a 96.40 a 161.15 a
TAIW 0.00018 0.00070 0.06525 −0.09936 0.01356 −0.27454 6.59593 2659.6 a 77.68 a 201.54 a
THA 0.00044 0.00064 0.10577 −0.16063 0.01316 −0.70520 12.86191 19948.5 a 70.19 a 656.27 a
Notes: a indicates the statistical significance at 1% level.

4.2. Return, Shock and Volatility Spillover Analysis

4.2.1. Stock Market Linkages between the USA and Asia from the Full Sample Period
Table 2 represents the return and volatility spillover between US and Asian stock markets
during the full sample period. The lagged stock returns were found to significantly affect the current
stock returns in all studied Asian stock markets except for Korea. This highlights the possibility of
short-term predictions of current returns through past returns in the Asian stock markets. Moreover,
the autoregressive term of the USA stock market was found to be significant as well. This depicts that
past returns help to predict current returns in the American stock market.
The estimate of return spillover from one market to another market can be estimated by using
the coefficient of lagged return of one market (i.e., the US) onto another market (i.e., India) and vice
versa. The return spillover from the USA to all Asian stock markets is significant. This implies that US
stock market prices play an important role in predicting the prices of all Asian stock markets during
the full sample period. These results are in line with the findings of Huyghebaert and Wang (2010),
which find a significant return spillover from the USA to Asian markets. This shows that the effect
of the returns of the American stock market are significantly transmitted to the Asian stock markets.
However, the return spillover from all Asian stock markets to the USA was found to be insignificant.
This implies that Asian stock market prices are not helpful in predicting the prices of the US stock
market during the full sample period.

6 We applied both Augmented Dickey–Fuller (ADF), and Phillip–Perron (PP) tests to examine the stationarity of all returns
series and found that all returns series are stationary, but we do not report these results in Table form for the sake of brevity.

11
 Table 2. Estimates of bivariate Vector Autoregressive-Generalized Autoregressive Conditional Heteroskedasticity (VAR-AGARCH) for the USA and Asian stock
markets during a full sample period.

IND USA INDO USA KOR USA MYS USA PAK USA PHL USA TAIW USA THA USA
Panel A: Mean Equation
JRFM 2020, 13, 226

4.90 × 2.48 × 5.13 × 2.40 × 1.55 × 2.08 × 1.08 × 2.46 × 8.79 × 2.07 × 3.53 × 2.20 × 1.28 × 2.64 × 6.11 × 2.71 ×
Constant 10−4 a 10−4 c 10−4 a 10−4 b 10−4 10−4 c 10−4 10−4 b 10−4 a 10−4 b 10−4 b 10−4 a 10−4 10−4 b 10−4 a 10−4 c
(0.001) (0.023) (0.001) (0.022) (0.252) (0.059) (0.195) (0.026) (0.000) (0.048) (0.030) (0.001) (0.296) (0.016) (0.000) (0.010)
−7.65 × −4.52 × −1.87 × −8.99 ×
0.100 a 0.136 a −0.010 0.017 0.016 0.168 a −0.022 0.168 a 0.118 a 0.054 a 0.014 0.093 a
rot−1 10−4 10−3 10−3 10−3
(0.000) (0.000) (0.260) (0.241) (0.126) (0.000) (0.232) (0.000) (0.000) (0.000) (0.232) (0.000)
(0.933) (0.648) (0.132) (0.423)
6.40 ×
0.231 a −0.041 a 0.301 a −0.032 b 0.420 a −0.037 b 0.201 a −0.034 b −0.030 b 0.421 a −0.030 b 0.386 a −0.037 b 0.224 a −0.037 b
rut−1 10−3 b
(0.000) (0.009) (0.000) (0.031) (0.000) (0.017) (0.000) (0.035) (0.040) (0.000) (0.014) (0.000) (0.021) (0.000) (0.021)
(0.017)
Panel B: Variance Equation
2.68 × 1.72 × 1.46 × 1.47 × 1.40 × 1.71 × 6.67 × 1.70 × 7.61 × 2.30 × 3.82 × −1.46 × 1.11 × 1.64 × 4.90 × 1.72 ×
Constant 10−6 a 10−6 a 10−5 a 10−6 a 10−6 a 10−6 a 10−7 a 10−6 a 10−6 a 10−6 a 10−5 a 10−6 a 10−6 a 10−6 a 10−6 a 10−6 a
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)

 2 7.04 × −1.84 × 8.93 ×

0.056 a 0.038 a 0.118 a 0.057 a 0.053 a 0.020 a 0.090 a 0.100 a 0.125 a 0.041 a 0.037 a 0.035 a 0.080 a
eot−1 10−3 a 10−3 a 10−3
(0.000) (0.000) (0.000) (0.000) (0.000) (0.001) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.002) (0.000) (0.315)

12
 2 8.17 × −5.68 × 2.30 × 6.81 × 4.97 × 5.98 × −6.98 × −4.96 × 6.71 × −8.59 × 7.89 × −9.45 ×
−0.011 c −0.011 −0.011 c −0.014 a
eut−1 10−3 a 10−3 10−3 10−3 b 10−3 10−3 a 10−3 c 10−3 a 10−3 10−3 10−3 b 10−3
(0.052) (0.108) (0.073) (0.000)
(0.001) (0.401) (0.538) (0.032) (0.524) (0.000) (0.083) (0.000) (0.110) (0.228) (0.012) (0.119)
2.77 × 1.71 × 5.47 ×
0.877 a −0.030 a 0.579 a 0.085 0.906 a −0.018 0.853 a 0.791 a 0.506 a −0.029 0.925 a −0.030 a 0.813 a
hot−1 10−3 10−4 10−3
(0.000) (0.000) (0.000) (0.130) (0.000) (0.290) (0.000) (0.000) (0.000) (0.142) (0.000) (0.000) (0.000)
(0.378) (0.369) (0.599)
−4.98 × 6.82 × −2.59 × −7.80 × 2.08 × −4.67 ×
0.896 a 0.908 a 0.900 a 0.010 0.894 a 0.892 a 0.036 a 0.903 a 0.897 a 0.900 a
hut−1 10−3 b 10−4 10−3 10−3 a 10−4 10−3 c
(0.000) (0.000) (0.000) (0.355) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.031) (0.915) (0.407) (0.000) (0.968) (0.076)
0.098 a 0.172 a 0.199 a 0.164 a 0.068 a 0.173 a 0.060 a 0.183 a 0.143 a 0.193 a 0.143 a 0.167 a 0.051 a 0.167 a 0.157 a 0.175 a
Asymmetry
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Panel C: Constant Conditional Correlation
0.198 a 0.104 a 0.185 a 0.082a 0.031b 0.054 a 0.142 a 0.134 a
p0,u
(0.000) (0.000) (0.000) (0.000) (0.030) (0.000) (0.000) (0.000)
Panel D: Diagnostic Tests
LogL 30500.4 30678.873 30692.048 33342.351 30597.4 30633.0 30931.2 30683.3
AIC −9.520 −9.912 −10.060 −11.074 −10.044 −10.537 −10.030 −10.174
SIC −9.126 −9.518 −9.671 −10.680 −9.040 −10.143 −9.780 −9.779
 Table 2. Cont.

IND USA INDO USA KOR USA MYS USA PAK USA PHL USA TAIW USA THA USA
Panel D: Diagnostic Tests
1921.200 462.680 20331.6 560.990
533.870 a 381.111 a 688.360 a 528.060 a 870.230 a 391.040 a a 500.410 a 3442.14 a 508.010 a 10842.7 a 438.390 a 295.380 a a a a
JB
JRFM 2020, 13, 226

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.000) (0.000) (0.000) (0.000)
16.143 22.788 9.012 15.021 5.829 18.049 13.179 14.149 55.870 c 15.244 21.845 b 15.008 18.533 c 19.566 c 33.450 c 16.279
Q (12)
(0.185) (0.303) (0.702) (0.240) (0.924) (0.114) (0.356) (0.291) (0.097) (0.228) (0.039) (0.241) (0.100) (0.076) (0.087) (0.179)
13.026 9.413 19.566 c 8.416 5.242 9.919 14.799 10.027 5.970 0.000 a 1.823 0.000 a 16.456 10.976 0.928 11.617
Q2 (12)
(0.367) (0.667) (0.076) (0.752) (0.949) (0.623) (0.253) (0.614) (0.918) (0.000) (0.923) (0.000) (0.171) (0.531) (0.965) (0.477)
Notes: The number of lags for VAR was decided using SIC (Schwartz information criterion) and AIC (Akaike information criterion) criteria. JB, Q(12), and Q2 (12) indicate the empirical
statistics of the Jarque–Bera test for normality, while Ljung–Box Q statistics with order 12 for autocorrelation were applied to the standardized residuals and squared standardized
residuals, respectively. USA, United States of America; IND, India; INDO, Indonesia; KOR, South Korea; MYS, Malaysia; PAK, Pakistan; PHL, the Philippines; TAIW, Taiwan; THA,
Thailand. Values in parentheses are the p-values. a ,b ,c indicate the statistical significance at 1%, 5%, and 10%, respectively.

13
JRFM 2020, 13, 226

ARCH coefficient captures the shock dependence, while the GARCH coefficient captures the
persistence of volatility in conditional variance equations. The findings reveal that the sensitivity of
past own shocks (ARCH term) is significantly positive for all Asian Stock Markets in the short run.
In addition, the sensitivity of past own volatility (the GARCH term) was found to be significant for all
stock markets (including the Asian and American Markets), thus the ARCH (1) volatility model was
determined to be more appropriate in this case. The coefficient of past own volatility was than the
coefficients of past own shocks in all Asian stock markets, implying that past own volatilities are more
critical for prediction of future volatility as compared to past own shocks.
The conditional volatility of India’s, South Korea’s, the Philippines’, Pakistan’s, and Thailand’s
stock markets was found to be significantly affected by shocks in the American stock market.
These results are similar to the findings of Syriopoulos et al. (2015), which show that past shocks in
the American market significantly affect the market volatility of India, Brazil, and Russia. Therefore,
this implies that shock in the American stock market leads to an increase in the volatility of the
majority of Asian markets. The past volatility of the American stock market significantly influenced
the conditional volatility of India’s, The Philippines’, Pakistan’s, and Thailand’s stock markets.
These results confirm the previous findings of Li and Giles (2015), which finds a significant volatility
spillover from the USA to emerging Asian stock markets. Further, Syriopoulos et al. 2015 found a
significant volatility spillover from the USA to India. In addition, the past volatility of the majority of
Asian Markets (Except for India and Taiwan) has not been significantly transmitted to the American
stock market. The asymmetric coefficients of all Asian stock markets are significant and positive,
showing that negative news (or unexpected shocks) for the American stock market has more ability to
increase the volatility of all Asian Stock markets as compared to positive news.
Besides, the asymmetric coefficient of the American stock market is positively significant,
demonstrating that negative unexpected shocks in Asian Stock markets will increase the volatility
more in the American Stock market as compared to a positive shock. Constant conditional correlation
(CCC) is positively significant for all pairs of stock markets. However, cross-market correlation is
weak in almost all pairs, indicating that investors can get substantial gains by having these pairs in the
same portfolio.

4.2.2. Stock Market Linkages between China and Asia from the Full Sample Period
Table 3 reports the return and volatility spillover between the Chinese and other Asian stock
markets during the full sample period. The current stock returns of Asian stock markets are significantly
affected by their own lagged stock returns. This highlights the possibility of short-term predictions
of current returns through past returns in the Asian stock markets. Moreover, Chinese stock returns
are also significantly influenced by their own single period lagged returns. These findings depict that
stock prices can be predicted in the short term in the Chinese stock market.
The return spillover is not significant from China to the majority of other Asian markets except for
the Indian, Philippines, and Thai stock markets. Besides, the return transmission from Asian markets
to the Chinese market is insignificant except for in the case of the Indian Stock market. Moreover,
there is a presence of bi-directional return transmission between the Indian and Chinese stock markets.
This implies that Chinese (Indian) stock market prices play an important role in predicting the prices
of Indian (Chinese) stock markets during the full sample period. The coefficient of past own shock
of all Asian markets (including China) was found to be significant; thus, past shocks affect current
conditional volatility in Asian stock markets. Besides, the sensitivity of past own volatility for all Asian
markets was found to be significant as well.

14
 Table 3. Estimates of bivariate VAR−AGARCH for China’s and other Asian stock markets during the full sample period.

IND CHN INDO CHN KOR CHN MYS CHN PAK CHN PHL CHN TAIW CHN THA CHN
Panel A: Mean Equation
5.06 × 3.25 × 5.37 × 3.50 × 2.71 × 3.26 × 1.69 × 3.74 × 8.64 × 3.39 × 3.32 × 3.50 × 2.18 × 3.51 × 5.05 × 2.12 ×
Constant 10−4 a 10−4 c 10−4 a 10−4 b 10−4 c 10−4 b 10−4 b 10−4 b 10−4 a 10−4 b 10−4 b 10−4 b 10−4 c 10−4 b 10−4 a 10−4
JRFM 2020, 13, 226

(0.000) (0.053) (0.000) (0.022) (0.065) (0.043) (0.036) (0.023) (0.000) (0.029) (0.031) (0.030) (0.089) (0.027) (0.000) (0.139)
6.27 ×
0.137 a 0.036 a 0.152 a 0.017 0.083 a 0.011 0.191 a 0.018 0.167 a 0.144 a −0.028 b 0.100 a 0.017 0.125 a 0.018
rot−1 10−3
(0.000) (0.004) (0.000) (0.237) (0.000) (0.345) (0.000) (0.456) (0.000) (0.000) (0.048) (0.000) (0.241) (0.000) (0.183)
(0.576)
−6.05 × 7.11 × 2.26 ×
−0.018 c 0.049 a 0.050 a −0.013 0.055 a 0.053 a 0.010 0.055 a 0.025 b 0.055 a 0.051 a −0.020 0.049 a
rct−1 10−3 10−3 10−3
(0.096) (0.001) (0.000) (0.269) (0.000) (0.000) (0.272) (0.000) (0.043) (0.000) (0.000) (0.103) (0.001)
(0.603) (0.295) (0.841)
Panel B: Variance Equation
2.14 × 1.14 × 4.30 × 1.19 × 1.00 × 1.37 × 5.84 × 1.26 × 7.29 × 1.15 × 1.36 × 1.19 × 9.17 × 1.21 × 1.01 × 1.19 ×
Constant 10−6 a 10−6 a 10−6 a 10−6 a 10−6 a 10−6 a 10−7 a 10−6 a 10−6 a 10−6 a 10−5 a 10−6 b 10−7 a 10−6 a 10−6 a 10−6 a
(0.000) (0.001) (0.000) (0.002) (0.000) (0.000) (0.000) (0.000) (0.000) (0.002) (0.000) (0.048) (0.000) (0.000) (0.001) (0.001)

 2 −1.95 × −1.77 × 7.03 × 3.80 × −9.47 × 8.20 × 3.74 ×

0.051 a 0.074 a 0.031a 0.074 a 0.100 a 0.053 a 0.033 a 0.070 a 0.023 a
eot−1 10−3 10−4 10−3 10−4 10−4 10−3 b 10−3 b
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.413) (0.950) (0.241) (0.668) (0.509) (0.016) (0.034)

 2 −5.03 × 9.29 × −1.32 × −6.92 × 3.56 × 8.77 × 9.94 × 7.51 ×

0.069 a 0.066 a 0.070 a 0.069 a 0.064 a 0.068 a 0.069 a 0.062 a

15
ect−1 10−3 c 10−3 c 10−3 10−3 10−3 10−4 10−3 a 10−3 b
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.081) (0.099) (0.551) (0.422) (0.453) (0.754) (0.000) (0.025)
9.48 × −6.06 × 2.91 × 1.39 × 2.22 × −1.34 × −8.97 ×
0.876 a 0.840 a 0.011 b 0.929 a 0.871 a 0.792 a 0.797 a 0.922 a 0.878 a
hot−1 10−3 a 10−3 a 10−3 b 10−3 10−3 10−3 10−3 a
(0.000) (0.000) (0.041) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.001) (0.000) (0.022) (0.458) (0.656) (0.508) (0.001)
7.61 × −4.86 × 2.24 × −1.29 × 1.53 × −8.00 × −3.06 ×
0.920 a 0.923 a 0.918 a 0.010 0.921 a 0.923 a 0.921 a 0.921 a 0.919 a
hct−1 10−3 b 10−3 10−3 10−3 10−3 10−3 a 10−3
(0.000) (0.000) (0.000) (0.278) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.033) (0.482) (0.314) (0.816) (0.816) (0.002) (0.497)
0.105 a 0.016 b 0.096 a 0.012 0.067 a 0.019 b 0.073 a 0.016 c 0.139 a 0.019 b 0.096 a 0.018 b 0.072 a 0.016 b 0.073 a 0.030 a
Asymmetry
(0.000) (0.048) (0.000) (0.138) (0.000) (0.023) (0.000) (0.055) (0.000) (0.012) (0.000) (0.035) (0.000) (0.050) (0.000) (0.001)
Panel C: Constant Conditional Correlation
0.158 a 0.164 a 0.211 a a 0.047 a 0.131 a 0.218 a 0.146 a
p0,c
(0.000) (0.000) (0.000) (0.000) (0.001) (0.000) (0.000) (0.000)
Panel D: Diagnostic Tests
LogL 28559.5 28690.5 28584.3 31339.5 28776.2 28564.8 28882.9 28821.945
AIC 9.191 −9.543 −9.621 −10.762 −9.824 −9.920 −9.766 −9.770
SIC −8.797 −9.149 −9.227 −10.368 −9.430 −9.520 −9.372 −9.375
 Table 3. Cont.

IND CHN INDO CHN KOR CHN MYS CHN PAK CHN PHL CHN TAIW CHN THA CHN
Panel D: Diagnostic Tests
1238.55 56026.9 1380.10
574.750 a 1348.92 a 1310.07 a 1391.71 a 1385.22 a 705.970 a 1651.18 a 1428.96 a 3079.52 a 1424.88 a 1248.20 a 1408.79 a 464.070 a a a a
JB
JRFM 2020, 13, 226

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.000) (0.000) (0.000)
17.399 48.803 a 10.709 45.754 a 9.252 46.376 a 15.229 44.283 a 54.789 a 51.513 a 14.166 50.473 a 24.449 b 39.022 a 31.740 a 48.687 a
Q (12)
(0.135) (0.000) (0.554) (0.000) (0.681) (0.000) (0.229) (0.000) (0.000) (0.000) (0.290) (0.000) (0.018) (0.000) (0.002) (0.000)
11.249 11.647 17.033 9.322 5.203 11.672 20.075 c 11.915 6.182 12.065 4.715 10.218 11.938 13.649 2.396 10.561
Q2 (12)
(0.508) (0.474) (0.148) (0.675) (0.951) (0.472) (0.066) (0.453) (0.907) (0.440) (0.967) (0.597) (0.451) (0.324) (0.999) (0.567)
Notes: The number of lags for VAR was decided using SIC (Schwartz information criterion) and AIC (Akaike information criterion) criteria. JB, Q(12) and Q2 (12) indicated the empirical
statistics of the Jarque–Bera test for normality, while Ljung–Box Q statistics of order 12 for autocorrelation applied to the standardized residuals and squared standardized residuals,
respectively. CHN, China; IND, India; INDO, Indonesia; KOR, South Korea; MYS, Malaysia; PAK, Pakistan; PHL, the Philippines; TAIW, Taiwan; THA, Thailand. Values in parentheses
are the p-values. a , b , c indicate the statistical significance at 1%, 5%, and 10%, respectively.

16
JRFM 2020, 13, 226

The conditional volatility of India, Indonesia, Taiwan, and Thailand is significantly affected by
shocks in the Chinese market. Also, the conditional volatility of the Chinese market is significantly
impacted by the shocks in the Philippines, Taiwanese, and Thai stock markets. The past volatility
of the Chinese stock market has not influenced the conditional volatility of the most of the Asian
stock markets except for the Indian and Taiwanese stock markets. These findings corroborate with the
results of Zhou et al. (2012), which report a significant spillover from China to the Taiwanese stock
market. However, the past volatility of the majority Asian markets (except for Pakistan, the Philippines,
and Taiwan) significantly affected the conditional volatility of the Chinese stock market.
The asymmetric coefficients of all Asian stock markets were found to be significant and positive,
showing that negative news of the Chinese stock market has more of an ability to increase the volatility
of all Asian stock markets as compared to positive news. Moreover, the asymmetric coefficient of
the Chinese stock market is significant and positive, showing that negative news in Asian markets
(except in Indonesia) has a greater ability to increase the volatility of the Chinese market as compared
to positive news. Constant conditional correlation is positively significant for all pairs of stock markets,
but CCC is weak in majority pairs.

4.2.3. Stock Market Linkages between the USA and Asia from the US Financial Crisis
Table 4 shows the mean and volatility spillover between the USA and Asian stock markets during
the US financial crisis. In Asian Stock markets (except for South Korea), past lagged returns significantly
influenced the current returns. This highlights the possibility of short-term prediction of current
returns through past returns in the Asian stock markets. Moreover, the American stock returns were
also significantly influenced by their own single period lagged returns in the majority of cases.
The return spillover effect from the USA to all Asian markets was seen to be significant during the
US financial crisis. This implies that US stock market prices played an important role in predicting the
prices of all Asian stock markets during the US financial crisis. These results confirm the previous
findings of Glick and Hutchison (2013), who reported a significant impact of American equity returns
on Asian equity returns during the US financial crisis. Moreover, no single Asian stock market
transmitted the return effect to the American market during the US financial crisis. The sensitivity
of past own shock was significant for the majority of Asian markets other than Indonesia, Korea,
and Taiwan. The coefficient of past own shocks of the American stock market was insignificant in the
majority estimations. Besides, the coefficient of own past volatility in all Asian markets was significant
except in the Philippines.
The past shocks in the American stock market significantly influenced the conditional volatility of
Korea, the Philippines, and Taiwan during the US financial crisis. However, past shocks in most of
the Asian stock markets (Except India) have not affected the conditional volatility of the American
stock market. The effect of past volatility in the USA on conditional volatility of the Asian stock
markets (except Korea) was found to be insignificant. These results match with the findings of
Li and Giles (2015), which observe an absence of volatility spillover from the USA to emerging Asian
stock markets during the US financial crisis. Moreover, the past volatility of majority Asian stock has
not significantly affected American stock market volatility. The asymmetric coefficient of all Asian
markets is significant and positive. Moreover, the asymmetric coefficient of the US market is significant
and positive in all cases. Constant conditional correlation is positively significant for all pairs of stock
markets, but CCC is weak in majority pairs.

17
 Table 4. Estimates of bivariate VAR−AGARCH for the American and Asian stock markets during the US Financial Crisis.

IND USA INDO USA KOR USA MYS USA PAK USA PHL USA TAIW USA THA USA
Panel A: Mean Equation
7.60 × 3.49 × −2.0 × 8.03 × 8.5 × −7.01 × 3.72 × −1.43 × 9.53 × 3.57 × 9.83 × 4.79 × 1.90 × −2.64 × 6.00 × −6.12 ×
Constant 10−4 10−4 10−5 10−5 10−4 c 10−5 10−4 10−4 10−4 a 10−4 10−4 a 10−4 a 10−4 10−5 10−4 10−5
JRFM 2020, 13, 226

(0138) (0.936) (0957) (0.834) (0.080) (0.865) (0.123) (0.742) (0.005) (0.262) (0.001) (0.066) (0.684) (0.955) (0.194) (0.886)
0.081 b 0.016 0.031 b −0.024 0.044 −0.010 0.145 a −0.046 0.125 a 0.015 0.101 a 0.016 0.063 c 0.020 0.065 b −0.012
rot−1
(0.031) (0.510) (0.036) (0.445) (0.141) (0.713) (0.000) (0.409) (0.001) (0.648) (0.000) (0.523) (0.076) (0.565) (0.027) (0.663)
1.60 ×
0.293 a −0.093 b 0.423 a −0.062 a 0.325 a −0.079 b 0.187 a −0.080 b 0.175 a 0.016 0.489 a 0.383 a −0.103 b 0.249 a −0.077 b
rut−1 10−3
(0.000) (0.016) (0.000) (0.091) (0.000) (0.041) (0.000) (0.042) (0.000) (0.706) (0.000) (0.000) (0.011) (0.000) (0.034)
(0.963)
Panel B: Variance Equation
2.91 × 3.39 × 1.33 × 5.10 × 3.98 × 4.34 × 1.80 × 3.40 × 4.22 × 3.93 × 3.87 × 2.98 × 1.15 × 3.73 × 3.72 × 4.59 ×
Constant 10−6 10−6 a 10−5 a 10−6 a 10−5 a 10−6 a 10−6 b 10−6 a 10−6 a 10−6 a 10−5 a 10−6 a 10−6 10−6 a 10−5 a 10−8
(0.161) (0.001) (0.000) (0.000) (0.000) (0.000) (0.038) (0.001) (0.007) (0.000) (0.000) (0.000) (0.352) (0.009) (0.000) (0.984)

 2 −7.24 × 7.53 × 4.98 ×

0.086 a 0.075 b 0.198 a 0.001 0.012 a 0.077 c 0.119 a 0.036 b 0.283 0.051 b 0.029 0.151 a 0.015
eot−1 10−4 10−3 10−3
(0.000) (0.016) (0.000) (0.877) (0.000) (0.078) (0.000) (0.034) (0.140) (0.044) (0.230) (0.000) (0.678)
(0.879) (0.439) (0.761)

 2 6.65 × −0.01 × 7.71 × −4.01 × −5.15 × −6.17 ×

0.011 −0.029 b 0.001 0.051 −0.013 0.018 −0.060 a −0.061 a 0.026 c −0.022
eut−1 10−3 a 10−3 10−3 b 10−3 b 10−3 10−3
(0.138) (0.014) (0.917) (0.213) (0.361) (0.146) (0.001) (0.000) (0.056) (0.113)

18
(0.053) (0.337) (0.052) (0.033) (0.690) (0.661)
8.26 ×
−0.868 a −0.048 b 0.750 a 0.060 0.426 a −0.012 0.755 a 0.013 c 0.854 a 0.099 0.043 0.951 a −0.029 c 0.449 a 0.110 b
hot−1 10−3
(0.000) (0.012) (0.000) (0.208) (0.000) (0.330) (0.000) (0.090) (0.000) (0.196) (0.171) (0.000) (0.062) (0.000) (0.024)
(0.470)
2.17 × 5.62 × 8.49 ×
0.894 a 0.945 a −0.176 0.906 a −0.035 0.915 a −0.021 0.907 a 0.906 a −0.017 0.910 a 0.031 0.905 a
hut−1 10−3 c 10−4 10−3
(0.000) (0.000) (0.394) (0.000) (0.312) (0.000) (0.322) (0.000) (0.000) (0.382) (0.000) (0.299) (0.000)
(0.087) (0.765) (0.115)
0.064 c 0.179 a 0.265 a 0.159 a 0.446 a 0.161 a 0.163 a 0.154 a 0.101 a 0.237 a 0.523 a 0.233 a 0.075 a 0.169 a 0.187 a 0.147 a
Asymmetry
(0.090) (0.000) (0.000) (0.000) (0.000) (0.000) (0.004) (0.000) (0.004) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Panel C: Constant Conditional Correlation
0.172 a 0.280 a 0.220 a 0.195 a 0.016c 0.065b 0.187a 0.218 a
p0,u
(0.000) (0.000) (0.000) (0.000) (0.086) (0.044) (0.000) (0.000)
Panel D: Diagnostic Tests
LogL 4230.27 4320.18 4436.36 4843.96 4850.88 6628.56 4391.36 4399.67
AIC −10.066 −10.502 −10.394 −11.480 −11.499 −11.518 −10.601 −10.710
SIC −9.773 −10.209 −10.101 −11.187 −11.205 −11.224 −10.308 −10.417
 Table 4. Cont.

IND USA INDO USA KOR USA MYS USA PAK USA PHL USA TAIW USA THA USA
Panel D: Diagnostic Tests
635.280 330.040 643.320
513.31 a 372.23 a 679.01 a 512.65 a 843.871 a 399.862 a 321.220 a 627.410 676.080 a 654.370 a 1598.28 a 629.470 a 352.580 a a a a
JB
JRFM 2020, 13, 226

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.000) (0.000) (0.000)
14.588 8.939 11.706 6.002 12.006 7.401 11.897 6.280 17.579 9.780 9.776 7.878 14.047 5.797 13.234 5.468
Q (12)
(0.264) (0.708) (0.469) (0.915) (0.445) (0.829) (0.454) (0.901) (0.129) (0.635) (0.636) (0.795) (0.298) (0.926) (0.352) (0.941)
7.779 13.979 5.303 14.983 6.336 13.673 5.422 14.540 6.417 32.448 a 2.113 20.023 c 6.158 13.940 8.596 14.294
Q2 (12)
(0.802) (0.301) (0.947) (0.242) (0.898) (0.322) (0.942) (0.268) (0.894) (0.001) (0.999) (0.067) (0.908) (0.305) (0.737) (0.282)
Notes: The number of lags for VAR was decided using SIC (Schwartz information criterion) and AIC (Akaike information criterion) criteria. JB, Q(12) and Q2 (12) indicate the empirical
statistics of the Jarque–Bera test for normality, while Ljung–Box Q statistics of order 12 for autocorrelation were applied to the standardized residuals and squared standardized residuals,
respectively. USA, United States of America; IND, India; INDO, Indonesia; KOR, South Korea; MYS, Malaysia; PAK, Pakistan; PHL, the Philippines; TAIW, Taiwan; THA, Thailand.
Values in parentheses are the p-values. a , b , c indicate the statistical significance at 1%, 5%, and 10%, respectively.

19
JRFM 2020, 13, 226

4.2.4. Stock Market Linkages between China and Asia during the US Financial Crisis
Table 5 reports the return and volatility spillover between China and Asian stock markets during
the US financial crisis. The current stock returns of the majority of Asian stock markets (Except in
South Korea) are significantly affected by their own lagged stock returns. This highlights the possibility
of short-term prediction of current returns through past returns in the Asian stock markets. However,
Chinese stock returns were not significantly affected by their lagged returns during the US financial
crisis. Therefore, there is no evidence of Chinese stock price prediction being possible through lagged
values during the US financial crisis.
The return transmission effect from China to all Asian markets was insignificant during the US
financial US crisis. However, most of the Asian markets did not transmit the return effect to the
Chinese stock market other than India, Indonesia, and Malaysia. The coefficient of past own shock
was found to be significant in the majority of Asian markets except for Indonesia, Korea, and Pakistan.
The sensitivity to past own shocks from Chinese stock markets was found to be insignificant in the
majority of markets during the US financial crisis. Moreover, the sensitivity of past own volatility in all
Asian markets was significant. The past shocks of China did not influence the conditional volatility
of the majority of Asian stock markets (except India) during the US financial crisis. The conditional
volatility of the Chinese stock market was not affected by shocks in most of the Asian stock markets
(except for Indonesia and Thailand).
There is no significant evidence of volatility spillover from Chinese to Asian stock markets except
in India and Taiwan. Besides, the volatility spillover was insignificant in the majority of Asian markets
(except Indonesia, Pakistan, and The Philippines) to the Chinese stock market. The asymmetric
coefficient of all Asian markets is significant and positive. Moreover, the asymmetric coefficient of
China is asymmetric, showing that that negative news of all Asian stock markets (except Pakistan)
has more ability to increase the volatility of the Chinese stock market as compared to positive news.
Constant conditional correlation is positively significant for all pairs of stock markets. However,
CCC has a medium level in the majority of pairs.

4.2.5. Stock Market Linkages between the USA and Asia from the Chinese Stock Market Crash
Table 6 reports the mean and volatility spillover between the USA and Asian stock markets
during the Chinese stock market crash. The autoregressive term of Asian market returns (Except
Korea, the Philippines, and Taiwan) can be seen to be significant in the majority of stock markets.
This shows the short-term predictability in stock price changes in the Asian stock markets. In addition,
stock returns of the American stock market were significantly influenced by their lagged returns during
the Chinese Crisis.

20
 Table 5. Estimates of bivariate VAR−AGARCH for China and Asian stock markets during the US Financial Crisis.

IND CHN INDO CHN KOR CHN MYS CHN PAK CHN PHL CHN TAIW CHN THA CHN
Panel A: Mean Equation
6.04 × −2.82 × 6.72 × −2.28 × −7.49 × −2.18 × 4.34 × −5.07 × 7.61 × 1.87 × 2.57 × −8.84 × 2.65 × 3.46 × 7.98 × 9.09 ×
Constant 10−4 10−4 10−4 10−4 10−5 10−5 10−4 10−5 10−4 b 10−4 10−4 10−5 10−4 10−5 10−4 c 10−5
JRFM 2020, 13, 226

(0.2680 (0.671) (0.220) (0.712) (0.867) (0.974) (0.146) (0.943) (0.017) (0.684) (0.615) (0.895) (0.580) (0.954) (0.091) (0.876)
8.15 × 5.75 ×
0.147 a 0.109 a 0.146 a 0.136 a 0.052 0.216 a 0.196 a 0.117 a −0.032 0.138 a 0.102 a 0.017 0.128 a 0.045
rot−1 10−3 10−4
(0.000) (0.002) (0.000) (0.001) (0.211) (0.000) (0.008) (0.000) (0.451) (0.001) (0.007) (0.689) (0.000) (0.268)
(0.861) (0.990)
−9.00 × 9.47 × −7.18 ×
−0.044 0.022 0.019 0.011 0.067 c 0.030 0.029 0.031 0.043 0.015 0.052 −0.025 0.054
rct−1 10−3 10−3 10−3
(0.135) (0.576) (0.614) (0.675) (0.086) (0.432) (0.259) (0.432) (0.282) (0.570) (0.183) (0.233) (0.147)
(0.728) (0.534) (0.773)
Panel B: Variance Equation
5.43 × 1.43 × 3.74 × 3.62 × 3.52 × 1.61 × 2.63 × 1.48 × −2.77 × 3.62 × 5.69 × 1.70 × 1.84 × 1.10 × 1.50 × 8.77 ×
Constant 10−6 c 10−5 a 10−5 a 10−6 10−6 10−5 a 10−6 c 10−5 a 10−6 10−6 c 10−5 a 10−5 b 10−6 10−5 a 10−5 a 10−6
(0.064) (0.002) (0.001) (0.389) (0.101) (0.004) (0.087) (0.004) (0.365) (0.064) (0.000) (0.043) (0.223) (0.050) (0.007) (0.144)

 2 7.41 × 5.31 × 2.75 × −6.75 × 5.44 × 1.27 × 7.35 ×

0.067 a −0.025 b −0.016 0.112 a −0.018 0.070 a 0.022 0.055 b −0.011 b
eot−1 10−3 10−3 10−3 10−4 10−3 10−3 10−3
(0.001) (0.013) (0.271) (0.000) (0.269) (0.001) (0.108) (0.019) (0.044)
(0.446) (0.821) (0.636) (0.819) (0.489) (0.870) (0.229)

 2 −7.19 × −2.74 × −6.98 × 9.48 × −7.46 × 1.48 × −6.48 ×

0.027 c 0.015 0.100 0.020 0.025 b 0.017 −0.018 0.013 0.012

21
ect−1 10−3 b 10−3 10−3 10−3 10−3 10−3 10−3
(0.070) (0.398) (0.113) (0.265) (0.047) (0.391) (0.184) (0.388) (0.468)
(0.017) (0.870) (0.748) (0.587) (0.613) (0.864) (0.704)
−7.47 × 3.24 × −5.07 × 4.85 ×
0.893 a −0.016 0.623 a 0.073 c 0.919 a 0.789 a 0.844 a 0.064 a 0.726 a −0.042 b 0.936 a 0.834 a
hot−1 10−4 10−3 10−3 10−3
(0.000) (0.210) (0.000) (0.049) (0.000) (0.000) (0.000) (0.015) (0.000) (0.023) (0.000) (0.000)
(0.924) (0.615) (0.540) (0.637)
1.15 × 7.69 × 2.14 ×
0.030 a 0.865 a 0.060 0.896 a 0.879 a 0.865 a 0.952 a 0.016 0.867 a 0.041 c 0.885a 0.044 0.876 a
hct−1 10−3 10−3 10−3
(0.002) (0.000) (0.157) (0.000) (0.000) (0.000) (0.000) (0.504) (0.000) (0.099) (0.000) (0.128) (0.000)
(0.952) (0.916) (0.909)
0.091 a 0.151 a 0.361 a 0.054 b 0.152 a 0.120 a 0.149 a 0.109 a 0.167 a 0.010 0.153 a 0.126 a 0.059 a 0.116 a 0.124 a 0.130 a
Asymmetry
(0.003) (0.000) (0.000) (0.044) (0.000) (0.001) (0.006) (0.001) (0.000) (0.451) (0.005) (0.000) (0.005) (0.000) (0.001) (0.000)
Panel C: Constant Conditional Correlation
0.307 a 0.290 a 0.401 a 0.287 a 0.103 a 0.270 a 0.373 a 0.269 a
p0,c
(0.000) (0.000) (0.000) (0.000) (0.001) (0.000) (0.000) (0.000)
Panel D: Diagnostic Tests
LogL 3980.79 4085.73 4161.31 4570.93 4646.76 4104.40 4125.88 4144.48
AIC −9.451 −9.746 −9.840 −10.871 −11.412 −9.940 −9.990 −10.001
SIC −9.158 −9.453 −9.547 −10.578 −11.118 −9.650 −9.770 −9.707
 Table 5. Cont.

IND CHN INDO CHN KOR CHN MYS CHN PAK CHN PHL CHN TAIW CHN THA CHN
Panel D: Diagnostic Tests
671.310 358.350 686.450
501.810 a 693.810 a 541.320 a 686.040 a 374.450 a 673.900 a 541.990 a 694.160 590.770 a 756.610 a 1903.28 a 678.760 a 356.750 a a a a
JB
JRFM 2020, 13, 226

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.000) (0.000) (0.000)
14.715 7.454 11.152 7.664 12.457 8.501 11.203 6.933 18.800 c 9.775 11.248 7.940 13.853 8.794 12.923 7.812
Q (12)
(0.257) (0.826) (0.516) (0.811) (0.410) (0.745) (0.512) (0.862) (0.093) (0.636) (0.508) (0.790) (0.310) (0.720) (0.375) (0.800)
8.980 13.615 5.438 14.090 6.297 14.103 6.502 14.401 6.331 20.886 b 27.233 a 13.701 5.748 15.255 6.937 14.507
Q2 (12)
(0.705) (0.326) (0.942) (0.295) (0.900) (0.294) (0.889) (0.276) (0.899) (0.052) (0.007) (0.320) (0.928) (0.228) (0.862) (0.270)
Notes: The number of lags for VAR was decided using SIC (Schwartz information criterion) and AIC (Akaike information criterion) criteria. JB, Q(12) and Q2 (12) indicate the empirical
statistics of the Jarque–Bera test for normality, while Ljung–Box Q statistics of order 12 for autocorrelation were applied to the standardized residuals and squared standardized residuals,
respectively. USA, United States of America; IND, India; INDO, Indonesia; KOR, South Korea; MYS, Malaysia; PAK, Pakistan; PHL, the Philippines; TAIW, Taiwan; THA, Thailand.
Values in parentheses are the p-values. a , b , c indicate the statistical significance at 1%, 5%, and 10%, respectively.

Table 6. Estimates of bivariate VAR−AGARCH for American and Asian stock markets during the Chinese Stock Market Crash.

IND USA INDO USA KOR USA MYS USA PAK USA PHL USA TAIW USA THA USA
Panel A: Mean Equation

22
3.17 × 5.16 × −8.21 × 4.52 × 2.08 × 5.26 × −5.44 × 4.01 × 3.33 × 4.55 × −3.09 × 4.77 × −3.28 × 5.05 × 1.19 × 5.36 ×
Constant 10−4 10−4 b 10−5 a 10−4 b 10−4 10−4 b 10−5 10−4 b 10−4 10−4 b 10−4 10−4 b 10−6 10−4 b 10−4 10−4 a
(0.227) (0.020) (0.743) (0.047) (0.395) (0.013) (0.721) (0.055) (0.225) (0.030) (0.341) (0.029) (0.988) (0.011) (0.557) (0.005)
0.077 b −0.044 0.107 a −0.028 −0.020 −0.018 0.089 b 0.022 0.284 a −0.020 0.032 −0.032 0.031 0.076 0.147 a −0.024
rot−1
(0.032) (0.166) (0.001) (0.359) (0.560) (0.588) (0.013) (0.685) (0.000) (0.367) (0.342) (0.188) (0.359) (0.120) (0.000) (0.519)
0.233 a −0.082 b 0.241 a −0.069 c 0.351 a −0.079 b 0.234 a −0.070 c 0.088 b −0.077 b 0.416 a −0.077 b 0.405 a −0.081 b 0.148 a −0.054
rut−1
(0.000) (0.024) (0.000) (0.065) (0.000) (0.016) (0.000) (0.080) (0.016) (0.044) (0.000) (0.046) (0.000) (0.030) (0.000) (0.173)
Panel B: Variance Equation
3.75 × 4.42 × 2.05 × 4.16 × 1.11 × −1.41 × 3.61 × 3.28 × 4.09 × 4.13 × 4.62 × 1.00 × 2.60 × 3.85 × 1.12 × 7.03 ×
Constant 10−6 b 10−6 a 10−6 a 10−6 a 10−5 a 10−6 10−7 c 10−6 a 10−6 a 10−6 a 10−6 b 10−6 10−6 a 10−6 a 10−6 a 10−7
(0.020) (0.000) (0.005) (0.000) (0.001) (0.670) (0.092) (0.004) (0.000) (0.000) (0.025) (0.714) (0.008) (0.000) (0.001) (0.518)

 2 2.40 ×
−0.034 b 0.079 a 0.086 a 0.149 a 0.073 0.088 a 0.018 −0.018 0.011 0.032 0.063 −0.020 0.071 −0.023 0.020
eot−1 10−3
(0.037) (0.000) (0.003) (0.001) (0.270) (0.000) (0.110) (0.402) (0.425) (0.277) (0.130) (0.271) (0.160) (0.146) (0.311)
(0.911)

 2 1.73 × −6.47 × 9.22 × 4.99 ×

0.051 0.065 c 0.020 −0.077 a 0.046 0.015 −0.014 0.037 0.030 0.068 c 0.061 b 0.014
eut−1 10−3 10−4 10−3 10−3
(0.102) (0.064) (0.534) (0.000) (0.216) (0.640) (0.145) (0.343) (0.162) (0.059) (0.046) (0.650)
(0.915) (0.950) (0.738) (0.296)
−7.32 ×
0.896 a −0.066 a 0.891 a −0.058 0.511 a 0.049 0.865 a 0.856 a −0.020 0.870 a −0.051 a 0.890 a −0.054 a 0.913 a −0.017
hot−1 10−3
(0.000) (0.000) (0.000) (0.202) (0.000) (0.243) (0.000) (0.000) (0.155) (0.000) (0.004) (0.000) (0.001) (0.000) (0.219)
(0.556)
 Table 6. Cont.

IND USA INDO USA KOR USA MYS USA PAK USA PHL USA TAIW USA THA USA
Panel B: Variance Equation
−0.012 0.754 a 0.021 0.696 a 0.204 0.672 a 0.303 a 0.642 a −0.012 a 0.797 a 0.070 0.711 −0.018 0.726 a 0.157 c 0.645 a
hut−1
(0.644) (0.000) (0.306) (0.000) (0.149) (0.000) (0.002) (0.000) (0.002) (0.000) (0.166) (0.290) (0.515) (0.000) (0.072) (0.000)
JRFM 2020, 13, 226

0.126 a 0.270 a 0.123 a 0.314 a −0.079 0.346 a 0.040 0.317 a 0.253 a 0.258 a 0.081 b 0.338 a 0.131 a 0.296 a 0.164 a 0.343 a
Asymmetry
(0.002) (0.000) (0.001) (0.000) (0.183) (0.000) (0.231) (0.000) (0.000) (0.000) (0.035) (0.000) (0.001) (0.000) (0.000) (0.000)
Panel C: Constant Conditional Correlation
0.324 a 0.123 a 0.281 a 0.146 a 0.059 a 0.103 a 0.214 a 0.224 a
p0,u
(0.000) (0.000) (0.000) (0.000) (0.078) (0.001) (0.000) (0.000)
Panel D: Diagnostic Tests
LogL 5580.558 5500.464 5640.723 5895.388 5471.88 5390.25 5605.08 5699.52
AIC −13.712 −13.614 −14.028 −14.525 −13.330 −13.434 −13.877 −13.881
SIC −13.418 −13.321 −13.734 −14.232 −13.037 −13.140 −13.583 −13.587

a a a a a a a a a a a a a 230.350 154.390
72.640 204.890 58.520 412.790 134.540 168.440 22.680 442.380 30.880 428.120 27.070 422.700 84.480 a 68.060 a a
JB
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
(0.000) (0.000)
16.833 8.760 12.383 9.384 4.583 10.350 7.511 9.705 11.341 10.270 9.383 12.352 20.039 c 13.680 6.178 7.190
Q (12)
(0.156) (0.723) (0.415) (0.670) (0.970) (0.585) (0.822) (0.642) (0.500) (0.592) (0.670) (0.418) (0.066) (0.322) (0.907) (0.845)

23
7.383 3.713 18.894 c 8.117 8.144 5.426 8.099 5.282 7.532 5.161 12.838 7.290 7.271 8.802 24.298 b 5.675
Q2 (12)
(0.831) (0.988) (0.091) (0.776) (0.774) (0.942) (0.777) (0.948) (0.821) (0.952) (0.381) (0.838) (0.839) (0.720) (0.019) (0.932)
Notes: The number of lags for VAR was decided using SIC (Schwartz information criterion) and AIC (Akaike information criterion) criteria. JB, Q(12) and Q2 (12) indicate the empirical
statistics of the Jarque–Bera test for normality, while Ljung–Box Q statistics of order 12 for autocorrelation were applied to the standardized residuals and squared standardized residuals,
respectively. USA, United States of America; IND, India; INDO, Indonesia; KOR, South Korea; MYS, Malaysia; PAK, Pakistan; PHL, the Philippines; TAIW, Taiwan; THA, Thailand.
Values in parentheses are the p-values. a , b , c indicate the statistical significance at 1%, 5%, and 10%, respectively.
JRFM 2020, 13, 226

The return spillover from the USA to all Asian Stock markets was significant during the Chinese
crisis. This implies that US stock market prices played an important role in predicting the prices of
all Asian stock markets during the Chinese stock market crash. Moreover, the return spillover from
Asia to the US stock market was insignificant. The coefficient of past shock was insignificant in the
majority of Asian markets except for in India, Korea, and Malaysia. The sensitivity of past own shocks
of the USA was insignificant in most of the cases. In addition, the coefficient of past own volatility
significantly affected the conditional volatility of all Asian markets.
The conditional volatility of the majority of Asian stock markets (except Malaysia and Thailand)
was not significantly affected by the shocks in the US stock market. In addition, past shocks in most of
the Asian stock markets (Except in India and Indonesia) did not influence the conditional volatility of
the US stock market. The volatility transmission from the USA to most of the Asian stock markets
(except Malaysia, Pakistan, and Thailand) was found to be insignificant during the Chinese Crisis.
On the other hand, volatility spillover from most of the Asian stock markets to the USA stock market
was evidently insignificant.
The asymmetric coefficients of all Asian stock markets (except Korea and Malaysia) were significant
and positive, showing that negative news from the US stock market has a greater ability to increase
the volatility of all Asian Stock markets as compared to positive news. However, the asymmetric
coefficient of the US stock market is significant and positive. Constant conditional correlation was
positively significant for all pairs of stock markets. However, CCC was weak in the majority of pairs.

4.2.6. Stock Market Linkages between China and Asia from the Chinese Stock Market Crash
Table 7 reports the return and volatility spillover between Chinese and Asian stock markets during
the Chinese stock market crash. There is significant evidence that lagged returns influence the current
stock returns of Asian Stock markets (Except in Korea, the Philippines, and Taiwan). This shows the
short-term predictability in stock price changes in the Asian stock markets. Moreover, Chinese stock
market returns were not affected by their lags during the Chinese stock market crash.
The return spillover was found to be insignificant from China to all Asian markets. However,
the return spillover was found to be insignificant from the majority of Asian markets to the Chinese
market, except for India and Taiwan, during the Chinese stock market crash. The coefficient of past
own shock did not significantly influence the conditional variance of the most of the Asian stock
markets except in India, Malaysia, and Thailand. Moreover, the sensitivity to past own shock of the
Chinese stock market was insignificant during the Chinese crash. However, the sensitivity of past own
volatility was found to be significant for all Asian stock markets.
The conditional volatility of India, Indonesia, Taiwan, and Thailand was significantly affected by
the shocks in the Chinese stock market. However, the shocks in the majority of Asian stock markets
(except India and the Philippines) did not influence the Chinese stock market. The past volatility
of China significantly impacted the conditional volatility of the stock markets of India, Indonesia,
Taiwan, and Thailand. However, volatility spillover was not found from most of the Asian stock
markets (except India, Taiwan, and Thailand) to the Chinese stock market during the Chinese stock
market crash.
The asymmetric coefficients of all Asian stock markets (except Malaysia and the Philippines) were
significant and positive, showing that negative news of the US stock market has a greater ability to
increase the volatility of Asian stock markets as compared to positive news. Asymmetric coefficients
of China were significant and positive in all pairs, demonstrating that negative news for any Asian
markets except for India had a greater ability to increase the volatility of Chinese stock markets as
compared to positive news during the Chinese crash. Constant conditional correlation was positively
significant for all pairs of stock markets, but CCC was weak in the majority of pairs.

24
 Table 7. Estimates of bivariate VAR−AGARCH for Chinese and Asian stock markets during the Chinese Stock Market Crash.

IND CHN INDO CHN KOR CHN MYS CHN PAK CHN PHL CHN TAIW CHN THA CHN
Panel A: Mean Equation
3.32 × 5.74 × 1.10 × 1.39 × 3.28 × 8.49 × −1.35 × 9.62 × 4.13 × 1.16 × 5.57 × 1.07 × 2.52 × 9.50 × 1.73 × 1.56 ×
Constant 10−4 10−5 10−5 10−4 10−4 10−5 10−5 10−5 10−4 10−4 10−5 10−4 10−4 10−5 10−4 10−4
JRFM 2020, 13, 226

(0.168) (0.835) (0.964) (0.615) (0.147) (0.715) (0.933) (0.715) (0.145) (0.663) (0.875) (0.699) (0.310) (0.714) (0.359) (0.561)
−1.15 ×
0.167 a 0.106 b 0.144 a −0.034 0.084 −0.014 0.125 a 0.062 0.288 a −0.012 0.056 0.086 0.106 b 0.155 a 0.021 c
rot−1 10−4
(0.000) (0.022) (0.000) (0.500) (0.350) (0.715) (0.000) (0.322) (0.000) (0.660) (0.141) (0.280) (0.014) (0.000) (0.620)
(0.997)
5.37 × −8.57 ×
−0.029 0.037 −0.012 0.069 −0.019 0.062 c 0.021 0.053 0.014 0.062 0.033 0.062 0.021 0.070 b
rct−1 10−3 10−4
(0.185) (0.356) (0.550) (0.330) (0.361) (0.077) (0.222) (0.181) (0.452) (0.120) (0.274) (0.113) (0.580) (0.052)
(0.836) (0.965)
Panel B: Variance Equation
3.77 × 1.79 × 4.23 × −4.44 × 3.43 × 9.89 × 1.42 × −2.42 × 4.07 × 8.60 × 6.97 × 1.31 × 9.22 × 3.22 × 2.64 × −8.98 ×
Constant 10−6 a 10−6 b 10−6 c 10−8 10−6 10−7 10−6 10−7 10−6 a 10−7 a 10−6 b 10−6 10−6 a 10−6 b 10−6 a 10−7 b
(0.000) (0.012) (0.010) (0.930) (0.139) (0.425) (0.126) (0.768) (0.000) (0.005) (0.032) (0.259) (0.000) (0.017) (0.001) (0.031)
−5 .50× 5.85 × 5.44 × −8.63 × −2.66 × 5.81 × −2.55 ×
 2 −0.063 a 0.052 0.024 0.102 a −0.013 0.058 0.018 b −0.028 −0.039 b
eot−1 10−3 a 10−4 10−3 10−4 10−5 10−3 10−3
(0.000) (0.119) (0.413) (0.003) (0.560) (0.144) (0.022) (0.209) (0.021)
(0.002) (0.901) (0.107) (0.677) (0.986) (0.347) (0.364)

 2 8.96 × 9.71 × 5.15 × 8.77 ×

0.070 a 0.035 b −0.018 a 0.012 0.023 −0.013 0.021 0.020 0.062 a 0.027 −0.033 a 0.017

25
ect−1 10−3 10−3 10−3 10−3
(0.000) (0.031) (0.007) (0.567) (0.290) (0.780) (0.137) (0.342) (0.008) (0.176) (0.000) (0.355)
(0.618) (0.611) (0.180) (0.656)
8.85 × −2.40 × 9.26 × −1.18 × −9.03 ×
0.880 a 0.016 a 0.789 a 0.879 a 0.768 a 0.841 a 0.824 a 0.659 a 0.021 b 0.818 a 0.015 b
hot−1 10−3 10−3 10−3 10−3 10−3
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.038) (0.000) (0.028)
(0.176) (0.665) (0.283) (0.517) (0.137)
−6.00 ×
−0.096 a 0.945 a 0.039 b 0.940 a −0.014 0.944 a 0.100 0.934 a 0.949 a −0.026 0.943 a −0.118 b 0.944 a 0.112 a 0.933 a
hct−1 10−3
(0.000) (0.000) (0.025) (0.000) (0.742) (0.000) (0.343) (0.000) (0.000) (0.341) (0.000) (0.017) (0.000) (0.000) (0.000)
(0.109)
0.171 a 0.029 0.163 a 0.059 b 0.044 c 0.045 c 0.087 0.065 b 0.263 a 0.058 c 0.061 0.048 b 0.276 a 0.047 b 0.229 a 0.049 b
Asymmetry
(0.000) (0.145) (0.000) (0.021) (0.077) (0.058) (0.112) (0.015) (0.000) (0.010) (0.210) (0.045) (0.000) (0.040) (0.000) (0.029)
Panel C: Constant Conditional Correlation
0.204 a 0.151 a 0.282 a 0.166 a 0.109 a 0.179 a 0.319 a 0.202 a
p0,c
(0.000) (0.000) (0.000) (0.000) (0.003) (0.000) (0.000) (0.000)
Panel D: Diagnostic Tests
LogL 5216.702 5157.83 5258.48 5520.00 5143.34 5026.77 5246.59 5354.32
AIC −12.108 −12.093 −12.437 −12.929 −11.856 −11.866 −12.268 −12.364
SIC −11.814 −11.799 −12.143 −12.635 −11.562 −11.573 −11.975 −12.070
 Table 7. Cont.

IND CHN INDO CHN KOR CHN MYS CHN PAK CHN PHL CHN TAIW CHN THA CHN
245.340 a 365.250 a 212.590 a 369.930 a 325.490 a 305.710 a 214.270 a 342.420 a 156.230 a 422.300 a 216.090 a 338.660 a 301.330 a 269.370 a 227.030 a 329.200 a
JB
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
31.405 a 16.802 10.948 15.522 27.305 a 16.590 15.069 16.940 46.439 a 20.284 c 26.154 c 15.115 13.912 11.123 24.373 b 15.957
JRFM 2020, 13, 226

Q (12)
(0.002) (0.157) (0.533) (0.214) (0.007) (0.166) (0.238) (0.152) (0.000) (0.062) (0.010) (0.235) (0.306) (0.518) (0.018) (0.193)
55.871 a 17.976 47.212 a 21.484 b 48.551 a 17.416 70.480 a 21.463 b 70.632 a 30.028 a 50.180 a 23.783 b 44.260 a 16.310 88.957 a 31.358 a
Q2 (12)
(0.000) (0.116) (0.000) (0.044) (0.000) (0.135) (0.000) (0.044) (0.000) (0.003) (0.000) (0.022) (0.000) (0.177) (0.000) (0.002)
Notes: The number of lags for VAR was decided using SIC (Schwartz information criterion) and AIC (Akaike information criterion) criteria. JB, Q(12) and Q2 (12) indicate the empirical
statistics of the Jarque–Bera test for normality, while Ljung–Box Q statistics of order 12 for autocorrelation were applied to the standardized residuals and squared standardized residuals,
respectively. USA, United States of America; IND, India; INDO, Indonesia; KOR, South Korea; MYS, Malaysia; PAK, Pakistan; PHL, the Philippines; TAIW, Taiwan; THA, Thailand.
Values in parentheses are the p-values. a , b , c indicate the statistical significance at 1%, 5%, and 10%, respectively.

26
JRFM 2020, 13, 226

4.3. Optimal Weights and Hedge Ratio Portfolio Implications

Table 8 indicates the optimal weights and hedge ratios for the pairs of Asia-US stock portfolios
during the full sample period, US financial crisis, and the Chinese stock market crash.7 The range
of optimal weights is 0.37 for IND/USA to 0.68 for MYS/USA during the period of the full sample,
indicating that for a $1 India-USA portfolio, 37 cents should be invested in Indian stocks and the
remaining 63 cents in the US stock market. The average optimal portfolio weights vary from 0.38 for
IND/USA to 0.80 for MYS/USA during the US financial crisis and range from 0.37 for PHL/USA to
0.69 for MYS/USA during the Chinese stock market crash. Overall, the optimal weights of US stock in
Asia-USA portfolios are higher during the Chinese stock market crash compared to the US financial
crisis. This implies that investors should have maintained more US stocks in their portfolio of Asia-USA
during the Chinese stock market crash compared to the Asian stocks during the US financial crisis.
Table 9 presents the optimal weights and hedge ratios for the pairs of Asia-China stock portfolio
during the full sample period, US financial crisis, and the Chinese stock market crash.8 The range of
optimal weights is from 0.56 for IND/CHN, KOR/CHN, PHL/CHN to 0.81 for MYS/CHN during the
full sample period. The average optimal portfolio weights vary from 0.53 for IND/CHN to 0.90 for
MYS/CHN during the US financial crisis and range from 0.52 for PHL/CHN to 0.82 for MYS/CHN
during the Chinese stock market crash. Overall, for the majority of Asia-China portfolios, the optimal
weights of Chinese stocks were almost equal or higher during the Chinese stock market crash and
the US financial crisis. This suggests that portfolio managers and investors should have maintained
almost the same investment in Chinese stock in their majority of the portfolio of Asia-China during
both the Chinese crash and the US financial crisis.
Table 8 presents the optimal hedge ratios for the pairs of Asia-USA stock portfolio during the full
sample period, US financial crisis, and the Chinese stock market crash. Regarding the hedge ratio,
the range of average hedge ratio is 0.04 for PAK/USA to 0.27 for IND/USA during the period of the full
sample, showing that a long position of $1 in Pakistani stocks can be hedged for a short position of
4 cents in US stocks. During the US financial crisis, the average optimal hedge ratios varied from 0.08
for PAK/USA to 0.36 for IND/USA. The average optimal hedge ratio ranged from 0.09 for PAK/USA
to 0.36 for IND/USA during the Chinese stock market crash. For the majority of pairs of Asia-USA,
the hedge ratios were lower in the US financial crisis compared with the Chinese stock market crash.
This suggests that few US stocks were required to minimize the risk of Asian stock investors during
the US financial crisis as compared to during the Chinese crash.
Table 9 provides the optimal hedge ratios for the pairs of a Asia-China stock portfolio during the
full sample period, US financial crisis, and the Chinese stock market crash. The range of average hedge
ratio is 0.04 for PAK/CHN to 0.21 for KOR/CHN during the period of the full sample. During the US
financial crisis, the average optimal hedge ratios varied from 0.03 for PAK/CHN to 0.32 for KOR/CHN.
The average optimal hedge ratio ranged from 0.09 for MYS/CHN to 0.26 for TAIW/CHN during the
Chinese stock market crash. Overall, for the Asia-China pairs, the hedge ratio was lower during the
Chinese stock market crash compared to the hedge ratios in the US financial crisis. This implies that
fewer Chinese stocks were needed to minimize the risk for Asian stock investors during the Chinese
stock market crash as compared to during the US crisis.

7 We calculated the optimal weights by using both VAR-GARCH and VAR-AGARCH models, but we reported the optimal
weights only from the VAR-AGARCH model for the purpose of brevity.
8 We calculated the optimal weights by using both VAR-GARCH and VAR-AGARCH models, but we reported the optimal
weights only from the VAR-AGARCH model for the purpose of brevity.

27
 Table 8. Optimal Weights and Hedge Ratios for Asia/USA pairs.

IND/USA INDO/USA KOR/USA MYS/USA PAK/USA PHL/USA TAIW/USA THA/USA

Full Sample Period
wSU
t
0.37 0.41 0.40 0.68 0.41 0.42 0.43 0.41
JRFM 2020, 13, 226

βSU
t
0.27 0.13 0.24 0.06 0.04 0.06 0.17 0.17
US Financial Crisis
wSU
t
0.38 0.51 0.54 0.80 0.52 0.57 0.52 0.52
βSU
t
0.36 0.18 0.21 0.11 0.08 0.09 0.15 0.22
Chinese Stock Market Crash
wSU
t
0.46 0.44 0.51 0.69 0.44 0.37 0.49 0.53
βSO
t
0.36 0.15 0.29 0.11 0.09 0.14 0.22 0.22
Note: wSU SU
t and βt refer to the optimal weights and hedge ratios, respectively.

Table 9. Optimal Weights and Hedge Ratios for Asia/China pairs.

IND/CHN INDO/CHN KOR/CHN MYS/CHN PAK/CHN PHL/CHN TAIW/CHN THA/CHN

Full Sample Period

28
wSC
t
0.56 0.57 0.56 0.81 0.57 0.56 0.59 0.59
βSC
t
0.15 0.15 0.21 0.09 0.04 0.12 0.20 0.13
US Financial Crisis
wSC
t
0.53 0.63 0.68 0.90 0.64 0.64 0.66 0.67
βSC
t
0.31 0.24 0.32 0.13 0.03 0.22 0.30 0.19
Chinese Stock Market Crash
wSC
t
0.65 0.61 0.66 0.82 0.58 0.52 0.66 0.73
βSC
t
0.17 0.13 0.23 0.09 0.11 0.18 0.26 0.14
Note: wSC SC
t and βt refer to the optimal weights and hedge ratios, respectively.
JRFM 2020, 13, 226

5. Conclusions
In this paper, we extend the previous work by examining the return and volatility transmissions
from the US and China to the eight emerging Asian stock markets including India, Indonesia, Korea,
Malaysia, Pakistan, the Philippines, Taiwan, and Thailand during the Chinese stock market crash by
using the VAR-AGARCH model. Moreover, we also examine the spillovers during the full sample
period and the 2008 US financial crisis to provide comparative insights to investors about whether the
impact of the Chinese crash on equity market spillovers is different from the crashes in other sample
periods. Lastly, we also estimate the optimal weights and hedge ratios during the full period and
all sub-periods.
Our comprehensive analysis reveals that both return and volatility spillover vary across different
pairs of stock markets and during financial crises. The findings of return spillover indicate a significant
spillover from the USA to Asian stock markets during the full sample period, the US financial crisis,
and the Chinese stock market crash. This implies that US stock market prices play an important
role in predicting the prices of the majority of Asian stock markets during the full period and all the
sub-periods. However, the return spillover is not significant from China to emerging Asian stock
markets during the US financial crisis and the Chinese stock market crash, implying that Chinese stock
prices cannot be used for predicting the prices of the majority of Asian stock markets during any of the
crisis periods in our study.
Our volatility spillover analysis reveals that the volatility was transmitted from the US to the
majority of Asian markets during the full sample period and the Chinese stock market crash, but such
a conclusion cannot be drawn for during the US financial crisis. This implies that portfolio investors of
Asian stock markets could have gotten the maximum benefits of diversification by holding US stocks
in their portfolio during the US financial crisis. However, the volatility spillover was transmitted from
China to a majority of Asian markets during the full sample period and US financial crisis, but such a
conclusion cannot be reached for during the Chinese crash, implying that portfolio investors of Asian
stock markets could have gotten the maximum benefits of diversification by holding Chinese stocks in
their portfolio during the Chinese stock market crash.
Based on optimal weights results, the weights of the US stocks in the Asia-USA portfolios are
higher during the Chinese crash compared to the US financial crisis, implying that investors should
keep more US stocks in their portfolio of the Asia-USA stocks during the Chinese stock market crash,
compared to the US financial crisis. For the majority of Asia-China portfolios, the optimal weights of
Chinese stocks were almost equal or higher during the Chinese stock market crash and the US financial
crisis. This suggests that portfolio managers and investors should have maintained almost the same
investment in the Chinese stocks in their portfolio of the Asia-China majority pairs during both the
Chinese crash and the US financial crisis.
Regarding the hedge ratios, for most of the Asia-USA pairs, the hedge ratios were smaller in the
US financial crisis than in the Chinese stock market crash. This suggests that few US stocks were
required to minimize the risk for Asian stock investors during the US financial crisis as compared to
during the Chinese crash. In contrast, for the Asia-China pairs, the hedge ratio was smaller during the
Chinese stock market crash compared to that in the US financial crisis. This implies that fewer Chinese
stocks were needed to minimize the risk for Asian stock investors during the Chinese stock market
crash as compared to the US crisis. Overall, our findings provide several important implications for
risk management and portfolio diversification that could be useful for investors and for policymakers
related to the US and Asian stock markets.

Author Contributions: Conceptualization, estimations, formal analysis, original draft preparation I.Y.;
Data collection, methodology writing, and review of draft S.A.; review, editing, and funding W.-K.W. All authors
have read and agreed to the published version of the manuscript.
Funding: This research has been supported by Air University, Asia University, China Medical University
Hospital, The Hang Seng University of Hong Kong, Research Grants Council (RGC) of Hong Kong (project

29
JRFM 2020, 13, 226

number 12500915), and the Ministry of Science and Technology (MOST, Project Numbers 106-2410-H-468-002 and
107-2410-H-468-002-MY3), Taiwan.
Acknowledgments: The first author gratefully acknowledge Arshad Hassan (department of Management
and Social Sciences, Capital University of Science and Technology, Islamabad) for their valuable suggestions.
The third author would like to thank Robert B. Miller and Howard E. Thompson for their continuous guidance
and encouragement.
Conflicts of Interest: The authors declare no conflict of interest.

References
Allen, Katie. 2015. Why is China’s stock market in crisis? The Guardian, July 8.
Allen, David E., Ron Amram, and Michael McAleer. 2013. Volatility spillovers from the Chinese stock market to
economic neighbours. Mathematics and Computers in Simulation 94: 238–57. [CrossRef]
Arouri, Mohamed El Hedi, Amine Lahiani, and Duc Khuong Nguyen. 2011. Return and volatility transmission
between world oil prices and stock markets of the GCC countries. Economic Modelling 28: 1815–25. [CrossRef]
Arouri, Mohamed El Hedi, Jamel Jouini, and Duc Khuong Nguyen. 2012. On the impacts of oil price fluctuations
on European equity markets: Volatility spillover and hedging effectiveness. Energy Economics 34: 611–17.
[CrossRef]
Arouri, Mohamed El Hedi, Amine Lahiani, and Duc Khuong Nguyen. 2015. World gold prices and stock returns
in China: Insights for hedging and diversification strategies. Economic Modelling 44: 273–282. [CrossRef]
Baele, Lieven. 2005. Volatility spillover effects in European equity markets. Journal of Financial and Quantitative
Analysis 40: 373–401. [CrossRef]
Beirne, John, Guglielmo Maria Caporale, Marianne Schulze-Ghattas, and Nicola Spagnolo. 2013. Volatility
spillovers and contagion from mature to emerging stock markets. Review of International Economics 21: 1060–75.
[CrossRef]
Bollerslev, T. 1990. Modelling the coherence in short-run nominal exchange rates: A multivariate generalized
ARCH model. The review of Economics and Statistics 72: 498–505. [CrossRef]
Bouri, Elie. 2015. Return and volatility linkages between oil prices and the Lebanese stock market in crisis periods.
Energy 89: 365–71. [CrossRef]
Cheung, Yan-Leung, Yin-Wong Cheung, and Chris C. Ng. 2007. East Asian equity markets, financial crises, and the
Japanese currency. Journal of the Japanese and International Economies 21: 138–52. [CrossRef]
Chien, Mei-Se, Chien-Chiang Lee, Te-Chung Hu, and Hui-Ting Hu. 2015. Dynamic Asian stock market convergence:
Evidence from dynamic cointegration analysis among China and ASEAN-5. Economic Modelling 51: 84–98.
[CrossRef]
Diebold, Francis X., and Kamil Yilmaz. 2009. Measuring financial asset return and volatility spillovers,
with application to global equity markets. The Economic Journal 119: 158–71. [CrossRef]
Dutta, Anupam, Elie Bouri, and Md Hasib Noor. 2018. Return and volatility linkages between CO2 emission and
clean energy stock prices. Energy 164: 803–10. [CrossRef]
Engle, Robert F., Giampiero M. Gallo, and Margherita Velucchi. 2012. Volatility spillovers in East Asian financial
markets: A MEM-based approach. Review of Economics and Statistics 94: 222–23. [CrossRef]
Forbes, K., and R. Rigobon. 2002. No contagion, only interdependence: Measuring stock market comovements.
Journal of Finance 57: 2223–61. [CrossRef]
Fung, Eric S., Kin Lam, Tak-Kuen Siu, and Wing-Keung Wong. 2011. A Pseudo-Bayesian Model for Stock Returns
in Financial Crises. Journal of Risk and Financial Management 4: 42–72. [CrossRef]
Gkillas, Konstantinos, Athanasios Tsagkanos, and Dimitrios I. Vortelinos. 2019. Integration and risk contagion
in financial crises: Evidence from international stock markets. Journal of Business Research 104: 350–65.
[CrossRef]
Glick, Reuven, and Michael Hutchison. 2013. China’s financial linkages with Asia and the global financial crisis.
Journal of International Money and Finance 39: 186–206. [CrossRef]
Guo, Xu, Michael McAleer, Wing-Keung Wong, and Lixing Zhu. 2017. A Bayesian approach to excess volatility,
short-term underreaction and long-term overreaction during financial crises. North American Journal of
Economics and Finance 42: 346–58. [CrossRef]

30
JRFM 2020, 13, 226

Hammoudeh, Shawkat M., Yuan Yuan, and Michael McAleer. 2009. Shock and volatility spillovers among equity
sectors of the Gulf Arab stock markets. The Quarterly Review of Economics and Finance 49: 829–42. [CrossRef]
He, Ling T. 2001. Time variation paths of international transmission of stock volatility—US vs. Hong Kong and
South Korea. Global Finance Journal 12: 79–93. [CrossRef]
Huang, Bwo-Nung, Chin-Wei Yang, and John Wei-Shan Hu. 2000. Causality and cointegration of stock markets
among the United States, Japan and the South China Growth Triangle. International Review of Financial
Analysis 9: 281–97. [CrossRef]
Huo, Rui, and Abdullahi D. Ahmed. 2017. Return and volatility spillovers effects: Evaluating the impact of
Shanghai-Hong Kong Stock Connect. Economic Modelling 61: 260–72. [CrossRef]
Huyghebaert, Nancy, and Lihong Wang. 2010. The co-movement of stock markets in East Asia: Did the 1997–1998
Asian financial crisis really strengthen stock market integration? China Economic Review 21: 98–112. [CrossRef]
In Francis, Sangbae Kim, Jai Hyung Yoon, and Christopher Viney. 2001. Dynamic interdependence and volatility
transmission of Asian stock markets: Evidence from the Asian crisis. International Review of Financial Analysis
10: 87–96.
Jin, Xiaoye. 2015. Volatility transmission and volatility impulse response functions among the Greater China stock
markets. Journal of Asian Economics 39: 43–58. [CrossRef]
Johansson, Anders C., and Christer Ljungwall. 2009. Spillover effects among the Greater China stock markets.
World Development 37: 839–51. [CrossRef]
Jouini, Jamel. 2013. Return and volatility interaction between oil prices and stock markets in Saudi Arabia. Journal
of Policy Modeling 35: 1124–44. [CrossRef]
Kim, Bong-Han, Hyeongwoo Kim, and Bong-Soo Lee. 2015. Spillover effects of the US financial crisis on financial
markets in emerging Asian countries. International Review of Economics & Finance 39: 192–210.
Kroner, Kenneth F., and Victor K. Ng. 1998. Modeling asymmetric comovements of asset returns. The Review of
Financial Studies 11: 817–44. [CrossRef]
Kroner, Kenneth F., and Jahangir Sultan. 1993. Time-varying distributions and dynamic hedging with foreign
currency futures. Journal of financial and Quantitative Analysis 28: 535–51. [CrossRef]
Lean, Hooi Hooi, Michael McAleer, and Wing-Keung Wong. 2015. Preferences of risk-averse and risk-seeking
investors for oil spot and futures before, during and after the Global Financial Crisis. International Review of
Economics and Finance 40: 204–16. [CrossRef]
Li, Yanan, and David E. Giles. 2015. Modelling volatility spillover effects between developed stock markets and
Asian emerging stock markets. International Journal of Finance & Economics 20: 155–77.
Lin, Boqiang, Presley K. Wesseh Jr., and Michael Owusu Appiah. 2014. Oil price fluctuation, volatility
spillover and the Ghanaian equity market: Implication for portfolio management and hedging effectiveness.
Energy Economics 42: 172–82. [CrossRef]
Ling, Shiqing, and Michael McAleer. 2003. Asymptotic theory for a vector ARMA-GARCH model. Econometric
Theory 19: 280–310. [CrossRef]
Liu, Y. Angela, and Ming-Shiun Pan. 1997. Mean and volatility spillover effects in the US and Pacific-Basin stock
markets. Multinational Finance Journal 1: 47–62. [CrossRef]
Malik, Farooq, and Shawkat Hammoudeh. 2007. Shock and volatility transmission in the oil, US and Gulf equity
markets. International Review of Economics & Finance 16: 357–68.
McAleer, Michael, Suhejla Hoti, and Felix Chan. 2009. Structure and asymptotic theory for multivariate asymmetric
conditional volatility. Econometric Reviews 28: 422–40. [CrossRef]
Michaud, Richard O., Gary L. Bergstrom, Ronald Frashure, and Brian Wolahany. 1996. Twenty years of international
equity investing. The Journal of Portfolio Management 23: 9–22. [CrossRef]
Miyakoshi, Tatsuyoshi. 2003. Spillovers of stock return volatility to Asian equity markets from Japan and the US.
Journal of International Financial Markets, Institutions and Money 13: 383–99. [CrossRef]
Sadorsky, Perry. 2012. Correlations and volatility spillovers between oil prices and the stock prices of clean energy
and technology companies. Energy Economics 34: 248–55. [CrossRef]
Syriopoulos, Theodore, Beljid Makram, and Adel Boubaker. 2015. Stock market volatility spillovers and portfolio
hedging: BRICS and the financial crisis. International Review of Financial Analysis 39: 7–18. [CrossRef]
The causes and consequences of China’s market crash. 2015. The Economist. Available online: https:
//[Link]/news/2015/08/24/the-causes-and-consequences-of-chinas-market-crash (accessed on
24 September 2020).

31
JRFM 2020, 13, 226

Vieito, João Paulo, Wing-Keung Wong, and Zhen-Zhen Zhu. 2015. Could The Global Financial Crisis Improve the
Performance of The G7 Stocks Markets? Applied Economics 48: 1066–80. [CrossRef]
Wan, Henry, Jr., and Wing-Keung Wong. 2001. Contagion or inductance? Crisis 1997 reconsidered. Japanese
Economic Review 52: 372–80. [CrossRef]
Xiong, Xiong, Yuxiang Bian, and Dehua Shen. 2018. The time-varying correlation between policy uncertainty
and stock returns: Evidence from China. Physica A: Statistical Mechanics and its Applications 499: 413–19.
[CrossRef]
Yang, Jian, James W. Kolari, and Insik Min. 2003. Stock market integration and financial crises: The case of Asia.
Applied Financial Economics 13: 477–86. [CrossRef]
Yilmaz, Kamil. 2010. Return and volatility spillovers among the East Asian equity markets. Journal of Asian
Economics 21: 304–13. [CrossRef]
Yousaf, Imran, and Arshad Hassan. 2019. Linkages between crude oil and emerging Asian stock markets:
New evidence from the Chinese stock market crash. Finance Research Letters 31: 207–217. [CrossRef]
Yousaf, Imran, Shoaib Ali, and Wing-Keung Wong. 2020. Return and Volatility Transmission between
World-Leading and Latin American Stock Markets: Portfolio Implications. Journal of Risk and Financial
Management 13: 148. [CrossRef]
Yousaf, Imran, and Shoaib Ali. 2020. Linkages between gold and emerging Asian stock markets: New evidence
from the Chinese stock market crash. Studies of Applied Economics 39. [CrossRef]
Zhou, Xiangyi, Weijin Zhang, and Jie Zhang. 2012. Volatility spillovers between the Chinese and world equity
markets. Pacific-Basin Finance Journal 20: 247–70. [CrossRef]
Zhu, Yanjian, Zhaoying Wu, Hua Zhang, and Jing Yu. 2017. Media sentiment, institutional investors and
probability of stock price crash: Evidence from Chinese stock markets. Accounting & Finance 57: 1635–70.
Zhu, Zhenzhen, Zhidong Bai, João Paulo Vieito, and Wing-Keung Wong. 2019. The Impact of the Global Financial
Crisis on the Efficiency of Latin American Stock Markets. Estudios de Economía 46: 5–30. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

32
 Journal of
Risk and Financial
Management

Article
Comparison of Financial Models for Stock
Price Prediction
Mohammad Rafiqul Islam and Nguyet Nguyen *
Department of Mathematics and Statistics, Youngstown State University, Youngstown, OH 44555, USA;
mislam02@[Link]
* Correspondence: ntnguyen01@[Link]; Tel.: +1-330-941-1805

Received: 30 July 2020; Accepted: 12 August 2020; Published: 14 August 2020

Abstract: Time series analysis of daily stock data and building predictive models are complicated.
This paper presents a comparative study for stock price prediction using three different methods,
namely autoregressive integrated moving average, artificial neural network, and stochastic
process-geometric Brownian motion. Each of the methods is used to build predictive models using
historical stock data collected from Yahoo Finance. Finally, output from each of the models is
compared to the actual stock price. Empirical results show that the conventional statistical model
and the stochastic model provide better approximation for next-day stock price prediction compared
to the neural network model.

Keywords: stock price prediction; auto-regressive integrated moving average; artificial neural
network; stochastic process-geometric Brownian motion; financial models

1. Introduction
Predicting modeling is one of the most popular mathematical methods in many fields such as
business, social science, engineering, and finance. In business, predictive modeling is also known as
predictive analytics. Among many, one of the most important applications of predictive modeling is
to predict the stock price. Modern predictive modeling can be categorized into two basic categories
such as statistical and soft computing techniques (Adebiyi et al. 2014). Autoregressive integrated
moving average (ARIMA) is one of the most popular and widely used statistical techniques for making
predictions using past observations (Meyler et al. 1998). In spite of having great popularity in making
predictions, this method has some limitations such as seasonality, non-stationarity, and other factors
(Tambi 2005). In contrast, as a machine learning method or soft computing technique, artificial neural
networks (ANNs) are one of the most accurate and widely used forecasting models for forecasting,
pattern recognition, and image processing (Khashei and Bijari 2010). Neural network models have
become more popular in forecasting over the last decade in business, economics, and finance
(Avcı 2007). According to Khashei and Bijari (2010), ANNs are distinguished and most effective
for predictive modeling because of their data-driven self-adaptive nature and they are universal
function approximators. The network can generalize, this means that once the network learns the data,
it can predict the unseen or future part of the data even if the given data is not smooth.
In addition to the above two methods, stochastic modeling that uses geometric Brownian motion
to predict the stock price is very popular. Brownian motion is a special type of motion of molecular
particles, first observed and described by the British-Scottish botanist in 1827. However, Louis Bachelier,
a French mathematician named this Brownian motion and proposed a model to predict stock prices
using Brownian motion in 1900. According to the geometric Brownian motion model, the returns
on a certain stock in successive, equal periods of time are independents and normally distributed
(Dmouj 2006). The equation of geometric Brownian motion has a constant volatility and drift, but in

JRFM 2020, 13, 181; doi:10.3390/jrfm13080181 33 [Link]/journal/jrfm

JRFM 2020, 13, 181

real-world scenario these are not constant and vary over time (Estember and Maraña 2016). Hence,
we consider time variant volatility and drift in our analysis.
There are many researchers using the three basic techniques: ARIMA, ANN, and stochastic
models to predict stock prices, which will be reviewed in the next section. However, in the literature,
there are no comparisons of using each of the three models to predict prices of one stock. Most of the
researchers compared performances of the two models ARIMA and ANN in stock price predictions,
but not all of the three methods. Therefore, in this paper, we build predictive models using all of the
above three modeling techniques and compare the models’ performance for stock price predictions,
which are discussed in the subsequent sections. Section 2 represents the literature review and related
works. In Section 3, we describe the general theories for each of the methods and then build the models
specifically for S & P 500 index. In Section 4, we describe the results from each of the three models and
model diagnostics. Section 5 contains the conclusions.

2. Literature Review
Prediction has long been a popular field in mathematical science, so there is plenty of related
research in the field. The first significant study of neural network models for stock price prediction
was done by (White 1988). His predictive model was based on IBM’s daily common stock and the
training predictions were very optimistic. Thereafter, a lot of research was performed to check the
neural networks’ accuracy of prediction to forecast the stock market. Hassan et al. (2007) proposed
a fusion model by combining the hidden Markov model (HMM), artificial neural network (ANN),
and genetic algorithms (GA) to forecast financial market behavior. They found that the performance of
the fusion tool is better than that of the basic model (Hassan and Nath 2005) where they used only
a single HMM. They also indicated that the performance of the fusion model is similar to that of the
ARIMA model. Zhang and Wu (2009) proposed an integrated model improved bacterial chemotaxis
optimization (IBCO) and back propagation artificial neural network to predict the S & P 500 index.
The IBCO based back propagation (or IBCO-BP) model is less computationally complex and has better
accuracy. Khashei and Bijari (2010) found that the performance of a neural network for some real time
series is not satisfactory. Hence, using ARIMA models, they suggested a novel hybrid type of artificial
neural network. The proposed model provided better predictions for three separate actual datasets
than just the neural artificial network model. Yao et al. (1999) compared the back propagation neural
network model and ARIMA model stock index forecasting. They found that the neural network results
in better accuracy in forecasting than the traditional ARIMA models. Adebiyi et al. (2014) compared
the forecasting performance by ARIMA and artificial neural network for stock data. They analyzed
daily stock prices for the Dell Incorporation and found a superiority of the neural network model over
the ARIMA model.
Merh et al. (2010) developed a three-layer feed-forward neural network model and auto-regressive
integrated moving average model to predict the future value of the stock price and revealed that
the ARIMA models perform better over ANN models. Lee et al. (2007) did a comparative study of
the forecasting performance by neural network models and the time series model (SARIMA) for the
Korean Stock Index data. They also found ARIMA models outperforming ANN models for the stock
price prediction. Agustini et al. (2018) used several stock indexes under the Jakarta Corporate Index
to build a predictive model with Brownian motion. They found a higher accuracy for prediction
with a mean absolute percentage error (MAPE) less than 20%. Rathnayaka et al. (2014) developed a
forecasting model using the geometric Brownian motion model and compared the predictions with
the results from the traditional time series model ARIMA. They used the Colombo Stock Exchange
(CSE), Sri Lanka data to build their models and found that the stochastic model prediction is more
significant than the traditional model.
The literature shows different opinions on the relative performances of the two of the three
models depending on data. Hence, further comparative studies of all the three models can assemble
a consistent methodology for stock price prediction. In this paper, we study the comparative

34
JRFM 2020, 13, 181

performances of the three models in predicting next-day stock prices for S & P 500 index data from
Yahoo Finance.

3. Methodology
The methodology section contains the basic four subsections. The first subsection describes the
data that were used to build the models. Then, each of the subsections describes general theories
and procedures to build the models and then how the models were fitted for a particular dataset.
The overall performance of each of the models was checked by the analysis of the residuals and
four different error measures, namely the absolute percentage error (APE), the average absolute
error (AAE), the average relative percentage error (ARPE), and the root-mean-square error (RMSE)
(Nguyet Nguyen and Wakefield 2018). The formula to calculate these errors are as follows:

N
r −r̄
APE = 1
r̄ ∑ iN i ,
i =1
N
r −r̄
AEE = ∑ iN i ,
i =1
N (1)
ri −r̄i
ARPE = 1
N ∑ N ,
i =1

N
ri −r̄i
RMSE = 1
N ∑ N .
i =1

3.1. Data
S & P 500 daily stock for the period 1 January 2015 to 31 December 2019 was used in this research.
We used the quantmod package (Ryan et al. 2020) in statistical software R, version 1.2.1335 to collect
the data directly from Yahoo Finance. Initially, the dataset contains six variables, namely daily Open,
High, Low, Close, Volume, and Adjusted Close price. Addition to the six variables, we created two more
variables, i.e., Average and Return. The Average variable is the average of daily Open, High, Low, and
Close price. All the predictive models were built to predict the Adjusted Close price for the next day
on the basis of the present day’s predictor variables. There were 63 trading days per quarter in 2019.
All the models were used to predict the next-day stock price for the last quarter of 2019. A total of 63
predictions were made.

3.2. Autoregressive Integrated Moving Average Process

3.2.1. ARIMA(p,d,q) Models

The time series analysis requires the stationarity of the data, meaning the statistical properties
such as mean, variance, and so on do not change over time. However, most of the real-world data,
like stock data, are non-stationary by nature. This non-stationarity can be taken care of by using the
Box–Jenkins ARIMA(p,d,q) approach (Makridakis and Hibon 1997). A time series {Yt } is said to follow
an Autoregessive Integrated Moving Average ARIMA(p,d,q) if the dth difference Wt = ∇d Yt is a stationary
ARMA( p, q) process (Cryer and Kellet 1991). A generalised ARI MA( p, d, q) model can be written as

Wt = φ1 Wt−1 + φ2 Wt−2 + . . . + φ p Wt− p + et + θ1 et−1 + θ2 et−2 + . . . + θq et−q ,

where φ1 , φ2 , . . . , φ p and θ1 , θ2 , . . . , θq are the autoregressive and moving average parameters,

respectively, and e s are the white noise. The autoregression AR( p), order p, and moving average
MA(q), order q, are determined from the analysis of the autocorrelation function. The number d
indicates the number of differences applied to the time series to remove the trend. The autoregressive
parameters φ s and moving average parameters θ  s are estimated from the model based on p, d, and q.

35
JRFM 2020, 13, 181

3.2.2. ARIMA(p,d,q) Model for S & P 500 Index

The adjusted close price of S & P 500 is a time series process { Xt } that has been analysed to build
the model. The process { Xt } is not a stationary process, if we see the following graphs.
Figure 1 shows an upward trend in the data. Inspecting the sample autocorrelation plot from
Figure 2, it is clear that the Auto Correlation Function (ACF) dies down very slowly and the Partial
Autocorrelation Function (PACF) cuts off at lag 1 with correlation one.

Figure 1. Time plot of the raw data.

Figure 2. Sample autocorrelation plot.

The slow dying-down nature of the ACF indicates that the process is non-volatile. That is,
the current value is relating with all the past values. These facts ensure the process is non stationary.
To make the process stationary, we transform the { Xt } series to the {Yt } = {log Xt } series and then to
a new series {Wt } = {∇2 Yt }, where Wt = Yt − 2Yt−1 + Yt−2 .
Figure 3 is the window plot of the second differenced log transformed stock price. From this
plot, the data looks stationary and randomized. Stationarity has been confirmed from the augmented
Dickey–Fuller test (Cheung and Lai 1995) with a p value of 0.01, where the alternative hypothesis was

36
JRFM 2020, 13, 181

stationary. Next, the autoregressive and moving average orders p and q were determined from the
PACF and ACF plot from Figure 4.

Figure 3. Window plot of the second differenced log transformed stock price.

Figure 4. ACF and PACF of the second differenced log transformed stock price.

The ACF cuts of at lag 1 which indicates that the process incorporates an MA process of order
q = 1 whereas the PACF gradually dies down. Therefore, the series Wt follows an MA(1) process or
the series Yt follows an I MA(2, 1) process i.e. Yt ∼ ARI MA(0, 2, 1). Other ARI MA( p, d, q) models
were also considered in this research, as shown in Table 1. The best model has been chosen based
on the Schwarz Bayesian Information Criterion (BIC) (Neath and Cavanaugh 2012) criteria, the more
negative, the more accurate model. The reason of not choosing Akaike Information Criterion (AIC) or
Bias Corrected Akaike Information Criterion (AICc) is that those models provide over-fitting and non
significant parameters.
From Table 1, the ARI MA(0, 2, 1) model has the most negative BIC value which fits the data most
perfectly. The Arima function with order ( p, d, q) = (0, 2, 1) was run in RStudio and the summary of
the model is displayed in Table 2.

37
JRFM 2020, 13, 181

Table 1. ARIMA (p,d,q) model comparison.

Model AIC BIC AICc

ARIMA(0,2,0) −7144.11 −7139.03 −7144.11
ARIMA(0,2,1) −7981.48 −7971.31 −7981.47
ARIMA(0,2,2) −7979.84 −7964.59 −7979.82
ARIMA(0,2,3) −7981.77 −7961.43 −7981.73
ARIMA(0,2,4) −7980.02 −7954.61 −7979.97
ARIMA(1,2,0) −7455.29 −7445.12 −7455.28
ARIMA(1,2,1) −7979.81 −7964.56 −7979.79
ARIMA(1,2,2) −7982.16 −7961.83 −7982.13
ARIMA(1,2,3) −7983.04 −7957.62 −7982.98
ARIMA(1,2,4) −7981.53 −7951.03 −7981.45
ARIMA(2,2,0) −7631.73 −7616.48 −7631.71
ARIMA(2,2,1) −7981.50 −7961.16 −7981.46
ARIMA(2,2,2) −7982.88 −7957.46 −7982.83
ARIMA(2,2,3) −7982.62 −7952.12 −7982.55
ARIMA(2,2,4) −7979.91 −7944.32 −7979.81
ARIMA(3,2,0) −7692.63 −7672.30 −7692.60
ARIMA(3,2,1) −7979.84 −7954.42 −7979.79
ARIMA(3,2,2) −7981.50 −7951.00 −7981.43
ARIMA(3,2,3) −7977.84 −7942.25 −7977.74
ARIMA(3,2,4) −7978.07 −7937.40 −7977.95
ARIMA(4,2,0) −7738.80 −7713.39 −7738.75
ARIMA(4,2,1) −7980.13 −7949.63 −7980.06
ARIMA(4,2,2) −7978.69 −7943.11 −7978.60
ARIMA(4,2,3) −7978.28 −7937.61 −7978.15
ARIMA(4,2,4) −7984.84 −7939.09 −7984.69

Table 2. ARIMA(0,2,1) model summary.

Model Arima(x = [Link], order = c(0, 2, 1))

MA(1) Coefficient −1.00
Standard Error 0.0027
Sigma-squared estimated as 0.0000718
Log likelihood 3992.74
AIC −7981.48
AICc −7981.47
BIC −7971.31
ME RMSE MAE MPE MAPE MASE ACF1
Training set error measures
0.00013 0.00846 0.00573 0.00220 0.10561 0.99516 −0.01682

The model in difference equation is given as

Wt = et + θet−1
∇2 Yt = et + θet−1
Yt − 2Yt−1 + Yt−2 = et + θet−1
Yt = 2Yt−1 − Yt−2 + et + θet−1 . (2)

Finally, substituting the MA parameter θ = −1 in Equation (2), the model for Yt = log Xt is
given as
Yt = 2Yt−1 − Yt−2 + et − et−1 . (3)

A fixed window of 1194 past observed stock prices have been used to predict each of the next-day
prices using the model in Equation (3). Hence, the training dataset moved and the end price of the

38
JRFM 2020, 13, 181

window was updated with the actual price. The results and diagnostics of this model are discussed in
Section 4.1.

3.3. Stochastic Model Geometric Brownian Motion

A process that generates some outcomes which are time-dependent but can not be said ahead of
time is known as a stochastic process. A stochastic process {W (t) : 0 ≤ t ≤ T } is a standard Brownian
motion on [0, T ] if

1. W (0) = 0
2. It has independent increments. That is, for any t1 , t2 , . . . , tn , W (t2 ) − W (t1 ), W (t3 ) −
W (t4 ) . . . , W (tn ) − W (tn− ) are independent random variables.
3. For every 0 ≤ s < t ≤ T, W (t) − W (s) ∼ N(0, t − s).

A stochastic process { X (t) : 0 < t < T } is said to be a general Brownian motion with a
X (t)−μt
drift parameter μ and diffusion coefficient σ2 if σ is a standard Brownian motion, written as
X (t) ∼ BM(μ, σ ).
2

The general Brownian motion still follows first two properties of the standard Brownian motion.
However, the third property is modified as X (t) − X (s) ∼ N(μ(t − s), σ2 (t − s)) for any 0 ≤ s < t < T.

3.3.1. Geometric Brownian Motion (GBM) Model

If X (t) ∼ BM(μ, σ2 ) then X (t) satisfies the stochastic differential equation (Yang and Aldous 2015)

dX (t) = μt + σdW (t), (4)

where, W (t) is the standard Brownian motion or Wiener process. If the stochastic process is defined as
X (t) = log S(t) then dS(t) = μS(t)dt + σS(t)dW (t) is the stochastic differential equation for the stock
price random process.
For a given time t > 0, the standard model for stock prediction can be given from the stochastic
differential equation by integration
 t  t
S ( t ) = S (0) + μ S(r )dr + σ S(r )dW (r ). (5)
0 0

A more explicit formula can be derived using Ito’s formula (Ševcovic et al. 2011) to the function
F (log S(t), t)
   
dF = ∂F ∂F 1 2 ∂2 F
∂t + μ ∂S(t) + 2 σ ∂2 S(t) dt +
σ ∂S∂F(t) dW (t),

which results
1 1 −1
d log S(t) = dS(t) + (dS(t))2
S(t) 2 S2 ( t )
1 −1
= μdt + σdW (t) + (μS(t)dt + σS(t)dW (t))2
2 S2 ( t )
1
= (μ − σ2 )dt + σdW (t).
2
For any time t > 0, the differential can be written as

1
log S(t) = log S(0) + (μ − σ2 )t + σW (t)
2
Or, S(t) = S(0)e(μ− 2 σ
1 2 ) t + σW ( t )
. (6)

39
JRFM 2020, 13, 181

3.3.2. Geometric Brownian Motion Model for S & P 500 Index: GBM(μ, σ2 ) Simulation
For a given time set, t0 = 0 < t1 < t2 < . . . < tn , the stock price S(t) at time t0 , t1 , . . . , tn can be
generated by √
S ( t i +1 ) = S ( t i ) e ( μ − 2 σ
1 2 )( t
i +1 − ti )+ σ (ti+1 −ti ) Zi+1
, (7)

where Z1 , Z2 , . . . Zn are independent and identically distributed standard normals and i = 0, (n − 1).
In our case, the time interval ti+1 − ti = 1 for all i = 0, (n − 1), since we are predicting the next-day
price. Hence, the model becomes

S ( t i +1 ) = S ( t i ) e ( μ − 2 σ
1 2 )+ σZ
i +1 . (8)

Using the model in Equation (8), we simulate a large number of prices, and from that we take
the average to predict the next-day price. For our data, this large number is 100, 000. A total of
63 predictions have been made using this model. A fixed window of 1194 past observed stock prices
have been used to predict each of the next-day prices. Hence, the training dataset moved, and the end
price of the window was updated with the actual price. The results and diagnostics of this model are
discussed in Section 4.2.

3.4. Artificial Neural Network

This section describes how an artificial neural network can be used to predict the stock price and
how to build a model based on the stock data for S & P 500 index.

3.4.1. Model Descriptions

Artificial neural network is one of the most popular machine learning techniques for nonlinear
approximations because of its ability to deal with a large number of functions with a high degree
of accuracy (Chen et al. 2003). The idea of ANN came from the structure of the animal brain,
more specifically, from the human neural system. It is based on the idea of how brain works, how the
neurons in the brain receive information from the input neurons, analyse it, and finally identify the
object or pattern. Fundamentally, the mechanism has three layers—input layer, hidden layers, and
output layer. Each layer consists of neurons or nodes. The hidden part may consist of many layers,
however, for the time series analysis and forecasting, the single hidden layer feed forward network is
the most widely used model structure (Zhang et al. 1998). A simple three layer neural network has the
following mathematical form
q p
Yt = W0 + ∑ Wj .g(W0,j + ∑ Wi,j .Yt−i ) + t , (9)
j =1 i =1

where, Wi,j and Wj for i = 1, 2, . . . , p, j = 1, 2, . . . , q are known as connection weights. The parameter
p and q are the number of input and output nodes respectively. The network involves an activation
function which plays a very important role because it converts the input signals to be used for the
neurons or nodes in the next layer, eventually the output neuron. The most widely used activation
functions are the logistic and hyperbolic functions (Khashei and Bijari 2010), which are shown in
Equations (10) and (11)
1
sig( x ) = (10)
1 − e− x
1 − e−2x
tan−1 ( x ) = . (11)
1 + e−2x

40
JRFM 2020, 13, 181

Most of the modelers prefer the hyperbolic tangent function as the activation functions because of
its faster convergence, and it makes the optimization easier. Hence, we used this activation function in
our model. There is no systematic rule of choosing the number of neurons or nodes, q in the hidden
layer (Khashei and Bijari 2010). In most of the cases it is data-dependent and chosen on the basis of
trial and error.

3.4.2. Artificial Neural Network for S & P 500 Index

The model proposed for S & P 500 in this section is a three layer model—input, hidden, and the
output layer. The input layer consists of a total of seven nodes which are daily Open, Close, High, Low,
Average, Volume, and Return. The variable Average is the average of daily Open, Close, High, and Low.
The Volume was converted to million units. Daily return was calculated by this formula rt = log SSt ,
t −1
where St is the adjusted closed price and day one return, r0 was considered zero. The output layer
has only one node that corresponds to the predicting variable Adjusted Close price. The number of
the nodes in the hidden layer was chosen based on the error measures in Equation (1) for different
combinations of the hidden nodes, which are displayed in Table 3. From this Table 3, we see that model
ANN(7-15-1) has the lowest APE, AAE, ARPE, and RMSE and the highest adjusted R2 value.

Table 3. Error measures for different network structures.

MODEL R2 APE AAE ARPE RMSE

7-2-1 0.3137 0.0231 7.0435 0.1118 0.3344
7-3-1 0.4158 0.0211 6.412 0.1018 0.319
7-4-1 0.0422 0.028 8.5128 0.1351 0.3676
7-5-1 0.0368 0.0292 8.8796 0.1409 0.3754
7-6-1 0.3716 0.0215 6.5496 0.104 0.3224
7-7-1 0.3431 0.0226 6.8687 0.109 0.3302
7-8-1 0.3907 0.0219 6.6781 0.106 0.3256
7-9-1 0.4036 0.0212 6.4713 0.1027 0.3205
7-10-1 0.4108 0.0214 6.5237 0.1036 0.3218
7-11-1 0.5155 0.0195 5.9284 0.0941 0.3068
7-12-1 0.5392 0.0187 5.6802 0.0902 0.3003
7-13-1 0.4777 0.0194 5.8945 0.0936 0.3059
7-14-1 0.4676 0.0196 5.9702 0.0948 0.3078
7-15-1 0.6216 0.0167 5.0928 0.0808 0.2843
7-16-1 0.5701 0.0182 5.5382 0.0879 0.2965

The original dataset had 1257 observations, but the dataset used in this method was modified
in this way—all the predictor and predicting variables have the same length of 1256, however,
the predictor variables started from day 1 to the 1256th day and the predicting variable day 2 to
the 1257th day. Then, the dataset was divided into two parts to run the model. The test dataset
contained the last 63 actual stock prices (adjusted close) which were compared to the predicted
prices. The best model was selected on the basis of the adjusted R2 and four error measures (Table 3).
The model architecture is shown in Figure 5 and the result of this model is discussed in Section 4.3.

41
JRFM 2020, 13, 181

Figure 5. Artificial neural network architecture.

4. Results
In this section, the result of from the above three models is discussed and a window of the
predicted and actual price is shown together with a graphical presentation. Finally, we assess how our
model is performing by model diagnostics.

4.1. Autoregressive Integrated Moving Average

4.1.1. ARIMA Model Result

The model ARI MA(0, 2, 1) in Equation (3) produces the prediction in a logarithmic scale, which
is then converted back to the original scale by the formula Predicted Price = e prediction . Total 63 trading
days have been predicted by the model and compared with the actual prices which has been shown in
Table 4 with individual prediction error calculated by the formula,

actual − predicted
error = × 100. (12)
actual
From Table 4, we see that the forecast errors are less than one dollar for the daily period from
12 November, 2019 to 30 December 2019, with the relative errors within the range of 0.00003 to 0.00292.
Figure 6 shows the graphical representation of the actual and predicted stock price by the model.
The black line represents the actual stock price and the red line represents the predicted stock price for
S & P 500. Figure 6 also shows that the ARIMA (0,2,1) predicted prices follow closely to the trend of
the actual prices.

42
JRFM 2020, 13, 181

Table 4. Prediction by ARIMA(0,2,1) model.

Date Actual Predicted Error

11-12-2019 305.69 305.17 0.17
11-13-2019 305.79 305.81 0.01
11-14-2019 306.23 305.91 0.10
11-15-2019 308.45 306.35 0.68
11-18-2019 308.68 308.57 0.04
11-19-2019 308.59 308.80 0.07
11-20-2019 307.44 308.71 0.41
11-21-2019 306.95 307.56 0.20
11-22-2019 307.63 307.07 0.18
11-25-2019 310.01 307.75 0.73
11-26-2019 310.72 310.14 0.19
11-27-2019 312.10 310.85 0.40
11-29-2019 310.94 312.23 0.41
12-2-2019 308.30 311.07 0.90
12-3-2019 306.23 308.43 0.72
12-4-2019 308.12 306.36 0.57
12-5-2019 308.68 308.25 0.14
12-6-2019 311.50 308.81 0.86
12-9-2019 310.52 311.64 0.36
12-10-2019 310.17 310.65 0.15
12-11-2019 311.05 310.30 0.24
12-12-2019 313.73 311.18 0.81
12-13-2019 313.92 313.86 0.02
12-16-2019 316.08 314.05 0.64
12-17-2019 316.15 316.21 0.02
12-18-2019 316.17 316.29 0.04
12-19-2019 317.46 316.31 0.36
12-20-2019 318.86 317.60 0.40
12-23-2019 319.34 319.00 0.11
12-24-2019 319.35 319.48 0.04
12-26-2019 321.05 319.49 0.49
12-27-2019 320.97 321.20 0.07
12-30-2019 319.20 321.12 0.60

4.1.2. ARIMA Model Diagnostics

The performance of the ARIMA(0,2,1) model was assessed by the analysis of the four error
measures state in Equation (1) and the residuals plot, which is depicted in Figure 7 and those four
error measures are tabulated in Table 5.

Table 5. Prediction error by ARIMA(0,2,1) model.

MODEL APE AAE ARPE RMSE

ARIMA(0,2,1) 0.0044 1.3651 0.0217 0.1472

From the results in Table 5, we see that all the error measures are comparatively very low to the
actual prices, this indicates that the model is performing better in its prediction. From Figure 7 it is
clear that the residuals do not follow any special pattern, they are a randomized plot. Correlations in
the few lags are significant. Overall, the model fits very well to predict the stock price.

43
JRFM 2020, 13, 181

Figure 6. ARIMA (0,2,1) model prediction.

Figure 7. ARIMA(0,2,1) model residual analysis.

4.2. Stochastic Model

4.2.1. Stochastic Model Result

The model proposed in Equation (8) requires the calculation of 63 distinct values of the means μ
and standard deviations σ of the daily returns. Both of the parameters were calculated on the basis
of the same number of returns each time. Predicted values, actual values, and individual errors are
shown in Table 6, and the errors were calculated by the same formula (12) used in the previous model.
The results in Tables 5 and 6 are almost the same except at some points. Thus, Figures 7 and 8 are
almost identical.
Figure 8 displays the graphical representation of the actual and predicted stock prices from the
stochastic model. The black line represents the actual stock price and the green line represents the
predicted stock price for S & P 500 index.

44
JRFM 2020, 13, 181

Table 6. Prediction by geometric Brownian motion.

Date Actual Predicted Error

11-12-2019 305.69 305.17 0.17
11-13-2019 305.79 305.81 0.01
11-14-2019 306.23 305.91 0.10
11-15-2019 308.45 306.35 0.68
11-18-2019 308.68 308.59 0.03
11-19-2019 308.59 308.80 0.07
11-20-2019 307.44 308.71 0.41
11-21-2019 306.95 307.58 0.21
11-22-2019 307.63 307.06 0.19
11-25-2019 310.01 307.74 0.73
11-26-2019 310.72 310.15 0.18
11-27-2019 312.10 310.86 0.40
11-29-2019 310.94 312.23 0.41
12-2-2019 308.30 311.09 0.90
12-3-2019 306.23 308.43 0.72
12-4-2019 308.12 306.36 0.57
12-5-2019 308.68 308.25 0.14
12-6-2019 311.50 308.82 0.86
12-9-2019 310.52 311.63 0.36
12-10-2019 310.17 310.64 0.15
12-11-2019 311.05 310.30 0.24
12-12-2019 313.73 311.21 0.80
12-13-2019 313.92 313.88 0.01
12-16-2019 316.08 314.06 0.64
12-17-2019 316.15 316.21 0.02
12-18-2019 316.17 316.27 0.03
12-19-2019 317.46 316.31 0.36
12-20-2019 318.86 317.59 0.40
12-23-2019 319.34 319.00 0.11
12-24-2019 319.35 319.47 0.04
12-26-2019 321.05 319.51 0.48
12-27-2019 320.97 321.20 0.07
12-30-2019 319.20 321.11 0.60

Figure 8. Stochastic model geometric Brownian motion prediction.

4.2.2. Stochastic Model Diagnostics

The performance of the stochastic model was assessed by the analysis of the four error measures
stated in Equation (1) and the residual plot which is depicted in Figure 9 and the calculation of the
four different error measures, as shown in Table 7.

45
JRFM 2020, 13, 181

Table 7. Prediction error by geometric Brownian motion.

MODEL APE AAE ARPE RMSE

GBM 0.0044 1.3341 0.0212 0.1455

Figure 9. GBM model diagnostics.

The standardized residual plot is random and the mean passes through the zero line. A few of the
residuals at the lower end are outside of the band in the Q-Q plot of the residuals. Still, both of the
plots depict the approximate normal behavior of the residuals.

4.3. Artificial Neural Network

4.3.1. ANN(7-15-1) Results

Both of the training and test datasets were converted to normal and the prediction was converted
back to the original scale by inverse transformation. The model required approximately 7000 steps
with an error of 0.1972. Actual prices, predicted prices, and the corresponding errors are displayed in
Table 8.

Table 8. Prediction by ANN (7-15-1) model.

Date Actual Predicted Error

11-12-2019 305.69 302.13 1.17
11-13-2019 305.79 302.52 1.07
11-14-2019 306.23 302.26 1.30
11-15-2019 308.45 302.39 1.97
11-18-2019 308.68 303.56 1.66
11-19-2019 308.59 304.19 1.43
11-20-2019 307.44 303.54 1.27
11-21-2019 306.95 302.52 1.44
11-22-2019 307.63 302.86 1.55
11-25-2019 310.72 305.06 1.82
11-27-2019 312.10 305.81 2.02
11-29-2019 310.94 306.38 1.47
12-2-2019 308.30 306.02 0.74
12-3-2019 306.23 303.59 0.86
12-4-2019 308.12 301.63 2.11
12-5-2019 308.68 304.03 1.51

46
JRFM 2020, 13, 181

Table 8. Cont.

Date Actual Predicted Error

12-6-2019 311.50 304.28 2.32
12-9-2019 310.52 306.05 1.44
12-10-2019 310.17 306.07 1.33
12-11-2019 311.05 305.23 1.87
12-12-2019 313.73 305.55 2.61
12-13-2019 313.92 306.48 2.37
12-16-2019 316.08 306.73 2.96
12-17-2019 316.15 307.56 2.72
12-18-2019 316.17 308.05 2.57
12-19-2019 317.46 308.40 2.85
12-20-2019 318.86 308.04 3.39
12-23-2019 319.34 306.20 4.11
12-24-2019 319.35 308.94 3.26
12-26-2019 321.05 309.71 3.53
12-27-2019 320.97 310.31 3.32
12-30-2019 319.20 309.93 2.90

The predicted errors in Table 8 are much higher than those in Tables 5 and 6. Precise comparisons
of the three models are given in the next section. The graph associated with this result is displayed in
Figure 10. The black line represents the actual stock price and the blue line represents the predicted
stock price for the S & P 500 index. From the graph, it is clear that the model is working better at the
beginning of the prediction interval.

Figure 10. ANN(7-15-1) prediction.

4.3.2. ANN(7-15-1) Model Diagnostics

The performance of the neural network ANN(7-15-1) was assessed by the analysis of the four error
measures stated in Equation (1) and the standardized residuals plot, which is depicted in Figure 11,
and the calculation of the four different error measures are shown in Table 9.

Table 9. Prediction error by ANN(7-5-1) model.

MODEL APE AAE ARPE RMSE

ANN(7-15-1) 0.0167 5.09279 0.08084 0.28432

47
JRFM 2020, 13, 181

Figure 11. ANN(7-15-1) model diagnostics.

The standardized residual plot does not show normal behavior. The error increases in an
exponential shape as the predicting interval increases.

4.4. Comparison
In this section, the combined output from the three models above is discussed. Table 10 shows a
sample of the empirical results obtained from the models and Figure 12 displays the result graphically.

Table 10. Sample results from the models—ARIMA(0,2,1), GBM (μ, σ2 ), and ANN(7-15-1).

Date Actual ARIMA GBM ANN

11-12-2019 305.69 305.17 305.17 302.13
11-13-2019 305.79 305.81 305.81 302.52
11-14-2019 306.23 305.91 305.91 302.26
11-15-2019 308.45 306.35 306.35 302.39
11-18-2019 308.68 308.57 308.59 303.56
11-19-2019 308.59 308.80 308.80 304.19
11-20-2019 307.44 308.71 308.71 303.54
11-21-2019 306.95 307.56 307.58 302.52
11-22-2019 307.63 307.07 307.06 302.86
11-25-2019 310.01 307.75 307.74 303.50
11-26-2019 310.72 310.14 310.15 305.06
11-27-2019 312.10 310.85 310.86 305.81
11-29-2019 310.94 312.23 312.23 306.38
12-2-2019 308.30 311.07 311.09 306.02
12-3-2019 306.23 308.43 308.43 303.59
12-4-2019 308.12 306.36 306.36 301.63
12-5-2019 308.68 308.25 308.25 304.03
12-6-2019 311.50 308.81 308.82 304.28
12-9-2019 310.52 311.64 311.63 306.05
12-10-2019 310.17 310.65 310.64 306.04
12-11-2019 311.05 310.30 310.30 305.23
12-12-2019 313.73 311.18 311.21 305.55
12-13-2019 313.92 313.86 313.88 306.48
12-16-2019 316.08 314.05 314.06 306.73
12-17-2019 316.15 316.21 316.21 307.56

48
JRFM 2020, 13, 181

Table 10. Cont.

Date Actual ARIMA GBM ANN

12-18-2019 316.17 316.29 316.27 308.05
12-19-2019 317.46 316.31 316.31 308.40
12-20-2019 318.86 317.60 317.59 308.04
12-23-2019 319.34 319.00 319.00 306.20
12-24-2019 319.35 319.48 319.47 308.94
12-26-2019 321.05 319.49 319.51 309.71
12-27-2019 320.97 321.20 321.20 310.31
12-30-2019 319.20 321.12 321.11 309.93

From Figure 12 it clear that that ARIMA(0,2,1) model’s output and GBM model’s output are very
close, sometimes they coincide, whereas the output from the ANN(7-15-1) model gets far from the
actual points as time increases.

Figure 12. Prediction by all three models against the actual stock price.

Looking at the error measures in Table 11, it is clear that the ARIMA model and stochastic model
perform better than the neural network model for predicting the next-day stock price.

Table 11. Error measures comparison from the three models.

MODEL APE AAE ARPE RMSE

ARIMA(0,2,1) 0.00438 1.33476 0.02119 0.14556
GBM 0.00438 1.33426 0.02118 0.14553
ANN(7-15-1) 0.01672 5.09279 0.08084 0.28432

5. Conclusions
This study represents a comparative study of three financial models ARIMA, ANN, and
Geometric Brownian Motion to predict the next-day stock prices. Results obtained from the analysis
of the S & P 500 index show that the conventional statistical model ARIMA and the stochastic
model-geometric Brownian motion model perform better than the artificial neural network models
for short term next-day stock price prediction. The results are in contradiction with the results in
Khashei and Bijari (2010), which concluded that the ARIMA was no better than the ANN model in time
series predictions. However, their proposed hybrid ANN model outperformed the traditional ANN
and the ARIMA models. Furthermore, our results are similar to the conclusions in Merh et al. (2010)
and Lee et al. (2007) which stated that ARIMA models outperform ANN models for stock price
predictions. On the other hand, Rathnayaka et al. (2014) found that the stochastic model prediction is

49
JRFM 2020, 13, 181

more significant than the traditional ARIMA model. In fact, on the basis of our results, the ARIMA
model and the stochastic model produce almost the same results. Thus, for short term prediction
using the time series data, the ARIMA model and the stochastic model can be used interchangeably.
For the ANN models, further studies, hybridization of existing models, and adding more independent
variables can improve the neural network models in predicting stock prices. One model can work
better than other models with particular time series data. Therefore, researchers or investors should
examine some different models to predict the prices of each stock to find the best prediction model.

Author Contributions: M.R.I. setup and ran models, processed data, and wrote the first draft. N.N. introduced
the methodology, refined the manuscript, and supervised the project. All authors have read and agreed to the
published version of the manuscript.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflict of interest.

References
Adebiyi, Ayodele Ariyo, Aderemi Oluyinka Adewumi, and Charles Korede Ayo. 2014. Comparison of arima
and artificial neural networks models for stock price prediction. Journal of Applied Mathematics 2014: 614342.
[CrossRef]
Agustini, W. Farida, Ika Restu Affianti, and Endah R. M. Putri. 2018. Stock price prediction using geometric
brownian motion. Journal of Physics: Conference Series 974: 012047.
Avcı, Emin. 2007. Forecasting daily and sessional returns of the ise-100 index with neural network models. Dogus
Universitesi Dergisi 8: 128–42. [CrossRef]
Chen, An-Sing, Mark T. Leung, and Hazem Daouk. 2003. Application of neural networks to an emerging financial
market: forecasting and trading the taiwan stock index. Computers & Operations Research 30: 901–23.
Cheung, Yin-Wong, and Kon S. Lai. 1995. Lag order and critical values of the augmented dickey–fuller test.
Journal of Business & Economic Statistics 13: 277–80.
Cryer, Jonathan D., and Natalie Kellet. 1991. Time Series Analysis. Berlin and Heidelberg: Springer.
Dmouj, Abdelmoula. 2006. Stock Price Modelling: Theory and Practice. Masters’s thesis, Vrije Universiteit,
Amsterdam, The Netherlands.
Estember, Rene D., and Michael John R. Maraña. 2016. Forecasting of stock prices using brownian motion–monte
carlo simulation. Paper presented at the 2016 International Conference on Industrial Engineering and
Operations Management Kuala Lumpur, Kuala Lumpur, Malaysia, March 8–10, pp. 704–13.
Hassan, Md Rafiul, and Baikunth Nath. 2005. Stock market forecasting using hidden markov model: A new
approach. Paper presented at the International Conference on Intelligent Systems Design and Applications
(ISDA’05), Pretoria, South Africa, December 3–5, Piscataway: IEEE, pp. 192–96.
Hassan, Md Rafiul, Baikunth Nath, and Michael Kirley. 2007. A fusion model of hmm, ann and ga for stock
market forecasting. Expert Systems with Applications 33: 71–80. [CrossRef]
Khashei, Mehdi, and Mehdi Bijari. 2010. An artificial neural network (p, d, q) model for timeseries forecasting.
Expert Systems with Applications 37: 479–89. [CrossRef]
Lee, Kyungjoo, Sehwan Yoo, and John Jongdae. 2007. Neural network model versus sarima model in forecasting
korean stock price index (kospi). Issues in Information System 8: 372–8.
Makridakis, Spyros, and Michele Hibon. 1997. Arma models and the box–jenkins methodology. Journal of
Forecasting 16: 147–63. [CrossRef]
Merh, Nitin, Vinod P. Saxena, and Kamal Raj Pardasani. 2010. A comparison between hybrid approaches of ann
and arima for indian stock trend forecasting. Business Intelligence Journal 3: 23–43.
Meyler, Aidan, Geoff Kenny, and Terry Quinn. 1998. Forecasting Irish Inflation Using Arima Models. Dublin: Central
Bank of Ireland.
Neath, Andrew A., and Joseph E. Cavanaugh. 2012. The bayesian information criterion: Background, derivation,
and applications. Wiley Interdisciplinary Reviews: Computational Statistics 4: 199–203. [CrossRef]
Nguyet Nguyen, Dung Nguyen, and Thomas P. Wakefield. 2018. Using the hidden markov model to improve the
hull-white model for short rate. International Journal of Trade, Economics and Finance 9. [CrossRef] [CrossRef]

50
JRFM 2020, 13, 181

Rathnayaka, R. M. Kapila Tharanga, Wei Jianguo, and DMK N. Seneviratna. 2014. Geometric brownian motion
with ito’s lemma approach to evaluate market fluctuations: A case study on colombo stock exchange. Paper
presented at the 2014 International Conference on Behavioral, Economic, and Socio-Cultural Computing
(BESC2014), Shanghai, China, October 30–November 2, Piscataway: IEEE, pp. 1–6.
Ryan, Jeffrey A., Joshua M. Ulrich, Wouter Thielen, Paul Teetor, Steve Bronder, and Maintainer Joshua M. Ulrich. 2020.
Package ‘quantmod’. Available online: [Link]
(accessed on 14 August 2020).
Ševcovic, Daniel, B. Stehlıková, and K. Mikula. 2011. Analytical and Numerical Methods for Pricing Financial
Derivatives. Hauppauge: Nova Science.
Tambi, Mahesh Kumar. 2005. Forecasting exchange rate: A univariate out of sample approach. The IUP Journal of
Bank Management 4: 60–74.
White, Halbert. 1988. Economic prediction using neural networks: The case of ibm daily stock returns. Paper
presented at the IEEE 1988 International Conference on Neural Networks, San Diego, CA, USA, July 24–27,
vol. 2, pp. 451–58.
Yang, Zhijun, and D. Aldous. 2015. Geometric Brownian Motion Model in Financial Market. Berkeley: University of California.
Yao, Jingtao, Chew Lim Tan, and Hean-Lee Poh. 1999. Neural networks for technical analysis: A study on klci.
International Journal of Theoretical and Applied Finance 2: 221–41. [CrossRef]
Zhang, Guoqiang, B. Eddy Patuwo, and Michael Y. Hu. 1998. Forecasting with artificial neural networks: The state
of the art. International Journal of Forecasting 14: 35–62. [CrossRef]
Zhang, Yudong, and Lenan Wu. 2009. Stock market prediction of S & P 500 via combination of improved bco
approach and bp neural network. Expert Systems with Applications 36: 8849–54.

c 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

51
 Journal of
Risk and Financial
Management

Article
Return and Volatility Transmission between
World-Leading and Latin American Stock Markets:
Portfolio Implications
Imran Yousaf 1 , Shoaib Ali 1 and Wing-Keung Wong 2,3,4, *
1 Air University School of Management, Air University, Islamabad 44000, Pakistan;
imranyousaf_12@[Link] (I.Y.); shoaibali_12@[Link] (S.A.)
2 Department of Finance, Fintech Center, and Big Data Research Center, Asia University, Taichung 41354,
Taiwan
3 Department of Medical Research, China Medical University Hospital, Taichung 40402, Taiwan
4 Department of Economics and Finance, The Hang Seng University of Hong Kong, Hong Kong 999077, China
* Correspondence: wong@[Link]; Tel.: +886-4-2332-3456

Received: 20 May 2020; Accepted: 3 July 2020; Published: 8 July 2020

Abstract: This study uses the BEKK-GARCH model to examine the return-and-volatility spillover
between the world-leading markets (USA and China) and four emerging Latin American stock
markets over the global financial crisis of 2008 and the crash of the Chinese stock market of 2015.
Regarding return spillover, our findings reveal a unidirectional return transmission from Mexico to
the US stock market during the global financial crisis. During the crash of the Chinese stock market,
the return spillover is found to be unidirectional from the US to the Brazil, Chile, Mexico, and Peru
stock markets. Moreover, the results indicate a unidirectional return transmission from China to
the Brazil, Chile, Mexico, and Peru stock markets during the global financial crisis and the crash of
the Chinese stock market. Regarding volatility spillover, the results show the bidirectional volatility
transmission between the US and the stock markets of Chile and Mexico during the global financial
crisis. During the Chinese crash, the bidirectional volatility transmission is observed between the
US and Mexican stock markets. Furthermore, the volatility spillover is unidirectional from China to
the Brazil stock market during the global financial crisis. During the Chinese crash, the volatility
spillover is bidirectional between the China and Brazil stock markets. Lastly, a portfolio analysis
application has been conducted.

Keywords: return spillover; volatility spillover; optimal weights; hedge ratios; US financial crisis;
Chinese stock market crash

JEL Classification: G10; G11; G12; G15

1. Introduction
The information transmissions (return and volatility) across equity markets are of greater interest to
investors and policymakers with increased financial integration all over the world. For example, if asset
volatility is transferred from one market to another during turbulence or crisis, then portfolio managers
need to adjust their asset allocation (Bouri 2013; Syriopoulos et al. 2015; Yousaf and Hassan 2019) and
financial policymakers need to change their policies to reduce the contagion risk (Yang and Zhou 2017).
The linkages between equity markets, especially during a crisis, can also have important implications
for asset allocations, portfolio diversification, asset valuation, hedging, and risk management.
In the literature, several studies have examined the linkages between equity markets during
the Asian crisis of 1997 (In et al. 2001; Chen et al. 2002; Chancharoenchai and Dibooglu 2006;

JRFM 2020, 13, 148; doi:10.3390/jrfm13070148 53 [Link]/journal/jrfm

JRFM 2020, 13, 148

Li and Giles 2014; Gulzar et al. 2019) and the global financial crisis (Taşdemir and Yalama 2014;
Bekiros 2014; Mensi et al. 2016; Gamba-Santamaria et al. 2017). However, the linkages between equity
markets are rarely examined during the crash of the Chinese stock market in 2015. The Chinese stock
market crashed in 2015 (Han and Liang 2016; Ahmed and Huo 2019; Yousaf and Hassan 2019). The CSI
300 index had reached up to 5178 points until mid-June in 2015. Then, it took roller-coaster ride and
dropped up to 34% in just 20 days, also losing 1000 points within just one week. Around 50% of the
Chinese stocks lost more than half of their pre-crash market value. This crash adversely affected the
many other financial markets around the globe (Fang and Bessler 2017). Despite the importance of the
Chinese crash to international portfolio managers, only Ahmed and Huo (2019) examined the volatility
transmission between the Chinese and Asian stock markets during the crash of the Chinese stock
market in 2015. The empirical research remains surprisingly limited on the area of linkages between
equity markets during the crash of the Chinese stock market.
The US and China are the most significant trading partners of the emerging Latin American
economies. From 2000 to 2017, the trade volume of China (US) is increased by 21 (2.5)-fold with
emerging Latin American economies. The trade volume of leading economies grew at a different rate
with the emerging Latin American (LA) economies; thus, spillover can also be changed between the
China-LA and US-LA pairs during the last two decades. Johnson and Soenen (2003) also suggest that
trade increases the financial integration between countries’ stock markets. Previously, several studies
have examined the spillovers between the US and Latin American stock markets (Meric et al. 2001;
Arouri et al. 2015; Ben Rejeb and Arfaoui 2016; Cardona et al. 2017; Gamba-Santamaria et al. 2017;
Ramirez-Hassan and Pantoja 2018; Yousaf and Ahmed 2018; Fortunato et al. 2019; Coleman et al. 2018).
However, the linkages between the China and Latin American stock markets have not yet been
explored, especially during the global financial crisis and the crash of the Chinese stock market.
Based on the above-mentioned literature gaps, this study aims to examine the return and volatility
spillover between the world-leading (the US and China) and emerging Latin American stock markets
during the full sample period, the global financial crisis, and the crash of the Chinese stock market.
Additionally, this study estimates the optimal weights and hedge ratios during all the sample periods.
Our study makes the following contributions to the literature. First, regarding return spillover,
the findings reveal a unidirectional return transmission from Mexico to the US stock market during
the global financial crisis. During the crash of the Chinese stock market, the return spillover is found
to be unidirectional from the US to the Brazil, Chile, Mexico, and Peru stock markets. Moreover,
the results indicate a unidirectional return transmission from China to the Brazil, Chile, Mexico,
and Peru stock markets during the global financial crisis and the crash of the Chinese stock market.
Regarding volatility spillover, the results show the bidirectional volatility transmission between the
US and the stock markets of Chile and Mexico during the global financial crisis. During the Chinese
crash, a bidirectional volatility transmission is observed between the US and Mexican stock markets.
Furthermore, the volatility spillover is unidirectional from China to the Brazil stock market during the
global financial crisis. During the Chinese crash, the volatility spillover is bidirectional between the
China and Brazil stock markets.
The contributions of this study are four-fold. First, this study provides a comprehensive analysis
of spillover between the world-leading and emerging LA stock markets during the crash of the Chinese
stock market. Second, it contributes to the literature of the China-LA stock markets by examining the
spillovers during the global financial crisis. Lastly, the BEKK-GARCH model is applied to estimate the
spillovers, optimal weights, and hedge ratios, which provide better statistical properties compared to
the many other GARCH models. The rest of the paper is organized as follows: Section 2 provides
a review of the literature. The empirical method is described in Section 3. Section 4 consists of the
data and the preliminary analysis. The empirical results are reported in Section 5. Finally, Section 5
concludes the discussion.

54
JRFM 2020, 13, 148

2. Literature Review
Markowitz’s modern portfolio theory can describe the relationship between different stock markets
in order to build an optimum portfolio. The rationale behind this concept is to combine risky assets
with less risky or risk-free assets in the portfolio (Markovitz 1959). For example, the leading stock
market shows a higher volatility during the financial crisis, and as a result the portfolio investors need
to diversify their portfolios by investing in weakly integrated emerging stock markets. Therefore, an
analysis of risk transmission between different equity markets is essential for portfolio managers to
identify opportunities for portfolio diversification across markets and over time.
Over the past decade, there has been a growing body of literature examining the information
transmissions (return and volatility) between the US and LA stock markets during the crisis
and non-crisis periods. Meric et al. (2001) report significant co-movements between the US
and LA (Brazil, Argentina, Chile, and Mexico) stock markets during the period 1984–1995.
Fernández-Serrano and Sosvilla-Rivero (2003) report the cointegration across the US and LA equity
markets. Sharkasi et al. (2005) investigate the spillover across the US and Brazil stock markets.
They provide evidence of co-movements between the US and Brazil stock markets.
Diamandis (2009) investigates the linkages and common trends between the US and four Latin
American (Argentina, Brazil, Chile, and Mexico) stock markets. Because the four Latin American
countries initiated a phase of financial liberalization in the late 1980s and early 1990s, this study also
explores whether the removal of foreign-exchange controls had any effect on the potential linkages.
Firstly, this study finds that the US stock market is partially integrated with four LA stock markets.
Secondly, the five stock markets have four significant common permanent components/trends which
influence their system in the long run. Thirdly, the results indicate significant short-term deviations
from standard stochastic patterns during the 1994–1996 Mexican crisis and the 2001 financial crisis.
Beirne et al. (2013) use the tri-variate GARCH-BEKK model to estimate the volatility transmission
from mature markets to 41 emerging (including 8 Latin American) stock markets. The volatility
transmission is observed to be significant from many mature markets to the emerging stock markets.
Additionally, there is evidence of changes in the parameters of volatility spillovers during turbulent
or crisis periods. Graham et al. (2012) estimate the integration between the US and 22 emerging
equity markets and find evidence of strong co-movements across the US, Brazil, and Mexico equity
markets. Hwang (2014) examined the spillover between the US and LA equity markets during the
global financial crisis. The study found that the integration between the US and LA equity markets
became stronger during the global financial crisis.
Using the VAR-GARCH model, Arouri et al. (2015) estimate the return and volatility transmissions
between the US and LA (Brazil, Argentina, Mexico, Chile, and Columbia) stock markets from 1993
through to 2012. The return spillover is seen to be significant from the US to the Argentina, Mexico,
and Colombia stock markets. It also provides evidence of a volatility transmission from the US to a few
LA stock markets. Syriopoulos et al. (2015) use the VAR-GARCH model and find that the return and
volatility spillover is significant between the US and BRICS (Brazil, Russia, India, China, and South
Africa) equity markets (at the sectoral level). Mensi et al. (2016) reveal the strong dynamic correlation
between US and BRICS equity markets during the global financial crisis of 2008.
Ben Rejeb and Arfaoui (2016) examine the volatility transmission between developed (US and
Japan) and emerging (Latin American and Asian) stock markets using standard GARCH models and
a quantile regression approach. This study reveals a significant presence of volatility transmission
in these markets. The volatility transmission is seen to be closely associated with the crisis period
and geographical proximity. A lower and upper quantiles analysis shows that interdependence
between markets decreases during a bearish trend, while it increases during bullish markets. Using the
GARCH model, Bhuyan et al. (2016) observes return and volatility transmissions from the US to BRICS
stock markets.
Al Nasser and Hajilee (2016) provide evidence of short-run integration between developed (US,
UK, and Germany) and emerging stock markets (Brazil, Mexico, Russia, China, and Turkey). However,

55
JRFM 2020, 13, 148

in the long run, the cointegration is only found to be significant between Germany and emerging Asian
stock markets. Gamba-Santamaria et al. (2017) examine the directional volatility transmission between
the US and the four LA stock markets (Brazil, Chile, Mexico, and Columbia) using the framework of
Diebold and Yilmaz (2012). Brazil is found to be the net volatility transmitter for most of the sample
period, whereas Columbia, Chile, and Mexico are the net receivers of volatility. Moreover, the US stock
market is observed to be the net transmitter of volatility to the four LA stock markets. Besides this, the
magnitude of volatility transmission is increased from the US to LA stock markets during the global
financial crisis of 2008.1
Yousaf and Ahmed (2018) study the influence of the US and Brazil on the Mexico, Argentina,
Chile, and Peru stock markets by using GARCH in a mean approach. The study concludes that the
return effects are dominantly transmitted from the US to the Mexico, Argentina, Chile, and Peru stock
markets. Moreover, the volatility transmission is found to be dominant from Brazil to the Mexico,
Argentina, Chile, and Peru stock markets. Cardona et al. (2017) use the MGARCH-BEKK model to
estimate the volatility transmission between the US and the six LA stock markets (Brazil, Argentina,
Mexico, Chile, Peru, and Colombia). They report the significant volatility transmission from the US to
all LA stock markets. Moreover, only Brazil transmits volatility effects to the US stock market.
Ramirez-Hassan and Pantoja (2018) provide evidence of co-movements between the returns of
the US and six LA stock markets after the global financial crisis of 2008. Fortunato et al. (2019) provide
evidence of return transmission from the US to the Brazil, Chile, Columbia, Mexico, and Peru equity
markets. Coleman et al. (2018) find the co-movements between the US and LA (Brazil, Chile, Mexico,
Peru, Venezuela, and Argentina) stock markets. Su (2020) reports the dominant risk transmission from
the G7 (US, Japan, UK, Germany, France, Italy, and Canada) countries to the BRICS (Brazil, Russia,
India, China, and South Africa) stock markets.
However, fewer studies have examined the spillovers between the China and Latin American
stock markets during the crisis and non-crisis periods. Garza-García and Vera-Juárez (2010) study the
impact of US and Chinese macroeconomic variables on the stock markets of Brazil, Mexico, and Chile.
The macroeconomic variables (the US and Chinese) are observed to be integrated with the LA stock
markets. Additionally, the US macroeconomic variables Granger affect the Brazilian and Mexican stock
markets. On the other hand, the Chinese macroeconomic variables Granger affect the stock markets of
Mexico and Chile.
Horvath and Poldauf (2012) find that the Chinese stock market is weakly correlated with the Brazil,
Australia, Canada, Germany, Japan, Hong Kong, South Africa, Russia, US, and UK stock markets.
Sharma et al. (2013) apply the VAR model to examine the linkages between the BRICS (Brazil, Russia,
India, China, and South Africa) stock markets. This study finds a return transmission from Brazil
(India) to the Russia, India (Brazil), China, and South Africa equity markets. Moreover, the return
transmission is only observed from China to the Russian stock market. Bekiros (2014) looks at the
contagion effect between Brazil, Russia, India, and China by using several multivariate GARCH models.

1 Our study is different from the study of Gamba-Santamaria et al. (2017) in the following aspects. Gamba-Santamaria et al. (2017)
examine the volatility spillover between the US and four Latin American markets (Brazil, Chile, Mexico, and Columbia)
during the US financial crisis, whereas our study is examining the volatility as well as return spillover between the leading
(US and China) markets and four Latin American markets (Brazil, Chile, Mexico, and Peru) during the US financial
crisis and the crash of the Chinese stock market. More specifically, firstly our study examines the return as well as
volatility spillovers, whereas Gamba-Santamaria et al. (2017) examine the directional volatility spillovers. Second,
our study is examining the spillovers between two world-leading (the US and China) markets and four LA markets,
whereas Gamba-Santamaria et al. (2017) examine the spillovers between US and four LA markets. Third, our study is
focusing on the spillovers during the global financial crisis and the crash of the Chinese stock market in 2015, whereas
Gamba-Santamaria et al. (2017) examine the spillovers during the US financial crisis. Fourth, our study is using the
BEKK-GARCH model, whereas Gamba-Santamaria et al. (2017) employ the approach of Diebold and Yilmaz (2012). Lastly,
our full data sample is from January 2001 to May 2020, whereas Gamba-Santamaria et al. (2017) use the sample period from
January 2003 to January 2016. Apart from the differences, the study of Gamba-Santamaria et al. (2017) is very beneficial for
understanding the linkages among the US and LA stock markets.

56
JRFM 2020, 13, 148

This study concludes that there exists a higher integration between Brazil, Russia, India, and China
after the global financial crisis.
Ahmad and Sehgal (2015) estimate the volatility of the BRIICKS (Brazil, Russia, India, Indonesia,
China, South Korea, and South Africa) stock markets by using the Markov regime-switching (MS) in
the mean-variance model. It suggests that investors should allocate investment in the China, Russia,
and India emerging stock markets. While investigating the relationship between the Chinese and
foreign stock markets (US, Brazil, India, and Germany), Cao et al. (2017) reported a bi-directional
causality between the China and foreign stock markets. Previous work does not provide evidence of
return and volatility spillover between leading (US and China) and Latin American stock markets
during the global financial crisis and the crash of the Chinese stock market. Therefore, this study
addresses the above-mentioned literature gaps.

3. Data and Methodology

In this section, we will discuss the data and methodology used in our paper. We first discuss
the data.

3.1. Data
This study uses the daily data of benchmark stock indices of the US (S&P 500); China (SSE
Composite Index); and four emerging LA stock markets—namely, Brazil (IBOVESPA index), Chile
(IPSA index), Mexico (S&P/BMV IPC Index), and Peru (S&P/BVL Peru General TR PEN Index).
The data of stock indices are taken from the Data Stream database. The index is assumed to be the
same on non-trading days (holidays except weekends) as on the previous trading day, as suggested by
Malik and Hammoudeh (2007), and Ali et al. (2020).
This study uses the full sample period from 1 January 2000, to 29 May 2020, and studies the
following two sub-samples: the first sub-period from 1 August 2007, to 30 July 2010, presenting the
period with the US financial crisis; and the second sub-period from 1 June 2015, to 31 May 2018,
presenting the period with the Chinese stock market crash. We note that Yousaf and Hassan (2019)
also use similar timeframes for the global financial crisis and the crash of the Chinese stock market.
This study follows He (2001) to use three-year data for each crisis for a short-run analysis. Changes in
the market correlations take place continuously, not only as a result of the crises but also due to the
consequences of many financial, economic, and political events. This study uses the same time for
both the crisis periods to make the coefficient comparable. The difference in the opening time of the
China and LA stock markets has been adjusted in the estimations.

3.2. Methodology
The econometric specification used in this study has two components. First, a vector
autoregression (VAR) with one lag is used to model the returns.2 This allows for autocorrelations and
cross-autocorrelations in the returns. Second, a multivariate BEKK-GARCH model is used to model the
time-varying variances and covariances developed by Engle and Kroner (1995).3 BEKK-GARCH has the
attractive characteristics that the conditional covariance matrices are positive definite (Chang et al. 2011).
Several studies have used the BEKK-GARCH model to estimate the spillover between different asset
classes; see, for example, Chang et al. (2011), Sadorsky (2012), Beirne et al. (2013), Chang et al. (2017),
Cardona et al. (2017), and Sarwar et al. (2020). Moreover, we will estimate the optimal weights and
hedge ratios using the BEKK-GARCH model.

2 The number of lags is selected on the basis of the AIC and SIC criteria.
3 We apply the BEKK-GARCH model on the valuable suggestion of a respected reviewer.

57
JRFM 2020, 13, 148

This study aims to examine the return and volatility spillover between the stock markets, and
thus we firstly focus on return spillover. For any pair of two series, the following are the specifications
for the conditional mean equation:

Rt = μ + ∅ Rt−1 + et with et = Ht1/2 ηt . (1)

y 
Rt = (Rxt , Rt ) is the vector of returns on the stock market indices x and y at time t, respectively; ∅
is the 2 × 2 matrix of parameters, measuring the impacts of own lagged and cross mean transmissions
 
y 
between two series; et = ext , et is the vector of error terms of the conditional mean equations for the
 
y 
two series at time t; ηt = ηxt , ηt indicates a sequence of independently and identically distributed
xy
Htx Ht
random errors; and Ht = xy y denotes the conditional variance-covariance matrix of return
Ht Ht
series of x and y. In addition, Ht1/2 is the 2 × 2 symmetric positive definite matrix.
The full BEKK–GARCH, which imposes positive definiteness restrictions for Ht , is given by:

Ht = C C + A et−1 et−1 A + B Ht−1 B, (2)

where A and B are (n × n) coefficient matrices and C C is the decomposition of the intercept matrix. Each
element (i,j)th in Ht depends on the corresponding (i,j)th element in (et−1 et−1 ) and Ht−1 . Accordingly,
past shocks and volatility are allowed to directly spill over from a market to another, and they are
captured by the coefficients of the A and B matrices. More specifically, the BEKK-GARCH matrices can
be expanded as follows:
 2 y
 y 2 xy y
hxt = Cx + α2x ext−1 + 2αx α yx ext−1 et−1 + α2yx et−1 + β2x hxt−1 + 2βx β yx ht−1 + β2yx ht−1 (3)

xy
 2   y
 y 2   xy y
ht = Cxy + αx αxy ext−1 + α yx αxy + αx α y ext−1 et−1 + α yx a y et−1 + βs βxy hxt−1 + β yx βxy + βx β y ht−1 + β yx β y ht−1 (4)

y
 2 y
 y 2 xy y
ht = C y + α2xy ext−1 + 2αxy α y ext−1 et−1 + α2y et−1 + β2xy hxt−1 + 2βxy β y ht−1 + β2y ht−1 (5)

The BEKK-GARCH parameters are estimated by the maximum likelihood method using the BFGS
algorithm. In addition to the return and volatility spillover, we also compute the optimal weights and
hedge ratios for each pair of stocks.
The conditional variance and covariances are used for calculating the optimal portfolio weights
and hedge ratios. This study follows Kroner and Ng (1998) in calculating the optimal portfolio weights
of different pairs of stock markets:
y xy
xy ht − ht
wt = xy y, (6)
hxt − 2ht + ht
⎧ xy
⎪
⎪ 0,I f wt < 0
⎪
⎪
xy ⎨ xy xy
wt = ⎪ ⎪wt ,I f 0 ≤ wt ≤ 1
⎪
⎪
⎩ xy
1,I f wt > 1
xy xy
where wt is the weight of stock(x) in a $1 stock(x)-stock(y) portfolio at time t; ht is the conditional
y
covariance between the two stock markets; hxt and ht are the conditional variance of stock(x) and stock(y),
xy
respectively; and 1 − wt is the weight of stock(y) in a $1 stock(x)-stock(y) portfolio. As suggested by
Kroner and Sultan (1993):
xy
xy h
βt = ty (7)
ht
xy
where βt represents the hedge ratio. This shows that a short position in the stock (y) market can hedge
a long position in stock (x).

58
JRFM 2020, 13, 148

4. Empirical Results and Implications

In this section, we will discuss our empirical results and implications. We first discuss our
preliminary analysis.

4.1. Descriptive Statistics

Table 1 reports the summary statistics of the daily returns for the US; China; and four emerging
LA stock markets—namely, Brazil, Chile, Mexico, and Peru. Among them, Brazil and Peru have
the highest mean return, and the US has the smallest mean return during the full sample period.
On the other hand, Chile has the smallest standard deviation, while Brazil has the largest standard
deviation. Thus, Peru provides the highest mean return, with a relatively smaller risk in the LA stock
markets. Overall, the skewness is significantly negative, the kurtosis is significantly higher than three
for all stocks, and the Jarque–Bera statistics reject normality hypothesis for all series, inferring that all
the returns are negatively skewed and fat-tailed. Moreover, Table 1 also confirms that there are 1%
significant autocorrelation and ARCH (autoregressive conditional heteroskedasticity) effects for all
returns. We also apply both Augmented Dickey–Fuller (ADF) and Phillip–Perron (PP) tests to examine
the stationarity of all the returns and exhibit the results in Table 2. The table indicates that all the series
are 1% significant, inferring that all the returns are stationary.

Table 1. Summary statistics.

Markets Mean Std. Dev. Skewness Kurtosis J-B Stat Q-Stat ARCH
US 0.00016 0.0124 −0.364 *** 14.045 *** 27181.3 *** 56.584 *** 548.40 ***
CHN 0.00040 0.0155 −0.330 *** 8.2116 *** 6121.9 *** 60.119 *** 189.01 ***
BRAZ 0.00047 0.0183 −0.403 *** 9.6439 *** 9937.1 *** 24.957 *** 686.82 ***
CHIL 0.00030 0.0105 −0.878 *** 19.883 *** 37,432.8 *** 148.49 *** 180.34 ***
MEXI 0.00024 0.0128 −0.086 * 8.3698 *** 6403.18 *** 108.33 *** 173.49 ***
PERU 0.00047 0.0133 −0.549 *** 15.441 *** 34605.3 *** 290.64 *** 796.97 ***
Notes: US—United States of America; CHN—China; BRAZ—Brazil; MEXI—Mexico; CHIL—Chile. Q-stat denotes
the Ljung–Box Q-statistics. ARCH test refers to the LM-ARCH test. ***, * indicate the statistical significance at 1%
and 10%, respectively.

Table 2. Unit root tests

ADF (t-Test) Phillips-Perron Test

Constant Constant
Markets None Constant None Constant
and Trend and Trend
US −79.73 *** −79.94 *** −79.96 *** −80.00 *** −80.01 *** −80.05 ***
CHN −32.44 *** −32.48 *** −32.49 *** −69.97 *** −69.89 *** −69.88 ***
BRAZ −72.04 *** −72.08 *** −72.08 *** −72.03 *** −72.08 *** −72.08 ***
CHIL −33.59 *** −33.64 *** −33.67 *** −63.11 *** −63.11 *** −63.10 ***
MEXI −50.70 *** −50.73 *** −50.73 *** −63.67 *** −63.68 *** −63.68 ***
PERU −31.06 *** −31.12 *** −31.15 *** −61.82 *** −61.75 *** −61.64 ***
Notes: US—United States of America; CHN—China; BRAZ—Brazil; MEXI—Mexico; CHIL—Chile;
ADF—Augmented Dickey Fuller. *** indicate the statistical significance at 10%, respectively.

4.2. Return and Volatility Spillover between the US and LA Stock Markets
We turn to apply the BEKK-GARCH model to examine the return and volatility spillovers between
the US and LA stock markets in the full sample period, the global financial crisis, and the crash of the
Chinese stock market and exhibit the results in Tables 3–5. We note that the 1% significant autocorrelation
and ARCH effects for all returns, as shown in Table 1, justify the use of the BEKK-GARCH model in
our analysis.

59
JRFM 2020, 13, 148

Table 3. Estimates of BEKK-GARCH for the US and Latin American stock markets during the full
sample period

Brazil and US Chile and US Mexico and US Peru and US

Coefficient p-Value Coefficient p-Value Coefficient p-Value Coefficient p-Value
Panel A. Mean Equation
μ1 0.075 *** 0.000 0.098 * 0.062 0.058 *** 0.000 0.047 *** 0.000
∅11 −0.031 ** 0.036 0.007 0.722 0.038 ** 0.027 0.137 *** 0.000
∅12 0.023 *** 0.005 0.000 0.938 0.021 0.109 −0.019 0.185
μ2 0.056 *** 0.000 0.062 *** 0.000 0.051 *** 0.000 0.058 *** 0.000
∅21 0.055 ** 0.028 0.111 *** 0.000 0.050 *** 0.005 0.081 *** 0.000
∅22 −0.077 *** 0.000 −0.057 *** 0.001 −0.071 *** 0.000 −0.042 *** 0.006
Panel B. Variance Equation
c11 0.219 *** 0.000 2.496 *** 0.000 0.117 *** 0.000 0.161 *** 0.000
c21 0.069 *** 0.004 0.124 *** 0.009 0.050 *** 0.006 0.056 *** 0.003
c22 0.118 *** 0.000 0.042 0.571 0.124 *** 0.000 0.122 *** 0.000
α11 0.225 *** 0.000 0.003 0.562 0.264 *** 0.000 0.309 *** 0.000
α12 −0.014 0.319 0.001 0.611 0.008 0.609 0.004 0.836
α21 0.036 0.421 0.151 ** 0.026 −0.026 0.390 0.010 0.541
α22 0.338 *** 0.000 0.341 *** 0.000 0.314 *** 0.000 0.300 *** 0.000
β11 0.966 *** 0.000 0.701 *** 0.000 0.959 *** 0.000 0.942 *** 0.000
β12 0.005 0.158 −0.003 0.400 0.005 0.349 −0.002 0.789
β21 −0.101 ** 0.050 0.095 0.203 0.090 ** 0.043 −0.004 0.538
β22 0.933 *** 0.000 0.944 *** 0.000 0.938 *** 0.000 0.947 *** 0.000
Panel C. Diagnostic Tests
LogL −16,174.1 −21,397.4 −13,816.1 −14,378.1
AIC 6.789 8.588 5.970 6.370
SIC 6.799 8.635 6.017 6.417
Q1 [20] 30.320 * 0.065 1.791 0.720 19.075 0.517 328.759 * 0.031
Q2 [20] 18.920 0.527 19.184 0.510 19.493 0.490 17.728 0.605
Q21 [20] 29.942 0.182 0.004 0.659 34.776 ** 0.021 30.071 0.198
Q22 [20] 27.782 0.115 22.616 0.308 25.161 0.185 33.181 0.146
Notes: US, United States of America; CHN, China; BRAZ, Brazil; CHIL, Chile; MEXI, Mexico. Variable order is the
Latin American stock market (1) and China (2). In the mean equations, μ denotes the constant terms, whereas ∅12
denotes the return spillover from the Latin American stock market to the US stock market. In the variance equation,
c denotes the constant terms, α denotes the ARCH terms, and β denotes the GARCH terms. In the variance equation,
α12 indicates the shock spillover from the Latin American stock market to the US stock market, whereas β12 denotes
the long−term volatility spillover from the Latin American stock market to the US stock market. Number of lags
for VAR is decided using the SIC and AIC criteria. JB, Q(20), and Q2 (20) indicate the empirical statistics of the
Jarque–Bera test for normality, Ljung–Box Q statistics of order 20 for autocorrelation applied to the standardized
residuals, and squared standardized residuals, respectively. Values in parentheses are the p-Value. ***, **, * indicate
the statistical significance at 1%, 5%, and 10%, respectively.

Tables 3–5 report the return and volatility spillovers between the US and LA stock markets during
the full sample period, the global financial crisis, and the crash of the Chinese stock market, respectively.
Referring to coefficients ∅11 and ∅22 in Panel A, the results show that the lagged returns significantly
influence the current returns in the US and the majority of LA stock markets during the full sample
period, the global financial crisis, and the crash of the Chinese stock market. It highlights the possibility
of the short-term prediction of current returns through past returns in the US and the majority of
the LA stock markets. Our results are consistent with the findings of Syriopoulos et al. (2015) and
Arouri et al. (2015), which observe a significant impact of past returns on current returns in the US
and LA stock markets.

60
JRFM 2020, 13, 148

Table 4. Estimates of BEKK-GARCH for US and Latin American stock markets during the global
financial crisis.

Brazil and US Chile and US Mexico and US Peru and US

Coefficient p-Value Coefficient p-Value Coefficient p-Value Coefficient p-Value
Panel A. Mean Equation
μ1 0.101 * 0.068 0.113 *** 0.000 0.055 0.246 −0.014 0.744
∅11 −0.044 0.450 0.118 *** 0.001 0.023 0.680 0.138 *** 0.000
∅12 0.021 0.581 0.019 0.440 0.096 ** 0.043 −0.009 0.770
μ2 0.019 0.665 0.046 0.361 0.064 0.144 0.029 0.518
∅21 0.040 0.578 0.020 0.308 0.020 0.707 0.063 0.107
∅22 −0.128 *** 0.007 −0.164 *** 0.000 −0.188 *** 0.000 −0.100 *** 0.003
Panel B. Variance Equation
c11 0.271 ** 0.044 0.287 *** 0.000 0.218 ** 0.017 0.291 *** 0.000
c21 0.098 0.546 0.040 0.417 −0.035 0.730 0.173 *** 0.000
c22 0.129 ** 0.039 0.153 *** 0.000 0.000 0.799 0.109 *** 0.007
α11 0.421 * 0.069 0.483 *** 0.000 0.117 0.220 0.453 *** 0.000
α12 0.139 ** 0.022 −0.055 0.333 −0.104 * 0.057 0.087 0.255
α21 −0.237 0.128 −0.020 0.550 0.249 *** 0.000 −0.088 0.581
α22 0.138 0.145 0.292 *** 0.000 0.295 *** 0.000 0.226 *** 0.002
β11 0.902 *** 0.000 0.841 *** 0.000 1.063 *** 0.000 0.896 *** 0.000
β12 -0.041 * 0.082 0.051 ** 0.034 0.218 *** 0.001 −0.034 0.197
β21 0.071 0.482 0.024 * 0.083 −0.183 *** 0.000 0.014 0.687
β22 0.990 *** 0.000 0.937 *** 0.000 0.797 *** 0.000 0.969 *** 0.000
Panel C. Diagnostic Tests
LogL −2766 −2438.432 −2514.181 −2804.767
AIC 7.792 7.026 7.132 8.025
SIC 8.018 7.253 7.359 8.251
Q1 [20] 15.405 0.753 15.396 0.753 15.749 0.732 18.138 0.578
Q2 [20] 19.980 0.459 22.469 0.316 25.023 0.201 19.998 0.458
Q21 [20] 23.671 0.257 17.388 0.628 15.064 0.773 13.734 0.844
Q22 [20] 37.237 ** 0.011 45.203 *** 0.001 33.878 ** 0.027 37.570 *** 0.010
Notes: US, United States of America; CHN, China; BRAZ, Brazil; CHIL, Chile; MEXI, Mexico. Variable order is the
Latin American stock market (1) and China (2). In the mean equations, μ denotes the constant terms, whereas ∅12
denotes the return spillover from the Latin American stock market to the US stock market. In the variance equation,
c denotes the constant terms, α denotes the ARCH terms, and β denotes the GARCH terms. In the variance equation,
α12 indicates the shock spillover from the Latin American stock market to the US stock market, whereas β12 denotes
the long-term volatility spillover from the Latin American stock market to the US stock market. Number of lags
for VAR is decided using the SIC and AIC criteria. JB, Q(20), and Q2 (20) indicate the empirical statistics of the
Jarque–Bera test for normality, Ljung–Box Q statistics of order 20 for autocorrelation applied to the standardized
residuals, and squared standardized residuals, respectively. Values in parentheses are the p-Value. ***, **, * indicate
the statistical significance at 1%, 5%, and 10%, respectively.

Regarding the interdependence of returns in the mean equation (see coefficients ∅12 and ∅21 in
Panel A), the results indicate the unidirectional return spillover from the US to the majority of LA stock
markets during the full sample period and the crash of the Chinese Stock Market. They imply that the
past US returns can be used to predict the current returns of the LA markets during the full sample
period and the crash of the Chinese Stock Market. These results are consistent with the previous
findings of Arouri et al. (2015), who find the unidirectional return spillover from the US to the LA stock
markets. Moreover, the return transmission is also significant from the Brazil to the US stock market
during the full sample period. In contrast, the return transmissions are not found to be significant
between the US and the majority of the LA stock (except Mexico) markets during the global financial
crisis. These results suggest that the US (LA) stock returns are not useful in predicting the returns in
the majority of the LA (US) stock markets during the global financial crisis. The results also reveal a
unidirectional volatility spillover from Mexico to the US stock market during the global financial crisis.
Based on the variance equation (see coefficients of α11 in Panel B), the results show that the
conditional volatility of the majority of LA stock markets depends on their past shocks during all the
sample periods. In addition, the coefficients of the past own shocks (α22 ) are highly significant for
the US in all the sample periods. Besides this, the sensitivity of past own volatility (β11 and β22 ) is
significant for the US and LA stock markets during all the sample periods. These results are consistent

61
JRFM 2020, 13, 148

with the findings of Syriopoulos et al. (2015), which find that the past own volatility is a significant
determinant of the future volatility of BRICS countries (including Brazil). Further, the coefficients of
past own volatility are higher compared to the coefficients of the past own shocks in the US and LA
stock markets, suggesting that the past own volatilities are more critical for the prediction of future
volatility than the past own shocks during all the sample periods.
Referring to the coefficient α12 and α21 in Panel B, the past shocks of the US stock market
significantly influence the conditional volatility of just the Chile stock market during the full sample
period. During the global financial crisis, the shock transmission is unidirectional from Brazil to
the US and bidirectional between the US and Mexican stock markets. Moreover, the conditional
volatility of the Mexican stock market is significantly affected by the US during the crash of the Chinese
stock market.

Table 5. Estimates of BEKK-GARCH for the US and Latin American stock markets during the crash of
the Chinese stock market.

Brazil and US Chile and US Mexico and US Peru and US

Coefficient p-Value Coefficient p-Value Coefficient p-Value Coefficient p-Value
Panel A. Mean Equation
μ1 0.090 0.105 0.030 0.210 0.013 0.585 0.088 ** 0.029
∅11 −0.047 0.172 0.077 * 0.082 −0.032 0.469 0.076 * 0.089
∅12 0.016 0.302 −0.035 0.134 0.002 0.937 −0.025 0.550
μ2 0.064 *** 0.004 0.075 *** 0.001 0.072 *** 0.001 0.061 0.158
∅21 0.129 * 0.064 0.120 *** 0.001 0.137 *** 0.000 0.086 * 0.093
∅22 −0.066 ** 0.040 −0.052 0.112 −0.059 * 0.083 −0.050 * 0.085
Panel B. Variance Equation
c11 0.268 *** 0.005 0.361 *** 0.004 0.599 *** 0.000 0.159 * 0.066
c21 0.151 ** 0.017 0.011 0.720 0.164 *** 0.000 0.117 0.122
c22 0.124 * 0.076 0.186 *** 0.000 0.124 0.150 0.089 0.790
α11 0.196 *** 0.001 0.522 *** 0.001 0.434 *** 0.000 0.278 ** 0.011
α12 0.008 0.821 −0.024 0.326 0.033 0.298 0.142 0.331
α21 0.023 0.744 −0.028 0.648 −0.115 * 0.052 0.019 0.850
α22 0.430 *** 0.000 0.421 *** 0.000 0.381 *** 0.000 0.313 0.416
β11 0.958 *** 0.000 0.686 *** 0.001 −0.359* 0.077 0.949 *** 0.000
β12 −0.008 0.571 0.018 0.411 −0.068 *** 0.000 −0.042 0.464
β21 0.013 0.697 0.076 0.227 0.528 *** 0.000 −0.014 0.766
β22 0.880 *** 0.000 0.879 *** 0.000 0.915 *** 0.000 0.917 *** 0.000
Panel C. Diagnostic Tests
LogL −2078 −1582.556 −1545.033 −1733.842
AIC 5.759 4.585 4.429 4.891
SIC 5.986 4.812 4.655 5.118
Q1 [20] 21.413 0.373 33.001 ** 0.034 21.955 0.343 31.804 ** 0.045
Q2 [20] 24.907 0.205 24.713 0.213 24.601 0.217 25.783 0.173
Q21 [20] 6.942 0.897 85.117 *** 0.000 29.827 * 0.073 16.276 0.699
Q22 [20] 8.249 0.890 9.945 0.969 10.383 0.961 8.909 0.984
Notes: US, United States of America; CHN, China; BRAZ, Brazil; CHIL, Chile; MEXI, Mexico. Variable order is the
Latin American stock market (1) and China (2). In the mean equations, μ denotes the constant terms, whereas ∅12
denotes the return spillover from the Latin American stock market to the US stock market. In the variance equation,
c denotes the constant terms, α denotes the ARCH terms, and β denotes the GARCH terms. In the variance equation,
α12 indicates the shock spillover from the Latin American stock market to the US stock market, whereas β12 denotes
the long-term volatility spillover from the Latin American stock market to the US stock market. Number of lags
for VAR is decided using the SIC and AIC criteria. JB, Q(20), and Q2 (20) indicate the empirical statistics of the
Jarque–Bera test for normality, Ljung–Box Q statistics of order 20 for autocorrelation applied to the standardized
residuals, and squared standardized residuals, respectively. Values in parentheses are the p-Value. ***, **, * indicate
the statistical significance at 1%, 5%, and 10%, respectively.

Regarding the cross-market volatility spillover (see coefficients β12 and β21 in Panel B), the results
indicate that the volatility transmission is unidirectional from the US to the Brazil and Mexican stock
markets during the full sample period. In contrast, the results reveal the bidirectional volatility
transmission between the US and two LA stock markets (Chile and Mexico), whereas there was
unidirectional volatility transmission from Brazil to the US stock market during the global financial crisis.

62
JRFM 2020, 13, 148

These results are in contrast with the findings of Wang et al. (2017), which report an insignificant volatility
spillover between the US and Brazil stock markets during the global financial crisis. The considerable
trade volumes between the US and two LA stock markets (Brazil and Mexico) explain the volatility
linkages between the stock markets of the concerned countries. Johnson and Soenen (2003) also suggest
that trade increases the financial contagion effects between the stock markets of concerned countries.
From the Latin American region, Mexico is the biggest trading partner of the US; therefore, volatility
linkages are also observed between Mexico and the US stock market during the global financial crisis.
These findings suggest that portfolio investors can get the maximum benefit of diversification by
making a portfolio of US and Peru stocks during the global financial crisis. Lastly, a bidirectional
volatility transmission is observed between the US and Mexican stock markets during the crash of the
Chinese stock market. It implies that portfolio investors can diversify risk by making a portfolio of the
US and LA stock markets (except Mexico) during the crash of the Chinese stock market.

4.3. Return and Volatility Spillover between China and the LA Stock Markets
Tables 6–8 represent the return and volatility transmissions between China and the LA stock
markets during the full sample period, the global financial crisis, and the crash of the Chinese stock
market. The difference in the opening time of the China and LA stock markets has been adjusted
where necessary in the estimations. Referring to the coefficient ∅11 in Panel A, the results indicate that
the lagged returns of the majority of LA stock markets (except Brazil) largely determine their current
returns during the full sample period and the crash of the Chinese stock market. During the global
financial crisis, the past returns significantly affect the current returns of the Chile and Peru stock
markets. This implies that the past returns can be used for the short-term prediction of the current
LA stock returns. These results confirm the previous findings of Arouri et al. (2015). Referring to the
coefficient ∅22 in Panel A, the lagged returns significantly influence the current returns in the Chinese
stock market during the full sample period. In contrast, the current returns of the Chinese stock market
are not influenced by their past returns during the global financial crisis and the crash of the Chinese
stock market. This implies that the past returns cannot be used for the short-term prediction of the
current Chinese stock returns during the crisis period.
Based on the cross-market return spillover (see the coefficients ∅12 and ∅21 in Panel A), the results
reveal the unidirectional return transmissions from China to the majority of LA stock markets during
all the sample periods. These results contradict the previous findings of Aktan et al. (2009) and
Sharma et al. (2013), who report the insignificant impact of the Chinese stock returns on the Brazilian
stock returns. In addition, the return transmission is also significant from Brazil to China during the
crash of the Chinese stock market.
From the variance equation (see coefficients α11 and α22 Panel B), the findings show that the
lagged shocks significantly influence the conditional volatility of the China and LA stock markets
during all the sample periods. Referring to the coefficients β11 and β22 , the results show that the current
conditional volatility depends on their past volatility in the China and LA stock markets during the
all sample periods. The critical finding is that the coefficients of past own volatility are seen to be
higher compared to the past own shocks. This difference suggests that past own volatilities rather
than past shocks are more important for the prediction of the current volatility in the China and LA
stock markets.
Refer to the coefficients α12 and α21 in panel B, the shock transmission is unidirectional from
Brazil and Peru to the Chinese stock market, whereas bidirectional shock transmission is observed
between the China and Mexican stock markets during the full sample period. The results reveal that
the past shocks in the Brazil and Mexican stock markets significantly affect the conditional volatility of
the Chinese stock market during the global financial crisis. On the other hand, the shock spillover is
insignificant between China and the majority of the LA stock markets during the crash of the Chinese
stock market.

63
JRFM 2020, 13, 148

Table 6. Estimates of BEKK-GARCH for the China and Latin American stock markets during the full
sample period.

Brazil and China Chile and China Mexico and China Peru and China
Coefficient p-Value Coefficient p-Value Coefficient p-Value Coefficient p-Value
Panel A. Mean Equation
μ1 0.076 *** 0.001 0.047 *** 0.000 0.038 *** 0.004 0.050 *** 0.000
∅11 0.033 ** 0.050 0.205 *** 0.000 0.101 *** 0.000 0.234 *** 0.000
∅12 0.013 0.208 0.020 0.190 0.014 0.213 0.015 0.287
μ2 0.044 *** 0.008 0.042 ** 0.026 0.050 *** 0.000 0.039 ** 0.045
∅21 0.132 *** 0.000 0.036 *** 0.000 0.075 *** 0.000 0.053 *** 0.000
∅22 0.037 *** 0.005 0.041 ** 0.019 0.036 ** 0.011 0.042 *** 0.009
Panel B. Variance Equation
c11 0.283 *** 0.000 0.186 *** 0.000 0.138 *** 0.000 0.219 *** 0.000
c21 0.009 0.699 −0.001 0.956 0.009 0.721 −0.013 0.500
c22 0.118 *** 0.000 0.121 *** 0.000 0.115 *** 0.000 0.108 *** 0.000
α11 0.273 *** 0.000 0.347 *** 0.000 0.279 *** 0.000 0.394 *** 0.000
α12 −0.023 * 0.059 0.013 0.490 −0.037 *** 0.002 −0.024 * 0.057
α21 0.000 0.899 0.001 0.950 0.021 * 0.087 −0.002 0.920
α22 0.250 *** 0.000 0.243 *** 0.000 0.240 *** 0.000 0.237 *** 0.000
β11 0.948 *** 0.000 0.919 *** 0.000 0.954 *** 0.000 0.904 *** 0.000
β12 0.015 ** 0.040 −0.002 0.741 0.009 *** 0.003 0.015 *** 0.007
β21 0.002 0.814 0.002 0.726 −0.004 0.283 0.001 0.840
β22 0.966 *** 0.000 0.968 *** 0.000 0.969 *** 0.000 0.970 *** 0.000
Panel C. Diagnostic Tests
LogL −19,187.432 −15,839.650 −17,037.161 −16,739.334
AIC 7.720 6.599 7.002 7.045
SIC 7.767 6.646 7.049 7.092
Q1 [20] 21.935 0.344 19.993 0.458 17.078 0.648 72.725 *** 0.000
Q2 [20] 82.861 *** 0.000 78.794 *** 0.000 83.815 *** 0.000 80.555 *** 0.000
Q21 [20] 26.742 0.133 8.890 0.984 26.056 0.187 18.240 0.572
Q22 [20] 22.787 0.299 22.412 0.319 25.134 0.196 25.146 0.196
Notes: US, United States of America; CHN, China; BRAZ, Brazil; CHIL, Chile; MEXI, Mexico. Variable order is the
Latin American stock market (1) and China (2). In the mean equations, μ denotes the constant terms, whereas ∅12
denotes the return spillover from the Latin American stock market to the Chinese stock market. In the variance
equation, c denotes the constant terms, α denotes the ARCH terms, and β denotes the GARCH terms. In the
variance equation, α12 indicates the shock spillover from the Latin American stock market to the Chinese stock
market, whereas β12 denotes the long−term volatility spillover from the Latin American stock market to the Chinese
stock market. Number of lags for VAR is decided using the SIC and AIC criteria. JB, Q(20), and Q2 (20) indicate
the empirical statistics of the Jarque–Bera test for normality, Ljung–Box Q statistics of order 20 for autocorrelation
applied to the standardized residuals, and squared standardized residuals, respectively. Values in parentheses are
the p-Value. ***, **, * indicate the statistical significance at 1%, 5%, and 10%, respectively.

Based on the cross-market volatility spillover effects (see coefficients β12 and β21 in Panel B),
the results demonstrate that there is unidirectional volatility transmission from Brazil, Mexico, and Peru
to China during the full sample period. These volatility transmissions can be explained through the
considerable trading volumes between China and two Latin economies (Brazil and Mexico) during the
full sample period. During the global financial crisis, the volatility effects are transmitted from the
China to Brazil stock markets. Therefore, the majority of LA stock markets provide an opportunity to
diversify the risk of Chinese equity portfolios during the global financial crisis. Lastly, the volatility
spillover is bidirectional between the China and Brazil stock markets during the crash of the Chinese
stock market. Due to the crash of the Chinese stock market, the slowdown of the Chinese economy
also affected its major trading partner Brazil and its stock market; therefore, volatility linkages are also
observed between China and Brazil. These findings propose that the portfolio investors of Chinese
stock markets can get the maximum benefit of diversification by adding Mexico, Chile, and Peru stocks
in their portfolios during the crash of the Chinese stock market.

64
JRFM 2020, 13, 148

Table 7. Estimates of BEKK-GARCH for the China and Latin American stock markets during the global
financial crisis.

BRAZIL and China Chile and China Mexico and China Peru and China
Coefficient p-Value Coefficient p-Value Coefficient p-Value Coefficient p-Value
Panel A. Mean Equation
μ1 0.106 * 0.081 0.116 *** 0.001 0.035 0.476 0.014 0.773
∅11 0.029 0.465 0.195 *** 0.000 0.034 0.340 0.211 *** 0.000
∅12 −0.003 0.938 −0.073 0.189 −0.034 0.424 −0.023 0.554
μ2 0.044 0.590 0.001 0.987 0.090 0.285 0.030 0.731
∅21 0.186 *** 0.000 0.030 * 0.083 0.101 *** 0.000 0.103 *** 0.000
∅22 0.032 0.422 0.022 0.572 0.033 0.431 0.027 0.488
Panel B. Variance Equation
c11 −0.201 0.105 0.142 ** 0.025 0.127 *** 0.008 0.344 *** 0.000
c21 0.134 0.577 1.932 *** 0.000 0.134 0.479 0.081 0.250
c22 0.143 0.385 0.000 0.920 0.195 ** 0.037 0.163 * 0.098
α11 0.297 *** 0.000 0.408 *** 0.000 0.253 *** 0.000 0.477 *** 0.000
α12 −0.041 0.409 −0.108 0.124 −0.041 0.392 −0.005 0.902
α21 −0.069 * 0.089 0.007 0.686 0.060 ** 0.020 −0.031 0.215
α22 0.188 *** 0.001 0.316 *** 0.000 0.209 *** 0.000 0.181 *** 0.000
β11 0.947 *** 0.000 0.893 *** 0.000 0.960 *** 0.000 0.870 *** 0.000
β12 0.006 0.580 0.021 0.130 0.003 0.820 0.005 0.777
β21 0.028 * 0.085 0.047 0.260 −0.009 0.292 0.005 0.515
β22 0.977 *** 0.000 −0.259 0.143 0.972 *** 0.000 0.979 *** 0.000
Panel C. Diagnostic Tests
LogL −3318.981 −2879.180 −3090.113 −3166.538
AIC 8.973 7.826 8.388 8.689
SIC 9.200 8.052 8.614 8.915
Q1 [20] 14.730 0.792 11.805 0.923 17.935 0.592 17.407 0.626
Q2 [20] 30.922 * 0.056 32.099 ** 0.042 30.614 * 0.060 31.335 * 0.051
Q21 [20] 22.021 0.339 12.016 0.916 14.074 0.827 9.990 0.968
Q22 [20] 28.559 0.127 37.567 0.161 26.806 0.141 28.213 0.104
Notes: US, United States of America; CHN, China; BRAZ, Brazil; CHIL, Chile; MEXI, Mexico. Variable order
is the Latin American stock market (1) and China (2). In the mean equations and μ denotes the constant terms,
whereas ∅12 denotes the return spillover from the Latin American stock market to the Chinese stock market. In the
variance equation, c denotes the constant terms, α denotes the ARCH terms, and β denotes the GARCH terms. In
the variance equation, α12 indicates the shock spillover from the Latin American stock market to the Chinese stock
market, whereas β12 denotes the long-term volatility spillover from the Latin American stock market to the Chinese
stock market. Number of lags for VAR is decided using the SIC and AIC criteria. JB, Q(20), and Q2 (20) indicate
the empirical statistics of the Jarque–Bera test for normality, Ljung–Box Q statistics of order 20 for autocorrelation
applied to the standardized residuals, and squared standardized residuals, respectively. Values in parentheses are
the p-Value. ***, **, * indicate the statistical significance at 1%, 5%, and 10%, respectively.

Table 8. Estimates of BEKK-GARCH for the China and Latin American stock markets during the crash
of the Chinese stock market.

BRAZIL and China Chile and China Mexico and China Peru and China
Coefficient p-Value Coefficient p-Value Coefficient p-Value Coefficient p-Value
Panel A. Mean Equation
μ1 0.055 0.282 0.039 * 0.063 0.008 0.765 0.059 * 0.087
∅11 0.036 0.292 0.181 *** 0.000 0.077 ** 0.017 0.229 *** 0.000
∅12 0.043 ** 0.048 0.026 0.350 0.012 0.791 0.033 0.333
μ2 0.022 0.495 0.020 0.474 0.020 0.420 0.022 0.410
∅21 0.105 *** 0.001 0.057 *** 0.001 0.074 *** 0.000 0.062 *** 0.003
∅22 0.032 0.360 0.040 0.228 0.039 0.255 0.026 0.470

65
JRFM 2020, 13, 148

Table 8. Cont.

BRAZIL and China Chile and China Mexico and China Peru and China
Coefficient p-Value Coefficient p-Value Coefficient p-Value Coefficient p-Value
Panel A. Mean Equation
μ1 0.055 0.282 0.039 * 0.063 0.008 0.765 0.059 * 0.087
∅11 0.036 0.292 0.181 *** 0.000 0.077 ** 0.017 0.229 *** 0.000
∅12 0.043 ** 0.048 0.026 0.350 0.012 0.791 0.033 0.333
μ2 0.022 0.495 0.020 0.474 0.020 0.420 0.022 0.410
∅21 0.105 *** 0.001 0.057 *** 0.001 0.074 *** 0.000 0.062 *** 0.003
∅22 0.032 0.360 0.040 0.228 0.039 0.255 0.026 0.470
Panel B. Variance Equation
c11 0.915 *** 0.000 0.232 * 0.057 0.390 *** 0.000 0.259 *** 0.009
c21 0.126 *** 0.001 0.049 0.132 0.033 0.553 0.089 ** 0.045
c22 0.000 0.865 0.000 0.799 0.058 0.237 0.051 0.320
α11 0.334 *** 0.000 0.452 *** 0.004 0.405 *** 0.000 0.345 *** 0.004
α12 −0.043 0.159 0.096 ** 0.021 −0.017 0.778 0.006 0.934
α21 0.051 0.336 0.005 0.845 −0.069 0.394 0.054 0.182
α22 0.236 *** 0.000 0.208 *** 0.000 0.229 *** 0.000 0.256 *** 0.000
β11 0.691 *** 0.000 0.844 *** 0.000 0.751 *** 0.000 0.890 *** 0.000
β12 −0.069 * 0.071 −0.065 0.109 −0.032 0.662 −0.032 0.410
β21 −0.072 * 0.056 0.005 0.525 0.031 0.260 −0.012 0.404
β22 0.968 *** 0.000 0.978 *** 0.000 0.974 *** 0.000 0.965 *** 0.000
Panel C. Diagnostic Tests
LogL −2506.043 −1947.155 −2014.255 −2107.485
AIC 7.227 5.934 5.989 6.295
SIC 7.454 6.160 6.216 6.521
Q1 [20] 21.462 0.370 29.255 * 0.083 24.444 0.224 23.882 0.248
Q2 [20] 23.562 0.262 27.467 0.123 24.664 0.215 26.564 0.148
Q21 [20] 10.480 0.959 61.006 *** 0.000 15.298 0.759 14.951 0.779
Q22 [20] 21.376 0.375 27.907 0.112 25.661 0.177 20.186 0.446
Notes: US, United States of America; CHN, China; BRAZ, Brazil; CHIL, Chile; MEXI, Mexico. Variable order
is the Latin American stock market (1) and China (2). In the mean equations and μ denotes the constant terms,
whereas ∅12 denotes the return spillover from the Latin American stock market to the Chinese stock market. In the
variance equation, c denotes the constant terms, α denotes the ARCH terms, and β denotes the GARCH terms. In the
variance equation, α12 indicates the shock spillover from the Latin American stock market to the Chinese stock
market, whereas β12 denotes the long-term volatility spillover from the Latin American stock market to the Chinese
stock market. Number of lags for VAR is decided using the SIC and AIC criteria. JB, Q(20), and Q2 (20) indicate
the empirical statistics of the Jarque–Bera test for normality, Ljung–Box Q statistics of order 20 for autocorrelation
applied to the standardized residuals, and squared standardized residuals, respectively. Values in parentheses are
the p-Value. ***, **, * indicate the statistical significance at 1%, 5%, and 10%, respectively.

4.4. Optimal Weights and Hedge Ratio Portfolio Implications

In the above-mentioned results, volatility transmission is observed between the several pairs of
stock markets during the different sample periods. Thus, investment in these pairs of stock markets
reduces the benefit of diversification. Therefore, the risk transmission across stock markets push
investors to adjust their asset allocation and to hedge their portfolio risk over time. For this reason,
this study estimates the optimal weights and hedge ratios.
Tables 9 and 10 report the optimal weights for the pairs of LA-US and LA-China during all the
sample periods. The findings reveal that the optimal weight is 0.11 for BRAZ/US during the full
sample period, revealing that for a $1 portfolio in Brazil-US, 11 cents should be invested in the Brazil
stock market and the remaining 89 cents in the US stock market. The interpretations of all the optimal
weights are not interpreted here for the sake of brevity. For the LA-US portfolio (see Table 9), the results
show that the average optimal weights are seen to be higher in the global financial crisis and the
crash of the Chinese stock market as compared to the full sample period. For the LA-US portfolio,

66
JRFM 2020, 13, 148

the investors are suggested to allocate a higher proportion of investment in LA stocks during the global
financial crisis and the crash of the Chinese stock market. For the pair of LA-China (see Table 10),
the results show that the optimal weights are higher during the global financial crisis and the crash of
the Chinese stock market compared to the full sample period. For the LA-China portfolio, investors
should increase their investment in LA stocks during the global financial crisis and the crash of the
Chinese stock market.

Table 9. Optimal weights and hedge ratios for Latin America (LA)/US

BRAZ/US CHIL/US MEXI/US Peru/US

Full Sample Period
wLU
t
0.11 0.51 0.29 0.27
βLU
t
0.93 0.63 0.25 0.28
US Financial Crisis
wLU
t
0.17 0.77 0.49 0.41
βLU
t
0.94 0.42 0.77 0.56
Chinese Stock Market Crash
wLU
t
0.09 0.54 0.46 0.39
βLU
t
0.98 0.34 0.59 0.43
t and βt represent the optimal weight and hedge ratio for the LA-US pair. L and U in superscripts denote
Note: wLU LU

the Latin American and US stock markets, respectively.

Table 10. Optimal weights and hedge ratios for LA/China

Header BRAZ/CHN CHIL/CHN MEXI/CHN Peru/CHN

Full Sample Period
wLC
t
0.41 0.70 0.61 0.63
βLC
t
0.14 0.07 0.07 0.08
US Financial Crisis
wLC
t
0.53 0.81 0.68 0.63
βLC
t
0.21 0.09 0.15 0.13
Chinese Stock Market Crash
wLC
t
0.43 0.68 0.64 0.64
βLC
t
0.23 0.11 0.07 0.11
t and βt represent the optimal weight and hedge ratio for LA-China pair. L and C in superscripts denote
Note: wLC LC

the Latin American and Chinese stock markets, respectively.

It is also essential to estimate the risk-minimizing optimal hedge ratios for portfolios of different
stocks. Referring to Table 9, the optimal hedge ratio range is 0.93 for BRAZ/US during the full sample
period, showing that a $1-long position in Brazil stocks can be hedged for 93 cents with a short position
in the US stocks. The interpretations of all the optimal hedge ratios are not interpreted here for the
sake of brevity. For the LA-US portfolio (see Table 9), the average optimal hedge ratios are found to be
higher for most of the pairs during the global financial crisis and the crash of the Chinese stock market
compared to the full sample period. It implies that less LA stocks are needed to minimize the risk of
US stock during crisis periods compared to the full sample period. For the LA-China portfolio (see
Table 10), the optimal hedge ratios are also higher during both crises, which implies that the lesser LA
stocks are required to minimize the risk of the Chinese stock market during both crises compared to
the full sample period.

5. Conclusions
This study examines the return and volatility spillover between the world-leading (the US and
China) and emerging Latin American (Brazil, Chile, Mexico, and Peru) stock markets during the full
sample period, the global financial crisis, and the crash of the Chinese stock market. Moreover, this study

67
JRFM 2020, 13, 148

also estimates the optimal weights and hedge ratios during all the sample periods. The BEKK-GARCH
model is applied to estimate the return and volatility spillover between the stock markets.
Regarding return spillover, the results reveal a unidirectional return spillover from the US to the
majority of the LA stock markets during the full sample period and the Chinese crash. This implies
that the US stock market prices play an important role in predicting the prices of the majority of LA
stock markets during the full sample period and the Chinese crash. During the global financial crisis,
the return transmissions are not significant between the US and the majority of Latin American stock
markets. This implies that the prices of the US (LA) stock markets do not contribute to the role of
price discovery in the LA (US) stock markets during the global financial crisis. For the China-LA
nexus, the results reveal a unidirectional return transmission from China to Brazil, Chile, Mexico,
and Peru stock markets during all the sample periods. Thus, the Chinese stock returns can be useful in
predicting the returns of the LA stock markets.
Regarding the volatility spillover between the US and LA stock markets, the results reveal the
bidirectional volatility transmission between the US and two stock markets of Chile and Mexico, as well
as the unidirectional volatility transmission from Brazil to the US stock market during the global
financial crisis. During the Chinese crash, a bidirectional volatility transmission is observed between
the US and Mexican stock markets. This implies that portfolio investors can diversify risk by making a
portfolio of the US and LA stock markets (except Mexico) during the crash of the Chinese stock market.
Regarding the volatility spillover between the China and LA stock markets, the volatility spillover
is unidirectional from the China to Brazil stock markets during the global financial crisis. Therefore,
the majority of the LA stock markets provide an opportunity to diversify the risk of Chinese equity
portfolios during the global financial crisis. During the Chinese crash, the volatility spillover is
bidirectional between the China and Brazil stock markets. These findings propose that the portfolio
investors of the Chinese stock markets can get the maximum benefit of diversification by adding
Mexico, Chile, and Peru stocks to their portfolios during the crash of the Chinese stock market.
These findings are also important because understanding the stock market volatility behavior can play
a vital role during the valuation of derivatives and for hedging purposes. Moreover, policymakers
should consider the “prices and volatilities of the world-leading stock market” as one of the critical
factors while devising the policies to stabilize their emerging financial markets.
Based on optimal weights, investors are suggested to allocate a higher proportion of investment
to the LA stocks in the LA-US portfolio during the global financial crisis and the crash of the Chinese
stock market. For the LA-China portfolio, investors should increase their investment in the LA stocks
during the global financial crisis and the crash of the Chinese stock market. Based on hedge ratios,
less LA stocks are needed to minimize the risk of the US and Chinese stocks during the periods of
both crises compared to the full sample period. Overall, these findings provide useful information
for policymakers and portfolio managers regarding optimal asset allocation, diversification, hedging,
forecasting, and risk management.
This study employs the BEKK-GARCH model to examine the linkages between the world-leading
countries and the emerging Latin American stock markets. Extensions could include other models to
examine the return and volatility spillover—for example, cointegration and causality (Lv et al. 2019;
Demirer et al. 2019), Copulas (Ly et al. 2019a, 2019b; Yuan et al. 2020), Stochastic Dominance
(Chiang et al. 2008; Abid et al. 2014; Guo et al. 2017; Wong et al. 2018), and many others. See,
for example, Chang et al. (2018), Woo et al. (2020), and the references therein for more information.

Author Contributions: Conceptulization, estimations, formal analysis, original draft preparation (I.Y.); Data
collection, methodology writing, and review of draft (S.A.); review, editing, and funding (W.-K.W.). All authors
have read and agreed to the published version of the manuscript.
Funding: The research is partially supported by Asia University, China Medical University Hospital, Hang
Seng University of Hong Kong, and the Ministry of Science and Technology (MOST) (Project Numbers
106-2410-H-468-002 and 107-2410-H-468-002-MY3).

68
JRFM 2020, 13, 148

Acknowledgments: The first author gratefully acknowledge Arshad Hassan (Professor/Dean, department of
Management and Social Sciences, Capital University of Science and Technology, Islamabad) for their valuable
suggestions. The third author would like to thank Robert B. Miller and Howard Thompson for their continuous
guidance and encouragement. All the errors remain with the authors.
Conflicts of Interest: The authors declare no conflict of interest.

References
Abid, Fathi, Pui Leung, Mourad Mroua, and Wing Wong. 2014. International Diversification Versus Domestic
Diversification: Mean-Variance Portfolio Optimization and Stochastic Dominance Approaches. Journal of
Risk and Financial Management 7: 45–66. [CrossRef]
Ahmad, Wasim, and Sanjay Sehgal. 2015. Regime Shifts and Volatility in BRIICKS Stock Markets: An Asset
Allocation Perspective. International Journal of Emerging Markets 10: 383–408. [CrossRef]
Ahmed, Abdullahi D., and Rui Huo. 2019. Impacts of China’s Crash on Asia-Pacific Financial Integration:
Volatility Interdependence, Information Transmission and Market Co-Movement. Economic Modelling 79:
28–46. [CrossRef]
Aktan, Bora, Pinar Evrim Mandaci, Baris Serkan Kopurlu, and Bulent Ersener. 2009. Behaviour of Emerging Stock
Markets in the Global Financial Meltdown: Evidence from Bric-A. African Journal of Business Management 3:
396–404.
Al Nasser, Omar M., and Massomeh Hajilee. 2016. Integration of Emerging Stock Markets with Global Stock
Markets. Research in International Business and Finance 36: 1–12. [CrossRef]
Ali, Shoaib, Imran Yousaf, and Muhammad Naveed. 2020. Role of credit rating in determining capital structure:
Evidence from non-financial sector of Pakistan. Studies of Applied Economics 38. [CrossRef]
Arouri, Mohamed El Hédi, Amine Lahiani, and Duc Khuong Nguyen. 2015. Cross-Market Dynamics and Optimal
Portfolio Strategies in Latin American Equity Markets. European Business Review 27: 161–81. [CrossRef]
Beirne, John, Guglielmo Maria Caporale, Marianne Schulze-Ghattas, and Nicola Spagnolo. 2013. Volatility
Spillovers and Contagion from Mature to Emerging Stock Markets. Review of International Economics 21:
1060–75. [CrossRef]
Bekiros, Stelios D. 2014. Contagion, decoupling and the spillover effects of the US financial crisis: Evidence from
the BRIC markets. International Review of Financial Analysis 33: 58–69. [CrossRef]
Bhuyan, Rafiqul, Mohammad G. Robbani, Bakhtear Talukdar, and Ajeet Jain. 2016. Information Transmission and
Dynamics of Stock Price Movements: An Empirical Analysis of BRICS and US Stock Markets. International
Review of Economics & Finance 46: 180–95.
Bouri, Elie I. 2013. Correlation and Volatility of the MENA Equity Markets in Turbulent Periods, and Portfolio
Implications. Economics Bulletin 33: 1575–93.
Cao, Guangxi, Yan Han, Qingchen Li, and Wei Xu. 2017. Asymmetric MF-DCCA Method Based on Risk
Conduction and Its Application in the Chinese and Foreign Stock Markets. Physica A: Statistical Mechanics
and Its Applications 468: 119–30. [CrossRef]
Cardona, Laura, Marcela Gutiérrez, and Diego A. Agudelo. 2017. Volatility Transmission between US and Latin
American Stock Markets: Testing the Decoupling Hypothesis. Research in International Business and Finance
39: 115–27. [CrossRef]
Chancharoenchai, Kanokwan, and Sel Dibooglu. 2006. Volatility Spillovers and Contagion During the Asian Crisis:
Evidence from Six Southeast Asian Stock Markets. Emerging Markets Finance and Trade 42: 4–17. [CrossRef]
Chang, Chia-Lin, Michael McAleer, and Roengchai Tansuchat. 2011. Crude Oil Hedging Strategies Using Dynamic
Multivariate GARCH. Energy Economics 33: 912–23. [CrossRef]
Chang, Chia-Lin, Michael McAleer, and Guangdong Zuo. 2017. Volatility Spillovers and Causality of Carbon
Emissions, Oil and Coal Spot and Futures for the EU and USA. Sustainability 9: 1789. [CrossRef]
Chang, Chia-Lin, Michael McAleer, and Wing-Keung Wong. 2018. Big Data, Computational Science, Economics,
Finance, Marketing, Management, and Psychology: Connections. Journal of Risk and Financial Management 11:
15. [CrossRef]
Chen, Gong-Meng, Michael Firth, and Oliver Meng Rui. 2002. Stock Market Linkages: Evidence from Latin
America. Journal of Banking & Finance 26: 1113–41.

69
JRFM 2020, 13, 148

Chiang, Thomas, Hooi Hooi Lean, and Wing-Keung Wong. 2008. Do REITs Outperform Stocks and Fixed-Income
Assets? New Evidence from Mean-Variance and Stochastic Dominance Approaches. Journal of Risk and
Financial Management 1: 1–40. [CrossRef]
Coleman, Simeon, Vitor Leone, and Otavio R. de Medeiros. 2018. Latin American Stock Market Dynamics and
Comovement. International Journal of Finance & Economics 24: 1109–29.
Demirer, Riza, Rangan Gupta, Zhihui Lv, and Wing-Keung Wong. 2019. Equity Return Dispersion and Stock
Market Volatility: Evidence from Multivariate Linear and Nonlinear Causality Tests. Sustainability 11: 351.
[CrossRef]
Diamandis, Panayiotis F. 2009. International Stock Market Linkages: Evidence from Latin America. Global Finance
Journal 20: 13–30. [CrossRef]
Diebold, Francis X., and Kamil Yilmaz. 2012. Better to Give than to Receive: Predictive Directional Measurement
of Volatility Spillovers. International Journal of Forecasting 28: 57–66. [CrossRef]
Engle, Robert F., and Kenneth F. Kroner. 1995. Multivariate Simultaneous Generalized ARCH. Econometric Theory
11: 122–50. [CrossRef]
Fang, Lu, and David A. Bessler. 2017. Is It China That Leads the Asian Stock Market Contagion in 2015? Applied
Economics Letters 25: 752–57. [CrossRef]
Fernández-Serrano, José L., and Simón Sosvilla-Rivero. 2003. Modelling the Linkages between US and Latin
American Stock Markets. Applied Economics 35: 1423–34. [CrossRef]
Fortunato, Graziela, Nathalia Martins, and Carlos de Lamare Bastian-Pinto. 2019. Global Economic Factors and
the Latin American Stock Markets. Latin American Business Review 21: 61–91. [CrossRef]
Gamba-Santamaria, Santiago, Jose Eduardo Gomez-Gonzalez, Jorge Luis Hurtado-Guarin, and Luis
Fernando Melo-Velandia. 2017. Stock Market Volatility Spillovers: Evidence for Latin America. Finance
Research Letters 20: 207–16. [CrossRef]
Garza-García, Jesus Gustavo, and Maria Eugenia Vera-Juárez. 2010. Who influences Latin American stock market
returns? China versus USA. International Research Journal of Finance and Economics 55: 22–35.
Graham, Michael, Jarno Kiviaho, and Jussi Nikkinen. 2012. Integration of 22 Emerging Stock Markets: A
Three-Dimensional Analysis. Global Finance Journal 23: 34–47. [CrossRef]
Gulzar, Saqib, Ghulam Mujtaba Kayani, Hui Xiaofeng, Usman Ayub, and Amir Rafique. 2019. Financial
Cointegration and Spillover Effect of Global Financial Crisis: A Study of Emerging Asian Financial Markets.
Economic Research-Ekonomska Istraživanja 32: 187–218. [CrossRef]
Guo, Xu, Xuejun Jiang, and Wing-Keung Wong. 2017. Stochastic Dominance and Omega Ratio: Measures to
Examine Market Efficiency, Arbitrage Opportunity, and Anomaly. Economies 5: 38. [CrossRef]
Han, Qian, and Jufang Liang. 2016. Index Futures Trading Restrictions and Spot Market Quality: Evidence from
the Recent Chinese Stock Market Crash. Journal of Futures Markets 37: 411–28. [CrossRef]
He, Ling T. 2001. Time Variation Paths of International Transmission of Stock Volatility—US vs. Hong Kong and
South Korea. Global Finance Journal 12: 79–93. [CrossRef]
Horvath, Roman, and Petr Poldauf. 2012. International Stock Market Comovements: What Happened during the
Financial Crisis? Global Economy Journal 12: 1850252. [CrossRef]
Hwang, Jae-Kwang. 2014. Spillover Effects of the 2008 Financial Crisis in Latin America Stock Markets. International
Advances in Economic Research 20: 311–24. [CrossRef]
In, Francis, Sangbae Kim, Jai Hyung Yoon, and Christopher Viney. 2001. Dynamic Interdependence and Volatility
Transmission of Asian Stock Markets. International Review of Financial Analysis 10: 87–96. [CrossRef]
Johnson, Robert, and Luc Soenen. 2003. Economic Integration and Stock Market Comovement in the Americas.
Journal of Multinational Financial Management 13: 85–100. [CrossRef]
Kroner, Kenneth F., and Victor K. Ng. 1998. Modeling Asymmetric Comovements of Asset Returns. Review of
Financial Studies 11: 817–44. [CrossRef]
Kroner, Kenneth F., and Jahangir Sultan. 1993. Time-Varying Distributions and Dynamic Hedging with Foreign
Currency Futures. The Journal of Financial and Quantitative Analysis 28: 535. [CrossRef]
Li, Yanan, and David E. Giles. 2014. Modelling Volatility Spillover Effects Between Developed Stock Markets and
Asian Emerging Stock Markets. International Journal of Finance & Economics 20: 155–77.
Lv, Zhihui, Amanda M. Y. Chu, Michael McAleer, and Wing-Keung Wong. 2019. Modelling Economic Growth,
Carbon Emissions, and Fossil Fuel Consumption in China: Cointegration and Multivariate Causality.
International Journal of Environmental Research and Public Health 16: 4176. [CrossRef] [PubMed]

70
JRFM 2020, 13, 148

Ly, Sel, Kim-Hung Pho, Sal Ly, and Wing-Keung Wong. 2019a. Determining Distribution for the Product of
Random Variables by Using Copulas. Risks 7: 23. [CrossRef]
Ly, Sel, Kim-Hung Pho, Sal Ly, and Wing-Keung Wong. 2019b. Determining Distribution for the Quotients of
Dependent and Independent Random Variables by Using Copulas. Journal of Risk and Financial Management
12: 42. [CrossRef]
Malik, Farooq, and Shawkat Hammoudeh. 2007. Shock and Volatility Transmission in the Oil, US and Gulf Equity
Markets. International Review of Economics & Finance 16: 357–68.
Markovitz, Harry M. 1959. Portfolio Selection: Efficient Diversification of Investments. Cowles Foundation Monograph
16. London: Yale University Press.
Mensi, Walid, Shawkat Hammoudeh, Duc Khuong Nguyen, and Sang Hoon Kang. 2016. Global Financial Crisis
and Spillover Effects among the U.S. and BRICS Stock Markets. International Review of Economics & Finance
42: 257–76.
Meric, Gulser, Ricardo P. C. Leal, Mitchell Ratner, and Ilhan Meric. 2001. Co-Movements of U.S. and Latin
American Equity Markets before and after the 1987 Crash. International Review of Financial Analysis 10:
219–235. [CrossRef]
Ramirez-Hassan, Andres, and Javier Orlando Pantoja. 2018. Co-Movements between Latin American and US
Stock Markets: Convergence After the Financial Crisis? Latin American Business Review 19: 157–72. [CrossRef]
Ben Rejeb, Aymen, and Mongi Arfaoui. 2016. Financial Market Interdependencies: A Quantile Regression
Analysis of Volatility Spillover. Research in International Business and Finance 36: 140–57. [CrossRef]
Sadorsky, Perry. 2012. Correlations and Volatility Spillovers between Oil Prices and the Stock Prices of Clean
Energy and Technology Companies. Energy Economics 34: 248–55. [CrossRef]
Sarwar, Suleman, Aviral Kumar Tiwari, and Cao Tingqiu. 2020. Analyzing Volatility Spillovers between Oil
Market and Asian Stock Markets. Resources Policy 66: 101608. [CrossRef]
Sharkasi, Adel, Heather J. Ruskin, and Martin Crane. 2005. Interrelationships among international stock market
indices: Europe, Asia and the Americas. International Journal of Theoretical and Applied Finance 8: 603–22.
[CrossRef]
Sharma, Gagan Deep, Mandeep Mahendru, and Sanjeet Singh. 2013. Are the Stock Exchanges of Emerging
Economies Inter-Linked: Evidence from BRICS. Indian Journal of Finance 7: 26–37.
Su, Xianfang. 2020. Measuring extreme risk spillovers across international stock markets: A quantile variance
decomposition analysis. The North American Journal of Economics and Finance 51: 101098. [CrossRef]
Syriopoulos, Theodore, Beljid Makram, and Adel Boubaker. 2015. Stock Market Volatility Spillovers and Portfolio
Hedging: BRICS and the Financial Crisis. International Review of Financial Analysis 39: 7–18. [CrossRef]
Taşdemir, Murat, and Abdullah Yalama. 2014. Volatility spillover effects in interregional equity markets: Empirical
evidence from Brazil and Turkey. Emerging Markets Finance and Trade 50: 190–202. [CrossRef]
Wang, Gang-Jin, Chi Xie, Min Lin, and H. Eugene Stanley. 2017. Stock Market Contagion during the Global
Financial Crisis: A Multiscale Approach. Finance Research Letters 22: 163–68. [CrossRef]
Wong, Wing-Keung, Hooi Lean, Michael McAleer, and Feng-Tse Tsai. 2018. Why Are Warrant Markets Sustained
in Taiwan but Not in China? Sustainability 10: 3748. [CrossRef]
Woo, Kai-Yin, Chulin Mai, Michael McAleer, and Wing-Keung Wong. 2020. Review on Efficiency and Anomalies
in Stock Markets. Economies 8: 20. [CrossRef]
Yang, Zihui, and Yinggang Zhou. 2017. Quantitative Easing and Volatility Spillovers Across Countries and Asset
Classes. Management Science 63: 333–54. [CrossRef]
Yousaf, Imran, and Junaid Ahmed. 2018. Mean and Volatility Spillover of the Latin American Stock Markets.
Journal of Business & Economics 10: 51–63.
Yousaf, Imran, and Arshad Hassan. 2019. Linkages between Crude Oil and Emerging Asian Stock Markets: New
Evidence from the Chinese Stock Market Crash. Finance Research Letters 31: 207–17. [CrossRef]
Yuan, Xinyu, Jiechen Tang, Wing-Keung Wong, and Songsak Sriboonchitta. 2020. Modeling Co-Movement among
Different Agricultural Commodity Markets: A Copula-GARCH Approach. Sustainability 12: 393. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

71
 Journal of
Risk and Financial
Management

Article
The Investment Home Bias with Peer Effect
Haim Levy
Department of Finance, Hebrew University, Jerusalem 9190401, Israel; mshlevy@[Link]

Received: 26 February 2020; Accepted: 7 May 2020; Published: 11 May 2020

Abstract: Observed international diversification implies an investment home bias (IHB). Can bivariate
preferences with a local domestic peer group rationalize the IHB? For example, it is argued that
wishing to have a large correlation with the Standard and Poor’s 500 stock index (S&P 500 stock
index) may induce an increase in the domestic investment weight by American investors and, hence,
rationalize the IHB. While this argument is valid in the mean-variance framework, employing bivariate
first-degree stochastic dominance (BFSD), we prove that this intuition is generally invalid. Counter
intuitively, employing “keeping up with the Joneses” (KUJ) preference with actual international data
even enhances the IHB phenomenon.

Keywords: investment home bias (IHB); bivariate first-degree stochastic dominance (BFSD); keeping
up with the Joneses (KUJ); correlation loving (CL)

JEL Classification: D81; C91

1. Introduction
The investment home bias (IHB) is well documented. For example, the US equity market accounts
for about 35% of the world equity market, yet about 75% of Americans’ equity investment is allocated
to the US market. Hence, the US equity IHB is in the magnitude of 40%. For most commonly employed
utility functions, the univariate expected utility maximization also does not support the relatively large
domestic investment weight; hence, the IHB puzzle emerges. It is advocated that the bivariate expected
utility maximization rationalizes partially or fully the IHB. In this study, we employ the “keeping
up with the Joneses” (KUJ) preference, where the investor’s wealth and the peer group’s wealth
are the two attributes of this utility function, analyzing the peer effect on the empirically observed
IHB phenomenon.
The basic idea and the intuition of the KUJ argument for rationalizing the IHB phenomenon
is as follows: suppose that you know your univariate utility function and that, for a given joint
distribution of returns corresponding to the various international markets, you derive with this utility
function the optimal investment weights in the domestic market as well as in the foreign markets
under consideration. Furthermore, suppose that the optimal domestic investment weight is, say, p%.
Now, suppose that you decide to consider, in addition to the joint distribution of returns, one more
factor: you also want the performance of your portfolio to be as close as possible to the performance
of a certain local stock index. For example, the American investor wants the return on her portfolio
to be as close as possible to the return on the S&P 500 index, which, for simplicity of the discussion,
is assumed to be the peer’s portfolio (the same analysis applies to any other local stock index). Thus,
if the investor benefits from having a relatively large correlation with the S&P index, she may have an
incentive to increase the domestic investment weight (which generally increases the correlation with
the S&P stock index, and if the domestic investment weight is 100%, this correlation is +1) beyond

JRFM 2020, 13, 94; doi:10.3390/jrfm13050094 73 [Link]/journal/jrfm

JRFM 2020, 13, 94

what is obtained by a maximization of a univariate expected utility function. Therefore, employing

KUJ preferences may rationalize the IHB.1
Indeed, Lauterbach and Reisman (2004) use the KUJ preference prove that the IHB is rationalized by
incorporating the peer effect. However, they use the mean-variance model with some approximations
to achieve this result. We analyze in this paper the impact of incorporating the peer effect on the IHB
in the most general bivariate expected utility case, not relying on the mean-variance framework and
where no approximations of the various mathematical formulas are employed. We define the precise
conditions, which guarantee the IHB rationalization, by adding the peer effect. We find that, in this
unrestricted analysis, the appealing intuitive explanation of the IHB rationalization by the peer effect
is generally wrong. Thus, we conclude that one should seek other economic explanations for the
observed IHB phenomenon. For some interesting economic suggestions, see Coeurdacier and Rey
(2013) and Berriel and Bhattarai (2013)2 or other behavioral explanations.
We employ in this study distribution-free bivariate first-degree stochastic dominance (BFSD),
with no assumptions on the shape of the bivariate preferences and no approximations. We prove that,
despite the above appealing intuition of the peer effect on the optimal domestic investment weight,
using the bivariate preferences, the IHB may increase or decrease relative to the univariate optimal
domestic investment weight. Moreover, we demonstrate with actual international data that adding the
peer effect, counter intuitively, even intensifies the IHB from the American investor’s point of view.
Hence, the IHB still exists.
The structure of the rest of this paper is as follows. Section 2 provides a brief literature review.
Section 3 presents bivariate first-degree stochastic dominance (BFSD) rule and the implied theoretical
results. We analyze the various factors affecting the IHB and show that bivariate preferences rationalize
the IHB phenomenon only in a limited and unrealistic case. Section 4 is devoted to the commonly
employed KUJ preferences, which is a specific set of all the bivariate preferences. We show empirically
that the peer effect with KUJ preferences even enhances the IHB puzzle. Section 5 concludes.

2. Literature Review
Vanpée and DeMoore (2012) show that the IHB exists in virtually all countries. The magnitude of
the IHB phenomenon is relatively large, characterizing various periods, assets, and countries. While
about three decades ago the American investment in the local market was more than 90%, implying a
very large IHB, in recent years the IHB phenomenon has been mitigated, yet it is still about 40%. When
it comes to fixed-income assets, the home bias is even larger. This phenomenon is not unique to the US
and characterizes many capital markets (for a report on the IHB in various countries, regarding equity
and fixed-income assets, see (Philips et al. 2012)). Actually, there is evidence that the home bias is even
worse than reported (see Baxter and Jermann 1997).
Researchers have analyzed various possible key explanations for the IHB. It is agreed that some
portion of the domestic overinvestment may be induced by international trade barriers, foreign exchange
risk, and regulation, as well as by a domestic peer group effect. However, with the increase in the
rapid flow of information and market efficiency observed over the last few decades, the trade barriers,
including possible asymmetrical information, have drastically declined. This may account for the
observed slight decrease in the domestic overinvestment phenomenon. However, since 1998, the equity
IHB of American investors has stabilized at about 40% (see Levy and Levy 2014).

1 Note, we analyze whether the peer effect increases the optimal domestic weight, which partially or fully rationalizes the IHB.
The reason is that it is possible that the peer effect increases the optimal domestic weight by, say, 1%, but the IHB is, say, 40%,
a case where other factors are needed to explain the observed IHB. In our study, we find empirically that the peer effect even
enhanced the IHB; hence, the distinction between partial and full IHB rationalization is irrelevant.
2 They consider portfolio diversification when macroeconomic factors are incorporated into a two-country general equilibrium
model, called the “Open Economy Financial Macroeconomics” model. They conclude that, with this equilibrium model,
the home bias is less of a puzzle. Berriel and Bhattarai (2013) also suggest a macroeconomic model (related to the positive
association between government spending and return on local stocks) to explain the home bias.

74
JRFM 2020, 13, 94

While most empirical studies analyze the IHB at the country level (see French and Poterba 1991;
Tesar and Werner 1995), Kang and Stultz (1997), who study the IHB puzzle in Japan, analyze it at the
individual firm level, showing that foreign investors hold disproportionally more Japanese shares
of firms in the manufacturing industries, large firms, and firms with good accounting performance.
Similarly, Dahlquist and Robertsson (2001) identify the characteristics of Swedish firms that attract
foreign investors. Lewis (1999), who analyzes the effect of each economic factor that is considered
as a barrier for efficient international diversification on the IHB, concludes that the trade barriers
cannot explain the magnitude of the existing IHB. Therefore, the IHB puzzle is still an interesting
research topic.3
Obviously, if the IHB does not incur economic loss, it does not constitute an economic puzzle.
Indeed, the intensity of the IHB economic cost changes over time. Levy (2016) analyzes the trend
in the IHB phenomenon over time. Moreover, he distinguishes between the economic home bias
(EHB), which measures the economic loss in terms of the differences in the certainty equivalent of
two alternative international diversification strategies (with and without a home bias) and the IHB,
which simply measures the deviations between the optimal international investment weights and
the actual investment weights. He reports that, while the EHB was very large in the past, in the
last 15 years, the EHB from the American investment point of view has become negligible, despite
the existence of about 40% IHB. This reduction in the EHB is induced by the increasing trend in
the international correlations. Thus, it seems that for the American investors the IHB is not a major
economic puzzle. However, he also reports that for other countries, e.g., France, the EHB is still very
large, and the economic puzzle exists. Moreover, in recent years, we have trend reversal in correlations,
and a decrease in the average correlation between various markets has been recorded. As a result of this
trend reversal, the EHB has recently increased, even for American investors. Thus, for most countries
and with the recent trend reversal in correlation also for the US, the IHB still constitutes an economic
puzzle that needs an explanation. The employment of the KUJ preference, namely incorporation of the
peer effect, is considered as one of the promising paths in explaining the IHB puzzle.
We employ in this paper a bivariate preference. Generally, with bivariate preference, the two
variables can take many forms, e.g., wealth and health, climate and income, etc. Our study deals
with investment choices. Hence, the two variables are the individual’s wealth and the peer group’s
wealth. The peer group’s wealth can be the return on a certain domestic portfolio, and in our case,
as mentioned above, we assume, for the simplicity of the discussion and without loss of generality,
that it is the return on S&P 500 stock index.
The common view is that the relevant bivariate utility function has a positive cross derivative
(we will elaborate on this issue below) and that investors want, among other things, the performance
of their portfolio to be as close as possible to the performance of the peer’s portfolio, i.e., a large
correlation with the S&P stock index is desired.4 Therefore, we focus our analysis on the positive
cross-derivative case. Obviously, despite the desire for having a relatively large correlation with the
S&P index, the investor will shift from a portfolio with a small correlation to a portfolio with a large
correlation, only if the bivariate expected utility increases by such a shift. We turn to analyze the
conditions under which indeed such shift takes place, namely that the IHB can be rationalized with the
peer effect.

3 It is interesting to note that, even in a case in which there are no transparent trade barriers, there is a tendency to invest in
firms that are geographically located close to the investor’s location. This phenomenon is well documented within the US
(see Coval and Moskowitz 1999, 2001; Huberman 2001). This indicates that the home bias is a complex phenomenon that is
not easy to explain with conventional economic factors.
4 Tsetlin and Winkler (2009) advocate that correlation aversion prevails. However, in their model, the two attributes of the
bivariate preference directly affect the utility of the decision maker, for example, income and quality of life. In our model,
the two attributes are different: the individual’s wealth and the peer group’s wealth. As relative wealth may affect the
individual’s utility, it is advocated in the literature that, when some conditions hold, correlation loving prevails.

75
JRFM 2020, 13, 94

3. Bivariate First-Degree Stochastic Dominance (BFSD) and the IHB

We would like to stress at the outset that most of the mathematical formulas given in the first
part of this section are not new and exist in the literature, albeit in different forms and in different
connotations. However, we use these mathematical results, to the best of our knowledge for the first
time to analyze the peer effect with KUJ preferences on the IHB phenomenon.

3.1. The Sufficient Conditions for BFSD Implying the IHB Rationalization
Consider an individual with a bivariate preference U (w, wP ), where w denotes the return on the
selected international portfolio by the investor under consideration, and wp denotes the return on
the peer’s portfolio. We compare two bivariate investment portfolios, F and G, where the domestic
investment weight in portfolio F is larger than the domestic investment weight in portfolio G (we will
elaborate later on the selected portfolios, F and G). Our aim is to examine the conditions under which F
dominates G by BFSD with  the above bivariate utility function, where we
 firstassume two assumptions
on the preferences: ∂U w, wp /∂w ≡ U1 ≥ 0 (monotonicity) and ∂2 U w, wp /∂w∂wp ≡ U12 ≥ 0 (later
on we consider also U12 ≤ 0, a case usually not considered in KUJ economic
  research but emerges as
important to our analysis). There is no constraint on the derivative ∂U w, wp /∂wp ≡ U2 , which can be
negative, zero, or positive.5 If such dominance exists, then all investors, regardless of the precise shape
of the bivariate preference, will switch from G to F. Hence, the optimal domestic investment weight
increases, and therefore the peer effect rationalizes the IHB phenomenon.
Note that the main ingredient of the KUJ preference is that the cross derivative (U12 ) is positive,
implying that the individual’s marginal utility increases with an increase in the peer group wealth
(see Ljungqvist and Uhlig 2000).6 Therefore, as explained before, it seems that the investor with a
positive cross derivative would incline to overinvest domestically, as she prefers her wealth to be
positively correlated with the peer’s wealth. While the above intuitive explanation is appealing, in the
following proposition, it is formally shown that generally only under some specific conditions, indeed
a positive cross derivative is tantamount to correlation loving, where correlation loving implies that,
by increasing the domestic investment weight, the bivariate expected utility increases. Namely, if the
conditions required in the proposition are intact, the investor increases her bivariate expected utility by
overinvesting domestically (relative to the optimal univariate expected utility maximization optimal
domestic investment weight), and by doing so, the correlation increases. Thus, if the proposition
required conditions hold in practice, we have by the KUJ preferences a rationalization of the IHB,
and the IHB puzzle may vanish. As we explain below, in practice, the required conditions for IHB
rationalization are not intact. Before stating the proposition, we need the following definition:

Definition 1. Definition of correlation loving (CL): The investor is CL if and only if, by increasing the correlation
between her portfolio and the peer’s portfolio, the expected bivariate utility increases.

Hence, CR investors who maximize the bivariate expected utility would increase the domestic
investment weight relative to the optimal univariate expected utility weight.

5 Note that a negative sign implies jealousy, and a positive sign implies altruism (see Dupor and Liu 2003).
6 Numerous studies suggest replacing the univariate expected utility analysis with the expected bivariate utility analysis with
various definitions of the two variables: past and present consumption, consumption of the individual, and consumption of
the peer group, the wealth obtained by the individual and the opponent in an ultimatum game, and so forth. For studies
that assume that the utility is derived not from the absolute wealth (or consumption) of the individual but from the relative
wealth (or consumption), in which the wealth’s position relative to the peer group plays an important role, as well as for
other factors that do not affect the classic univariate expected utility but affect the bivariate expected utility, see, for example,
Abel (1990), Constantinides (1990), Bolton (1991), Rabin (1993, 1998), Galí (1994), Campbell and Cochrane (1999), Bolton and
Ockenfels (2000), Dupor and Liu (2003), Zizzo (2003), and Demarzo et al. (2008).

76
JRFM 2020, 13, 94

Proposition 1. Suppose that the investor faces two alternate bivariate prospects, F(w, wp ) and G(w, wp ),
where w as well as wp can take only two different outcomes. As there are only two outcomes, they can be
rearranged to have either a correlation of +1 or a correlation of −1. Diversification between w and wp is not
allowed, implying that the marginal distributions are identical (namely, Fw = Gw and FwP = GwP , regardless
of the outcomes arrangement; see, for example, Table 1). Under these specific conditions, the investor with a
bivariate preference is CL if and only if the cross derivative is positive, namely U12 ≥ 0. Specifically, under the
conditions of the proposition, with CL, the prospect with a correlation of +1 yields a higher bivariate expected
utility than any other possible prospect. (For proof, with some other notation, see (Eeckhoudt et al. 2007)).

Thus, if the conditions of the proposition were intact, the American investor who likes her
investment performance to be as close as possible to the S&P index would have a higher expected utility
by increasing the domestic investment weight. Actually, under the conditions of the proposition, having
a correlation of +1 with the S&P index is optimal, implying that investing 100% domestically is optimal,
which creates a negative IHB puzzle (because in practice less than 100% is invested domestically).
In short, if the conditions of Proposition 1 are intact, we have:

CL ⇔ U12 > 0 (1)

Note that investing more intensively domestically, hence increasing the correlation between the
investor’s portfolio and the peer’s portfolio, generally does not imply CR as defined above. The reason
is that, with investment in practice, by increasing the domestic investment weight, although the
correlation increases, generally, other parameters of the portfolio may also change, the marginal
distributions may change (hence, the conditions of the proposition are violated), and the bivariate
expected utility may decrease.
Therefore, the American investor may decide not to decrease the domestic investment weight,
despite the desire to have large correlation with the S&P index. However, by the above definition,
the investor is CL only if, after considering all effects, the bivariate expected utility increases.
However, note that, by Proposition 1, the marginal distributions are kept unchanged, and the
correlation can take only the extreme values of either +1 or −1. This is because in Eeckhoudt et al. (2007)
original proposition, each variable can get only two possible values. Hence, by reordering these values,
the marginal distributions are kept unchanged. Also, diversification between w and wp is not allowed,
because if it is allowed, the marginal distribution of the individual’s wealth, w, generally will not
be kept constant. Thus, the statement given in Proposition 1 is suitable to some choices, where the
variables are, for example, wealth and health, with only two outcomes (say, bad and good health,
high and low income, etc.). As we shall see below, with international diversification, we have more
than two outcomes corresponding to each prospect, and diversification is allowed. Hence, the marginal
distributions generally change when the selected diversification changes. Therefore, a positive cross
derivative in our analysis does not necessarily imply CL. As a result, we may even obtain an IHB
phenomenon enhanced with bivariate preferences relative to the univariate IHB, despite the fact that a
positive cross derivative is assumed.
Let us turn now to the conditions for BFSD of the distribution of returns of the portfolio with
the IHB over the distribution of returns with no IHB. The two portfolios that we compare, F and G,
have bivariate density functions, denoted by f (w, wP ) and g(w, wP ), respectively. As we focus on
the possible IHB rationalization, it is assumed, as explained before, that F stands for a portfolio with
an IHB, that is, the domestic weight in this portfolio is larger than the corresponding weight in G.
Thus, if the domestic investment weight in G is equal to the optimal theoretical univariate expected
utility maximization domestic weight (say, the international market portfolio), the BFSD of F over G
implies that the peer effect rationalizes the IHB phenomenon, as all investors would prefer F over

77
JRFM 2020, 13, 94

 
G.7 Assuming that ∂2 w, wp /∂w∂wp ) ≡ U12 > 0 with KUJ preference8 to explain various observed
economic phenomena is very common. As seen in Proposition 1, this assumption is an important
ingredient also needed to rationalize the IHB phenomenon, so long as the conditions of Proposition 1
hold. Therefore, we examine the role of the cross derivative on the BFSD relation. To examine possible
rationalization of the observed IHB with KUJ preferences, we extend the expected utility univariate
analysis to the bivariate expected utility analysis by adding the peer effect.
The expected bivariate utility of portfolios F and G is given by:
 w  wp  
EF U = w wp
U (w, wp ) f w, wp dwdwp
 w  wp   (2)
EG U = w wp
U (w, wp ) g w, wp dwdwp

where w and w denote the minimal and maximal values of w (which can be −∞ and ∞); similarly, wP
and wP denote the minimal and maximal values of wP . Thus,
 w  wp     
Δi ≡ EF U − EG U = U (w, wp ) f w, wp − g w, wp dwdwp
w wp

Integrating by parts the above equation with respect to both variables yields:
 wp  w w
Δi ≡ EF U − EG U = w w U12 [F(w, wp ) − G(w, wp )]dwdwp + w U1 [G(w)−
 wp p

F(w)]dw + w U2 [G(wp ) − F(wp )]dwp (3)

p
≡ A+B+C

where Δi denotes the expected utility difference corresponding

  the ith investor, F(w, wP ) and G(w, wP )
to
are the two bivariate cumulative distributions, F(w) = F w, wp is the marginal cumulative distribution
   
function of w, F wp = F w, wp is the marginal cumulative distribution function of wP , and U1 , U2 ,
and U12 denote the partial derivatives: U1 ≡ ∂U/∂w, U2 ≡ ∂U/∂wP , and U12 ≡ ∂2 U/∂w∂wP ,
respectively. For the derivation of Equation (3) with slightly different notations, see Levy and Paroush
(1974, p. 131) and Atkinson and Bourguignon (1982, pp. 185–86).9 Note that, as the marginal utility of
the peer’s portfolio is identical under the various
  investment
  strategies (with and without intensive
domestic investment). Namely, we have G wp = F wp , therefore term C in Equation (3) is equal to
zero. Thus, the rest of the paper relate only to terms A and B.
Let us first analyze the relation between Equation (3) (with C = 0) and the conditions given
in Proposition 1. If each of the two random variables, w and wp , has only two possible different
outcomes, the correlation is either +1 or −1. Also, when diversification between w and wp is not
allowed, the marginal distributions are equal (namely, G(w) = F(w), see also the example given in

7 Obviously, we have a different optimum portfolio for each utility function, but, as we shall see below, the analysis is intact,
independent of the assumed preference.
8 The KUJ and CUJ literature is very extensive; hence, we mention here only a few of these studies. Abel (1990) and Galí (1994)
use this bivariate framework to explain optimal choices. Ljungqvist and Uhlig (2000) examine the role of tax policies in
economics with CUJ utility functions. Campbell and Cochrane (1999) assume that the preference is a function of the relative
consumption, when the individual’s consumption is measured relative to the weighted average of the past consumption
of all individuals. In these models, when the peer group’s variable (e.g., consumption) is a lagged variable, the model is
commonly called the CUJ model, and when the individual’s variable and the peer group variable relate to the same time
period (e.g., return on investment), it is commonly called the KUJ model. In this paper, we analyze the optimal portfolio
investment decision in the KUJ set-up.
9 Note that Equation (2) is reduced to the well-known univariate formula employed to derive the FSD rule, where U12 = U2 = 0.
For more details, see Hadar and Russell (1969) and Hanoch and Levy (1969). Although we focus in this paper on FSD,
one can assume risk aversion and employ stronger investment rules; for example, see Rothschild and Stiglitz (1970) and
Levy (2015).

78
JRFM 2020, 13, 94

Table 1. Hence, in this specific case also term B is equal to zero, and we are left with term A. If F
 the +1 correlation
represents
  and G the −1 correlation, we must have with the two outcomes case that
F w, wp ≥ G w, wp (see example 1 in Table 1. Hence, in this case by Equation (3), with B = C = 0, the
condition U12 ≥ 0 is a sufficient condition for dominance of the joint distribution with the +1 correlation
over the joint distribution with the −1 correlation (see Equation (3)).10 It is easy to verify that, in this
specific case, U12 ≥ 0 is a necessary and sufficient condition for dominance.11 Thus, Equation (3) is
perfectly consistent with Proposition 1, so long as the conditions given in the proposition are intact.
However, Equation (3) corresponds to the general case, as it covers the more realistic scenarios where
more than two outcomes are possible. Diversification is allowed, and the marginal distributions are
not necessarily equal, hence term B is not necessarily equal to zero. As we shall see in this general and
realistic case, U12 ≥ 0 is neither a necessary nor a sufficient condition for dominance. We turn now to
analyze the possible dominance of the portfolio with the IHB over a portfolio with no IHB in the most
general case.
To examine possible rationalization of the IHB phenomenon with bivariate preferences, let us first
take a deeper look at the marginal distributions corresponding to the international diversification issue
analyzed in this paper. Returning to Equation (3), note that, as mentioned above, it is reasonable to
assume that the third term on the right-hand side of Equation (3), term C, is equal to zero, as the investor
in the capital market generally cannot affect the peer group investment decision; hence, the peer’s
group marginal distribution is identical under F and G. This condition conforms to the requirement in
Proposition 1, even in the case where more than two outcomes exist. This is a reasonable assumption
with the investment choices that we analyze in this study but not with ultimatum games in which
the individual decision affects the opponent’s outcome. Moreover, in the portfolio investment case,
this term is equal to zero, regardless of whether the peer group portfolio is domestic or international.
Thus, regarding the issue that we investigate in this paper (investment with a particular stock index as
the peer group’s portfolio), as advocated above, the sign of the derivative U2 is irrelevant. Namely,
term C = 0 and there is no need to assume jealousy (U2 < 0) or altruism (U2 > 0) to obtain our results
corresponding to the portfolio investment case. Thus, as for the analysis of the IHB, term C is equal to
zero, and Equation (3) is reduced to:
Δi = A + B. (4)

However, note that generally we cannot assume that also term B is equal to zero, as by changing
the diversification strategy, we change the marginal distribution of the individual’s wealth. Thus,
with international portfolio diversification, the condition of equal marginal distributions (see term B of
Equation (3)) required by Proposition 1 does not hold.
To be able to determine whether the peer effect induces an increase in the optimal domestic
investment relative to the univariate expected utility optimal domestic investment weight, we need to
be more specific regarding the definitions of portfolios F and G under consideration. We examine here
the possible existence of BFSD by considering the two specific portfolios with direct implication to the
IHB issue analyzed in this paper. These two portfolios are denoted by FA and GM , as defined below.

Definition 2. GM is the portfolio with the international market weights. If the American investor holds this
market portfolio, she would invest 35% (which is the weight of the American market in the world market)
domestically; hence, the IHB does not exist. Distribution FA stands for the actual aggregate portfolio held by the
American investors. Namely, the actual domestic weight held by the American investor is 75%; hence, holding
this portfolio implies an IHB of 40%.

10 Actually, it is required to have at least one strict inequality with the distribution functions as well as with the cross derivative
to avoid the trivial case of having Δi = 0. In the rest of the paper, when we write such inequalities, we always mean that
there is at least one strict inequality, but to avoid a complex writing, we will not write it down everywhere.
11 If U12 < 0, in some range, one can always find a bivariate preference, such that outside this range the cross derivative is close
to zero; hence, Δi is negative. Therefore, to guarantee that Δi is non-negative, the cross derivative cannot be negative.

79
JRFM 2020, 13, 94

Assuming that term C is equal to zero, and rewriting Equation (4) in terms of the above two
portfolios, we obtain:
 wp  w     
Δi ≡ EFA U − EGM U = w w U12 FA w, wp − GM w, wp dwdwp
w p
(5)
+ w U1 [GM (w) − FA (w)]dw ≡ A + B

Suppose that without the peer effect the market portfolio is optimal. If Δi > 0, the ith investor
under consideration who considers also the peer effect prefers the actual portfolio to the market
portfolio; hence, the investor increases the bivariate expected utility by increasing the domestic
investment weight. However, to have BFSD and IHB rationalization, we need to have that Δi > 0 for
all investors i = 1, 2, . . . n, regardless of the precise shape of their preferences. A few conclusions,
some of them in contradiction to the common view regarding the role of the cross derivative, can be
drawn from Equation (5).
First, if U12 ≥ 0 is assumed (as needed in Proposition 1 to justify the rationalization of the IHB),
we find that there is no BFSD; hence, in this setting there is no IHB rationalization.
  The reason is
that if U12 ≥ 0, term A of Equation (5) is positive only if FA w, wp ≥ GM w, wp , but this implies that
FA (w, ∞) = FA (w) ≥ G(w, ∞) = G(w), and therefore, term B is negative. The sum A + B may be
negative, implying that there is no BFSD.
Surprisingly, in contrast to Proposition 1, the condition U12 ≤ 0 may allow BFSD; hence, it may
allow IHB rationalization. We have BFSD and IHB rationalization with U12 ≤ 0 if the following two
conditions hold:    
FA w, wp ≤ GM w, wp (6a)

FA (w) ≤ GM (w) (6b)

But as condition (a) implies condition (b), unlike the positive cross-derivative case, these two
conditions can simultaneously hold. Therefore, if condition (a) on the joint distribution holds, both terms
A and B are positive (with U12 < 0) and therefore Δi ≥ 0 for i = 1, 2, . . . n.

Example 1. The marginal distributions and the BFSD.

In this example, we demonstrate the relation between the BFSD and the positive cross derivative
in the case where the conditions of Proposition 1 are intact, and then we demonstrate the more realistic
case, where the marginal distributions are not kept constant; hence, BFSD does not exist, despite the
positive cross-derivative assumption.
Suppose that the S&P index return wp is equal to 3 or 4, each outcome with an equal probability
of 0.5. We consider investing in either portfolio F or portfolio G, both yielding return w of either 2
or 5 with equal probability of 0.5. However, F has a correlation of +1 with the S&P index (with joint
returns of (2, 3) with a probability of 0.5 and joint returns of (5, 4) with a probability of 0.5). G has a
negative correlation of −1 with the S&P index with joint returns of (2, 4) with a probability of 0.5 and
joint returns (5, 3) with a probability of 0.5 (see Table 1). All other joint probabilities are equal to zero.
Denoting the joint distribution corresponding to the correlation +1 by FA and the joint distribution
corresponding to correlation −1 by GM , we have with the above example with the joint probabilities
the following relationship:
     
FA w, wp ≥ GM w, wp for all values w, wp (7)

with at least one strict inequality (see lower part of Table 1 Part a), e.g.,

FA (2, 3) = 0.5 > GM (2, 3) = 0,

80
JRFM 2020, 13, 94

and it is easy to verify that with the marginal distributions, we have:

GM (w) = FA (w)

and    
GM wp = FA wp

for all possible values(all marginal cumulative distributions get the values of 0.5 at the lower outcome
and 1 at the larger outcome). Therefore, the conditions of Proposition 1 are intact and terms B and C of
Equation (3) are equal to zero and we are left only with term A; hence, if the cross derivative is positive,
the joint distribution yielding (2, 3) and (5, 4) dominates the joint distribution (2, 4) and (5, 3) for all
bivariate preferences with a positive cross derivative. Thus, term A of Equation (3) is positive, and term
B is equal to zero; hence, we have BFSD of the joint distribution with correlation +1 over the joint
distribution with correlation −1. So far, this example conforms to the conditions given in Proposition
1 and provides the IHB rationalization by adding the peer effect, so long as the cross derivative is
positive. Let’s turn to another example, where one of the conditions of Proposition 1 is violated.

Table 1. Joint Probability and Joint Cumulative Probability Functions Corresponding to the two
Examples Given in the Text.

Correlation +1 Correlation −1
Part (a): The First Example Given in the Text
The Probability Functions
FA GM
w\wp 3 4 w\wp 3 4
2 0.5 0 2 0 0.5
5 0 0.5 5 0.5 0
The Cumulative Bivariate Probability Functions
w/wp 3 4 w/wp 3 4
2 0.5 0.5 2 0 0.5
5 0.5 1 5 0.5 1
Part (b): The Second Example Given in the Text
The probability Functions
w/wp 3 4 w/wp 3 4
2 0.5 0 2 0 0.5
5 0 0.5 10 0.5 0
The Cumulative Bivariate Probability Functions
w/wp 3 4 w/wp 3 4
2 0.5 0.5 2 0 0.5
5 0.5 1 10 0.5 1

Make now the following change in the previous example: the outcomes with the −1 correlation
are 2 and 10 with equal probability of 0.5 rather than 2 and 5 as we have in the previous example. Thus,
we simply replaced the outcome 5 by outcome 10 (see Table 1 Part b). All the other outcomes are kept
unchanged. Thus, under the choice with a correlation of −1, we have the joint outcomes of (2, 4) or
(10, 3) with an equal probability of 0.5, and with the choice corresponding to correlation +1, we have as
before the joint outcomes (2, 3) or the outcomes (5, 4) with an equal probability of 0.5. It is easy to
verify that as before, also with this change we have:

81
JRFM 2020, 13, 94

   
FA w, wp ≥ GM w, wp (8)

with at least one strict inequality, e.g.,

FA (2, 3) = 0.5 > GM (2, 3) = 0.

Hence, if the cross derivative is positive, term A of Equation (5) is also positive. However, with this
change, the marginal distribution of the individual wealth also changes, and as cumulative distribution
of (2, 10) is located to the right of the cumulative distribution of (2, 5), namely, GM (w) ≤ FA (w) for all
values w, and there is at least one strict inequality, e.g., GM (w = 5) = 0.5 > FA (w = 5) = 1, see Table 1
Part b (this means that the univariate prospect A dominates prospect M by FSD12 ); hence, term B of
Equation (5) is negative. Therefore, A + B may be positive, zero, or negative, and we do not have
BFSD, and as a result, the IHB cannot be rationalized in this case, despite the assumed positive cross
derivative. Actually, it is easy to find a specific bivariate preference with U12 → 0 and U1 very large
such that A + B < 0, implying that F does not dominate G.
Finally, even if we stick to the original set of numbers by diversification of say, 0.5 in each asset
(w and wp ), the marginal distribution of the individual’s wealth becomes either 2.5 = 0.5 × 2 + 0.5 × 3
or 4.5 = 0.5 × 5 + 0.5 × 4 (with equal probability) in the case of a correlation of +1, and either 3 = 0.5 × 2
+ 0.5 × 4 or 4 = 0.5 × 5 + 0.5 × 3 (with an equal probability) in the case of correlation of −1 (see Table 1
Part a). Thus, allowing diversification between the two variables the marginal distribution of w is
not kept constant, and the conditions required by Proposition 1 are violated, and once again, we do
not have BSFD, even where the cross derivative is positive. Moreover, by changing the investment
weights, we change the marginal distribution of w. Hence, we do not have BFSD of the +1 correlation
joint distribution over the −1 correlation joint distribution.
In sum, with the first example, we have IHB rationalization if diversification is not allowed. In the
second example, we do not have IHB, even if diversification is not allowed, let alone if it is allowed.
We show below that the case given in Table 1 Part b, where the marginal distributions are not kept
constant, conforms to actual stock market data; therefore, KUJ with a positive cross derivative does not
rationalize empirically the IHB phenomenon.

3.2. Discussion
We analyze above the dominance condition between two specific portfolios, one with an IHB and
one without an IHB. If indeed these are the two portfolios considered by all investors, assuming a
positive cross derivative, we do not have dominance of the portfolio with the IHB over the portfolio
with no IHB. Thus, a positive cross derivative is not a sufficient condition for BFSD. However, with a
negative cross derivative, we may have dominance and, therefore, may obtain IHB rationalization
by incorporating the peer effect. The existence of such IHB rationalization depends on the joint
distribution of the investor’s selected portfolio and the peer’s portfolio. If all investors consider
the same two  portfolios
 (the
 market
 portfolio and the actual portfolio defined above), and if the
condition FA w, wp ≤ GM w, wp holds, we have IHB rationalization, so long as the cross derivative is
negative. However, generally there is no reason to assume that with actual data this condition is intact.
Therefore, we generally do not have IHB rationalization with the above two portfolios, regardless of
the sign of the cross derivative.

12 Generally, if F(x) ≤ G(x) for all values x and there is at least one strict inequality, we say that F dominates G by first degree
stochastic dominance (FSD).

82
JRFM 2020, 13, 94

In the above analysis, we consider two specific portfolios. In practice, not all investors hold
these two portfolios, and each investor has her optimal univariate portfolio and her optimal bivariate
portfolio. In this realistic case, only if we have for all investors the optimal domestic investment
weight whereby the bivariate utility function is larger than the optimal domestic investment weight
corresponding to the univariate framework, we have an IHB rationalization. As such case cannot be
analyzed without the information on the preferences of all investors, we analyze below the peer effect
on the IHB with the commonly employed KUJ preferences.

4. The IHB with Some Specific KUJ Preferences

We proved above that unless the marginal distributions are kept unchanged with a positive cross
derivative, we do not have BFSD, and with a negative cross derivative, the required condition is
unlikely to hold; therefore, generally we do not have IHB rationalization, regardless of the sign of the
cross derivative. However, it is possible that, despite not having dominance which corresponds to all
possible bivariate preferences, for some important and commonly employed KUJ preferences (but not
for all bivariate preferences, as we do not have BFSD), with various risk aversion parameters, the peer
effect may induce an increase in the domestic investment weight relative to the univariate optimal
domestic investment weight. As we shall see below, this is not the case and also the KUJ preferences
empirically do no rationalize the IHB.
The procedure for testing the peer effect with KUJ preferences is as follows: We first solve with
actual data for the optimum diversification with a univariate preference for various degrees of risk
aversion and then solve for the optimum diversification with a KUJ preference with a similar form of
the univariate preference but with the incorporation of the peer effect. Using this two-step analysis,
we measure the marginal peer group effect on the optimal domestic investment weights. It is shown
below empirically that, with the commonly employed KUJ preference with a positive cross derivative,
adding the peer group effect does not rationalize the home bias phenomenon. Moreover, the optimal
domestic investment weight with the KUJ preference (with a domestic peer group) is found to be
even smaller than the optimal domestic investment weight with no peer group effect. Thus, despite
the assumed positive cross derivative, which under some conditions implies CL, we find empirically
that, counter intuitively, the IHB phenomenon is even enhanced when the peer effect is incorporated.
Thus, shifting from univariate utility function to the bivariate utility function with peer effect does not
rationalize the IHB and other explanations are called for. Let us elaborate.

4.1. The Optimal Diversification with a Univariate Utility

The empirical analysis given in this section is done from the American investor’s point of view.
Thus, all the returns are in dollar terms. Similar analysis can be done from other points of view.
With a univariate CRRA utility function, we have:

U (w) = [w1−α /(1 − α)], withα > 0. (9)

With n available international assets, we solve for the vector of the investment proportions p
(where 0 ≤ pi ≤ 1), which maximizes the following expected utility:


n−1
E{[(pUS yUS + p2 y2 + . . . .(1 − pn ) yn )1−α /(1 − α)]} (10)
i=1

where yUS is the rate of return on the US index (a random variable), and yi is the return on foreign
market i, where i = 2, 3, . . . . . n, and there are n stock indices, the US index plus n − 1 foreign indices.
Let us specify the corresponding bivariate preferences employed in this study.

83
JRFM 2020, 13, 94

4.2. The KUJ Preferences (for α > 1)

We discuss below various bivariate functions with a positive cross-derivative, which are natural
extensions of the myopic univariate reference. We first employ the bivariate utility suggested by
Abel (1990). As this function has a positive cross derivative only for α > 1, we also employ a slightly
different preference that conforms to the KUJ requirement for the whole range of the risk aversion
parameter (α > 0). Abel suggests a general bivariate formula that, under some conditions for the
various parameters, reduces to:

γ 1−α
U (w, wP ) = [w/wP ] /(1 − α) whereα > 0 andγ ≥ 0. (11)

With γ = 0 this function is reduced to the univariate CRRA function. Hence, this is a neutral
extension to the univariate myopic function. Thus, this bivariate preference is a natural extension of the
univariate CRRA function, U (w) = [w1−α /(1 − α)], which is commonly employed in economic research
(see (Merton and Samuelson 1974)) to the bivariate case, in which the peer group effect is incorporated
into the analysis. Note also that, when wp is constant, Equation (11) collapses to Equation (9). As we
wish to have monotonicity with regard to w and a positive cross derivative (because with positive cross
derivative there is an appealing intuition for an increase in the domestic investment weight), let us
examine these two derivatives. Monotonicity always exists with this function, because:

γ −α −γ −γ(1−α)
U1 = (w/wP ) wP = w−α wP ≥ 0.

The cross derivative, which is given by:

−γ(1−α)−1
U12 = −γ(1 − α)w−α wP ,

is positive only for α > 1. Thus, we have a KUJ preference with a positive cross-correlation for α > 1.
It is worth mentioning that with this function we also have jealousy, namely, U2 = ∂U/∂wp < 0.
Generally, in economics, the range of the risk aversion parameter is α ≥ 0. To avoid the above
constraint on α (namely α > 1), we also employ the following function:

1−β
U (w, wP ) = [w1−α /(1 − α)][wP /(1 − β)] withβ < 1. (12)

With this preference, the cross derivative is positive for all values α > 0, as required by the
myopic preference. Once again, when wp is constant, this function collapses to the univariate myopic
preference (multiplied by a positive constant). The partial derivative with respect to w of this bivariate
1−β
function, given by ∂U/∂w ≡ U1 = w−α wP /(1 − β) ≥ 0, is positive only for β ≤ 1 and for any value α.
As we assume monotonicity in the numerical solution to the optimum investment, we take the β ≤ 1
constraint into account.
The cross-derivative is positive for the whole range of relevant parameters:

−β
∂2 U/∂w∂wP ≡ U12 = w−α wP > 0.

It is easy to see that, with this function, U2 may be negative, zero, or positive, depending on
whether α is larger than, equal to, or smaller than 1. Therefore, with this function, we allow for jealousy
as well as altruism.
Let us write down the equations employed in the derivation of the optimal investment weights with
the various preferences. With n available international assets, we solve for the vector p (where 0 ≤ pi ≤ 1),
which maximizes the expected utility. For the preference suggested by Abel (see Equation (11)), we have:


n−1 1−α
γ
E{[pus yus + p2 y2 + . . . .(1 − pn ) yn ]/(wP ) /(1 − α)]} (13)
i=1

84
JRFM 2020, 13, 94

and with the KUJ preference given by Equation (12) we have:


n−1
E{[(pus yus + p2 y2 + . . . .(1 − pn ) yn )1−α /(1 − α)][ yus 1−β /(1 − β)]}, (14)
i=1

where yUS is the return on the US index (a random variable), and yi is the return on foreign market i,
where i = 2, 3, . . . . . n, and there are n stock indices, the US index plus n − 1 foreign indices.

4.3. Data and Results

The purpose of the empirical analysis given below is to examine the peer effect with some specific
preferences on the IHB phenomena with actual data in the case where BFSD does not necessarily exist.
It is not intended to find the optimal investment portfolio for investment, a case where statistical
adjustment should be made in the empirical distribution, which is very noisy
We employ the 11 countries’ actual ex-post joint distribution to analyze the peer group effect on the
domestic investment weight from the American investor’s point of view.13 Obviously, the empirical
distributions of return of all countries are not identical, and the marginal distribution of the various
diversified portfolios under consideration are not identical; hence, the conditions of Proposition 1
are not intact; hence, the BFSD does not exist with positive cross derivative. Yet, it is possible that,
with some specific bivariate preferences, the domestic investment weight increases. Thus, we examine
whether the above two specific KUJ preferences with a positive cross derivative explain the IHB,
despite the fact that we do not have BFSD. We use in the empirical study annual rates of return taken
from the Bloomberg MSCI stock indices.
We report the optimum investment weights corresponding to Abel’s bivariate preference function
(see Equation (13)). The results corresponding to the bivariate function given by Equation (14) are very
similar, so for brevity sake they are not reported here in detail. Both forms of the bivariate functions
reveal the same surprising results: the commonly employed KUJ preference with a positive cross
derivative induces a decrease rather than an increase in the US investment weight. Thus, the adding
the peer effect does not rationalize the IHB.
Table 2 reports the results obtained using Abel’s preference (see Equation (13)). The table employs
the actual 25-year historical data (see Appendix A for the annual rates of return of the 11 countries for
the years 1988–2012 and Appendix B for the correlation matrix). Recalling that, for γ = 0, the bivariate
function is reduced to the univariate CRRA function, we see that, by adding the KUJ peer group effect
with the domestic peer group, the US investment weight decreases quite sharply.

Table 2. The US optimal investment weights in the US market with Abel’s bivariate preferences with
empirical data for the period 1988–2012.

α
(CRRA Univariate Preference γ = 0) 0.5 1 2
γ
1* 0.00 0.00 0.00 0.00
2 0.49 0.39 0.29 0.13
5 0.95 0.82 0.72 0.57
* Only for α > 1 do we have a KUJ preference with U12 > 0. We also have with this function U1 > 0 (monotonicity)
and U2 < 0 (jealousy).

13 We are aware of the large potential statistical errors involved in the derivation of the optimal investment weight with
historical data (see Britten-Jones 1999; Levy and Roll 2010). However, the goal of this empirical analysis is not to derive the
optimal investment weights for ex-ante investment purposes but rather to demonstrate that, with empirical data covering 11
international markets and 25 years, it is possible that, with some commonly employed KUJ preferences with a positive cross
derivative, the peer group effect induces a decrease rather than an increase in the domestic investment weight, which is in
contradiction to the equal marginal distribution case and to the economic intuition.

85
JRFM 2020, 13, 94

Specifically, with the 11 countries for α = 2, we find with no peer group effect (a univariate
utility function) that the optimal investment weight in the US is 0.49. Employing Equation (13)
(Abel’s preference), we find that the domestic investment in the US decreases rather than increases.
For example, for α = 2 and γ = 1, the US domestic weight decreases from 0.49 to 0.29. The same
is true for the other KUJ preference suggested in this study. Employing Equation (14), we find that
with the KUJ domestic peer group effect, for α = 2 and β = 0.5, the optimal investment weight in
the US decreases from 0.49 with no peer group effect to 0.39 with the peer group effect (this figure is
not reported in a table). Thus, with the KUJ preference with a positive cross derivative, adding the
domestic peer group effect decreases rather than increases the optimal US investment weight. Hence,
with these specific KUJ preferences, we cannot rationalize the home bias empirically, which confirms
the assertion that a positive cross derivative and CL are generally not equivalent. With a positive
cross derivative, the investor wishes to increase the correlation by increasing the domestic investment
weight but may not do it because other parameters (e.g., mean return) may induce a decrease in the
bivariate expected utility by such action. Thus, if all investors have CRRA preference with various risk
aversion parameters, the IHB is even intensified with the peer effect, despite the fact that a positive
cross derivative implies that, other things being held unchanged, the American investor wants her
portfolio to be as close as possible to the S&P stock index.
We turn now to examine whether incorporating the peer effect with a negative cross derivative
1−β
1−α wp
may increase the optimal domestic investment weight. We employ the function [ w1−α ]/[ 1−β ]
(with α < 1 and β < 1). It is easy to verify that the derivative with respect to w is positive (monotonicity)
and that the cross derivative is negative. Once again, counter intuitively, we find that with this bivariate
utility function with a negative cross derivative, the optimal domestic investment weight of the
American investor in the US increases rather than decreases, due to the peer effect. For example,
for α = 2 and γ = 0.5, we obtain with this function that the domestic investment weight increases from
0.49 to 0.54. Therefore, we conclude that, with historical data, a positive cross derivative is neither
sufficient (see Proposition 1) nor necessary (as demonstrated empirically with a preference with a
negative cross derivative) for IHB rationalization.

5. Concluding Remarks
It is well documented in the literature that, in the univariate expected utility framework, the optimal
international diversified portfolio generally reveals an investment home bias (IHB), which constitutes
a major economic puzzle. In another research strand, it is suggested that investor’s welfare is generally
determined by relative wealth, relative consumption, relative success in investment, and so on, leading
to the development of the bivariate expected utility paradigm, in which keeping up with the Joneses
(KUJ) and catching up with the Joneses (CUJ) preferences are probably the most widely employed
preferences in the bivariate framework. In this study, we combine these two research strands by
investigating whether switching from a univariate preferences framework to multivariate preferences
framework enables the rationalization of the empirically observed IHB phenomenon.
As with the bivariate preference, with a positive cross derivative, other things being held constant,
the investor wishes the performance of her portfolio to be as close as possible to the performance of
a certain local stock index (the peer effect), it is suspected that with a peer effect the investor tends
to overinvest domestically (relative to the univariate expected utility domestic optimal investment
weight). Hence, the employment of the bivariate preference with a positive cross derivative may
rationalize the IHB. We find, theoretically and empirically, that this intuitive explanation is misleading.
While it is proven in the literature that, under some approximation, employing the mean-variance
framework with a peer effect indeed rationalizes the IHB, we show in this paper that for unrestricted
preferences not confining the analysis to the mean-variance model, counter intuitively, the IHB cannot
be rationalized by the peer effect, even when the cross derivative of the bivariate preference is assumed
to be positive.

86
JRFM 2020, 13, 94

We employ the bivariate first-degree stochastic dominance (BFSD) rule and prove theoretically
that bivariate preferences with a positive cross derivative rationalizes the observed IHB, only in the
unrealistic case in which the marginal distributions of all possible portfolio under consideration are
identical. Of course, this does not hold in practice, as not all international markets are identical,
and therefore also the marginal distributions of various selected diversified portfolios are not identical.
Thus, even with peer effect, overinvesting domestically may be an inferior investment strategy,
hence the IHB cannot be explained by the peer effect.
With actual empirical international stock market data (obviously, with unequal empirical marginal
distributions), we find that the commonly employed KUJ preference with a positive cross derivative,
which intuitively implies a desire to increase the correlation by overinvesting domestically, decreases
rather than increases the domestic investment weight, hence the peer effect even enhances the IHB
puzzle. Moreover, once again counter intuitively, we find that, with a bivariate preference with a
negative cross derivative, the optimal domestic investment increases. Thus, a positive cross derivative
is neither necessary nor sufficient for IHB rationalization.
In sum, employing a general bivariate utility function with peer effect, with no constraints on
the preference employed, generally cannot rationalize the empirically observed IHB. Employing the
commonly employed specific KUJ bivariate preferences also does not rationalize the IHB. As the IHB
is an empirical fact, to rationalize this phenomenon, one needs to seek other explanations and other
research strands, as the intuitive explanation of the peer effect for rationalizing the IHB phenomenon
is misleading.

Funding: This research received no external funding.

Acknowledgments: I would like to thank three anonymous referees of this Journal for their helpful comments
which greatly improved the paper.
Conflicts of Interest: The authors declare no conflict of interest.

87
 Appendix A

Table A1. The annual rates of return 1988–2012.

JRFM 2020, 13, 94

Year USA Canada Germany France The Netherlands Norway Sweden UK Australia Japan Emerging Markets
1988 0.16 0.18 0.21 0.39 0.16 0.43 0.49 0.06 0.38 0.36 0.40
1989 0.31 0.25 0.47 0.37 0.37 0.46 0.33 0.22 0.11 0.02 0.65
1990 −0.02 −0.12 −0.09 −0.13 −0.02 0.01 −0.20 0.10 −0.16 −0.36 −0.11
1991 0.31 0.12 0.09 0.19 0.19 −0.15 0.15 0.16 0.36 0.09 0.60
1992 0.07 −0.11 −0.10 0.03 0.03 −0.22 −0.14 −0.04 −0.10 −0.21 0.11
1993 0.10 0.18 0.36 0.22 0.37 0.43 0.38 0.24 0.37 0.26 0.75
1994 0.02 −0.02 0.05 −0.05 0.13 0.24 0.19 −0.02 0.06 0.22 −0.07
1995 0.38 0.19 0.17 0.15 0.29 0.07 0.34 0.21 0.12 0.01 −0.05
1996 0.24 0.29 0.14 0.22 0.29 0.29 0.38 0.27 0.18 −0.15 0.06
1997 0.34 0.13 0.25 0.12 0.25 0.07 0.13 0.23 −0.10 −0.24 −0.12
1998 0.31 −0.06 0.30 0.42 0.24 −0.30 0.15 0.18 0.07 0.05 −0.25
1999 0.22 0.54 0.21 0.30 0.07 0.32 0.81 0.12 0.19 0.62 0.66
2000 −0.13 0.06 −0.15 −0.04 −0.04 0.00 −0.21 −0.12 −0.09 −0.28 −0.31
2001 −0.12 −0.20 −0.22 −0.22 −0.22 −0.12 −0.27 −0.14 0.03 −0.29 −0.02

88
2002 −0.23 −0.13 −0.33 −0.21 −0.20 −0.07 −0.30 −0.15 0.00 −0.10 −0.06
2003 0.29 0.55 0.65 0.41 0.29 0.50 0.66 0.32 0.51 0.36 0.56
2004 0.11 0.23 0.17 0.19 0.13 0.54 0.37 0.20 0.32 0.16 0.26
2005 0.06 0.29 0.11 0.11 0.15 0.26 0.11 0.07 0.18 0.26 0.35
2006 0.15 0.18 0.37 0.35 0.32 0.46 0.45 0.31 0.33 0.06 0.33
2007 0.06 0.30 0.36 0.14 0.21 0.32 0.01 0.08 0.30 −0.04 0.40
2008 −0.37 −0.45 −0.45 −0.43 −0.48 −0.64 −0.49 −0.48 −0.50 −0.29 −0.53
2009 0.27 0.57 0.27 0.33 0.43 0.89 0.66 0.43 0.77 0.06 0.79
2010 0.15 0.21 0.09 −0.03 0.02 0.12 0.35 0.09 0.15 0.16 0.19
2011 0.02 −0.12 −0.17 −0.16 −0.12 −0.09 −0.15 −0.03 −0.11 −0.14 −0.18
2012 0.16 0.10 0.32 0.23 0.21 0.20 0.23 0.15 0.22 0.08 0.19
 Appendix B

Table A2. The correlation matrix for the period 1988–2012.

JRFM 2020, 13, 94

USA Canada Germany France The Netherlands Norway Sweden UK Australia Japan Emerging Markets
USA 1 0.68 0.79 0.83 0.85 0.48 0.76 0.86 0.58 0.44 0.52
Canada 0.68 1 0.77 0.76 0.75 0.84 0.88 0.79 0.81 0.67 0.78
Germany 0.79 0.77 1 0.90 0.89 0.70 0.80 0.84 0.72 0.59 0.68
France 0.83 0.76 0.90 1 0.88 0.66 0.84 0.83 0.75 0.62 0.67
The Netherlands 0.85 0.75 0.89 0.88 1 0.73 0.77 0.93 0.75 0.45 0.66
Norway 0.48 0.84 0.70 0.66 0.73 1 0.79 0.75 0.83 0.55 0.76
Sweden 0.76 0.88 0.80 0.84 0.77 0.79 1 0.80 0.79 0.80 0.74
UK 0.86 0.79 0.84 0.83 0.93 0.75 0.80 1 0.78 0.42 0.65
Australia 0.58 0.81 0.72 0.75 0.75 0.83 0.79 0.78 1 0.63 0.83
Japan 0.44 0.67 0.59 0.62 0.45 0.55 0.80 0.42 0.63 1 0.67
Emerging Markets 0.52 0.78 0.68 0.67 0.66 0.76 0.74 0.65 0.83 0.67 1

89
JRFM 2020, 13, 94

References
Abel, Andrew B. 1990. Asset prices under habit formation and catching up with the Joneses. American Economic
Review 80: 38–42.
Atkinson, Anthony B., and François Bourguignon. 1982. The comparison of multi-dimensioned distributions of
economic status. Review of Economic Studies 49: 183–201. [CrossRef]
Baxter, Marianne, and Urban J. Jermann. 1997. The international diversification puzzle is worse than what you
think. American Economic Review 87: 170–80.
Berriel, Tiago C., and Saroj Bhattarai. 2013. Hedging against the government: A solution to the home asset bias
puzzle. American Economic Journal: Macroeconomics 5: 102–34. [CrossRef]
Bolton, Gery E. 1991. A comparative model of bargaining: Theory and evidence. American Economic Review 81:
1096–136.
Bolton, Gery E., and Axel Ockenfels. 2000. ERC: A theory of equity, reciprocity, and competition. American Economic
Review 90: 166–93. [CrossRef]
Britten-Jones, Mark. 1999. The sampling error in estimates of mean-variance efficient portfolio weights. Journal of
Finance 54: 655–71. [CrossRef]
Campbell, Y. John, and John H. Cochrane. 1999. By force of habit: A consumption based explanation of aggregate
stock market behavior. Journal of Political Economy 107: 205–51. [CrossRef]
Coeurdacier, Nicolas, and Helene Rey. 2013. Home bias in open economy financial macroeconomics. Journal of
Economic Literature 51: 63–115. [CrossRef]
Constantinides, George M. 1990. Habit formation: A resolution of the equity premium puzzle. Journal of Political
Economy 98: 519–43. [CrossRef]
Coval, Joshua D., and Tobias J. Moskowitz. 1999. Home bias at home: Local equity preference in domestic
portfolios. Journal of Finance 54: 2045–73. [CrossRef]
Coval, Joshua D., and Tobias J. Moskowitz. 2001. The geography of investment: Informed trading and asset prices.
Journal of Political Economy 109: 811–41. [CrossRef]
Dahlquist, Magnus, and Goran Robertsson. 2001. Direct foreign ownership, individual investors, and firm
characteristics. Journal of Financial Economics 59: 413–40. [CrossRef]
Demarzo, Peter M., Ron Kaniel, and Ilan Kremer. 2008. Relative wealth concerns and financial bubbles. Review of
Financial Studies 21: 19–50. [CrossRef]
Dupor, Bill, and Wen-Fang Liu. 2003. Jealousy and equilibrium overconsumption. American Economic Review 93:
423–28. [CrossRef]
Eeckhoudt, Louis, Béatrice Rey, and Harris Schlesinger. 2007. A good sign for multivariate risk taking. Management
Science 53: 117–24. [CrossRef]
French, Kenneth R., and James M. Poterba. 1991. International diversification and international equity markets.
American Economic Review 81: 222–26.
Galí, Jórdi. 1994. Keeping up with the Joneses: Consumption externalities, portfolio choice, and asset prices.
Journal of Money, Credit and Banking 26: 1–8. [CrossRef]
Hadar, Josef, and William R. Russell. 1969. Rules for ordering uncertain prospects. American Economic Review 59:
25–34.
Hanoch, Giora, and Haim Levy. 1969. The efficiency analysis of choices involving risk. Review of Economic Studies
36: 335–46. [CrossRef]
Huberman, Gur. 2001. Familiarity breeds investment. Review of Financial Studies 14: 659–80. [CrossRef]
Kang, Jun-Koo, and Rene M. Stultz. 1997. Why is there a home bias? An analysis of foreign portfolio equity
ownership in Japan. Journal of Financial Economics 46: 3–28. [CrossRef]
Lauterbach, Beni, and Haim Reisman. 2004. Keeping up with the Joneses and the home bias. European Financial
Management 10: 225–34. [CrossRef]
Levy, Haim. 2015. Stochastic Dominance: Investment Decision-Making under Uncertainty, 3rd ed. New York: Springer.
Levy, Haim. 2016. Economic Home Bias. Journal of Money, Credit and Banking 26. forthcoming.
Levy, Moshe, and Haim Levy. 2014. The home bias is here to stay. Journal of Banking and Finance 47: 27–40.
[CrossRef]
Levy, Haim, and Jacob Paroush. 1974. Toward multivariate efficiency criteria. Journal of Economic Theory 7: 129–42.
[CrossRef]

90
JRFM 2020, 13, 94

Levy, Moshe, and Richard Roll. 2010. The market portfolio may be mean-variance efficient after all. Review of
Financial Studies 23: 2464–491. [CrossRef]
Lewis, Karen K. 1999. Trying to explain home bias in equities and consumption. Journal of Economic Literature 37:
571–608. [CrossRef]
Ljungqvist, Lars, and Harald Uhlig. 2000. Tax policy and aggregate demand management under catching up with
the Joneses. American Economic Review 90: 356–66. [CrossRef]
Merton, Robert C., and Paul Samuelson. 1974. Fallacy of lognormal approximation to optimal portfolio
decision-making over many periods. Journal of Financial Economics 1: 67–94. [CrossRef]
Philips, Christopher B., Francis M. Kinniry, and Scott J. Donaldson. 2012. The Role of Home Bias in Global Asset
Allocation Decisions. Vanguard: Vanguard Research.
Rabin, Matthew. 1993. Incorporating fairness into game theory and economics. American Economic Review 83:
1281–302.
Rabin, Matthew. 1998. Psychology and economics. Journal of Economic Literature 36: 11–46.
Rothschild, Michael, and Joseph E. Stiglitz. 1970. Increasing risk: I. A definition. Journal of Economic Theory 2:
225–43. [CrossRef]
Tesar, Linda, and Ingrid Werner. 1995. Home bias and home turnover. Journal of International Money and Finance 14:
467–92. [CrossRef]
Tsetlin, Ilia, and Robert L. Winkler. 2009. Multivariate utility satisfying a preference for combining good with bad.
Management Science 55: 1942–952. [CrossRef]
Vanpée, Rosanne, and Lieven DeMoore. 2012. Bond and Equity Home Bias and Foreign Bias: An International Study.
Working Paper. Bruxelles: Hogesschool Universiteit Brussel.
Zizzo, Daniel John. 2003. Money burning and rank egalitarianism with random dictators. Economics Letters 81:
263–66. [CrossRef]

© 2020 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

91
 Journal of
Risk and Financial
Management

Article
Equity Option Pricing with Systematic and
Idiosyncratic Volatility and Jump Risks
Zhe Li
Business School, Nanjing Normal University, Nanjing 210023, China; zheli@[Link]

Received: 21 November 2019; Accepted: 12 January 2020; Published: 17 January 2020

Abstract: Recently, a large number of empirical studies indicated that individual equity options
exhibit a strong factor structure. In this paper, the importance of systematic and idiosyncratic
volatility and jump risks on individual equity option pricing is analyzed. First, we propose a new
factor structure model for pricing the individual equity options with stochastic volatility and jumps,
which takes into account four types of risks, i.e., the systematic diffusion, the idiosyncratic diffusion,
the systematic jump, and the idiosyncratic jump. Second, we derive the closed-form solutions for
the prices of both the market index and individual equity options by utilizing the Fourier inversion.
Finally, empirical studies are carried out to show the superiority of our model based on the S&P
500 index and the stock of Apple Inc. on options. The out-of-sample pricing performance of our
proposed model outperforms the other three benchmark models especially for short term and deep
out-of-the-money options.

Keywords: equity option pricing; factor models; stochastic volatility; jumps

JEL Classification: G13

1. Introduction
Most of the existing literature studies on option pricing are for index options, and there
are very few about equity options. One approach to modeling equity options is to employ the
state-of-the-art model in the index option literature, a stochastic volatility model with jumps (see,
for example, Bates 1996, 2000; Bakshi et al. 1997; Duffie et al. 2000; Eraker et al. 2003; Broadie et al. 2007;
Christoffersen et al. 2012; Andersen et al. 2015; Bardgett et al. 2019), but to ignore any underlying
factor structure.
In Bakshi and Kapadia (2003a), the research results indicated that the volatility risk premium is
negative in index options by examining the statistical properties of delta hedged option portfolios,
i.e., a portfolio of a long call option position hedged by a short position in the stock. On the one hand,
stock returns have a significant market component; the emergence of market volatility risk premiums
is bound to have an impact on individual equity option pricing. On the other hand, from the economic
point of view, the risk neutral distributions of individual equities are systematically different from
the market index. Thus, it is necessary to explore how volatility risk is priced in individual equity
options, which also can produce additional insights into the pricing structure of individual equity
options (see Bakshi et al. 2003). As is well known, the beta of a stock represents the sensitivity of the
risk of the individual equity with respect to the systematic risk of the market and is very useful for
portfolio construction in the capital asset pricing model. Therefore, under the assumption that stock
returns include a market component and an idiosyncratic component, Bakshi and Kapadia (2003b)
developed a factor model for equity option valuation and investigated the pricing of market volatility
risk in individual equity options. Their empirical results showed that volatility risk premiums in
equity options are smaller than in index options.

JRFM 2020, 13, 16; doi:10.3390/jrfm13010016 93 [Link]/journal/jrfm

JRFM 2020, 13, 16

Afterwards, Fouque and Kollman (2011) proposed a continuous-time capital asset pricing
model (CAPM) where the dynamics of the market index have a stochastic volatility driven by
a fast mean reverting process. Moreover, they derived the analytical approximation pricing
formulas for both the market index and individual equity call options using a singular perturbation
method. Meanwhile, a calibration method for the beta parameter was also presented based on
the estimated model parameters of both the market index and individual equity option prices.
Subsequently, Fouque and Tashman (2012) extended the constant beta-parameter factor model of
Fouque and Kollman (2011) by considering a piecewise-linear relationship between the individual
asset and the market index and proposed a regime switching factor model for the pricing market index
and individual equity options. Supposing that stock return is linearly related to market index return in
terms of the beta parameter, Carr and Madan (2012) developed a factor model for individual equity
option pricing under a purely discontinuous Lévy process via fast Fourier transform, in which the
variance gamma process for the dynamics of both the market index and stock was taken as an example
for illustration. By supposing a continuous-time CAPM with Lévy processes, Wong et al. (2012) also
derived analytical solutions to the index and equity options and explored the corresponding static
hedging with index futures. Christoffersen et al. (2018) empirically studied the equity volatility levels,
skews, and term structures by using equity option prices and principal component analysis. The results
indicated that the equity options had a strong factor structure, and then, they developed an equity
option pricing model with a CAPM factor structure and stochastic volatility, which allowed for mean
reverting stochastic volatility for the dynamics of both the market factor and individual equity.
Recently, Xiao and Zhou (2018) proposed a GARCH-jump model for individual stock returns that
took into account four types of risks: the systematic and idiosyncratic jumps and the systematic and
idiosyncratic diffusive volatility. By using a dataset consisting of the S&P 500 index and 15 individual
stock prices, their empirical results indicated that idiosyncratic jumps were a key determinant of
expected stock return.1 Instead of using only stock returns, Kapadia and Zekhnini (2019) used
both stock and option data to decompose the four risk premiums associated with systematic and
idiosyncratic diffusive and jump risks and also documented that idiosyncratic jumps are important
determinants of the mean returns of a stock from both an ex post and ex ante perspective.
Motivated by the above mentioned insights, we propose to price individual equity options in
stochastic volatility jump-diffusion models with a market factor structure, which can be seen as a
generalized version of Christoffersen et al. (2018). Specifically, in our proposed model, the individual
equity prices are driven by the market factor, as well as an idiosyncratic component that also has
stochastic volatility and jump. Due to the model belonging to the affine class, we derive the closed-form
solutions for the prices of both the market index and individual equity options by utilizing the Fourier
inversion. In addition, we provide the empirical results to test the pricing performance of the proposed
factor model based on the S&P 500 index and the stock of Apple Inc. (AAPL) on options. Toward
this end, we empirically compare the pricing performance of the proposed model with those of
the other three classical two factor stochastic volatility models being taken as benchmark models.
Empirical results presented here confirm that the equity option pricing model considering systematic
and idiosyncratic volatility and jump risks may offer a good competitor to the models of Bates (2000),
Christoffersen et al. (2009), or Christoffersen et al. (2018) for some other option markets.
The remainder of the paper proceeds as follows. In Section 2, we present a novel factor
model for equity option valuation and derive the corresponding closed-form solutions. In Section 3,

1 In fact, the work of Xiao and Zhou (2018) is a complement to the recent studies that disentangle the four types of risks
in equity premiums, such as Bégin et al. (2020), who developed a GARCH-jump model in which an individual firm’s
systematic and idiosyncratic risk have both a Gaussian diffusive and a jump component. Their empirical results showed
that normal diffusive and jump risks have drastically different effects on the expected return of individual stocks by using
20 years of returns and options on the S&P 500 and 260 stocks.

94
JRFM 2020, 13, 16

empirical studies are carried out to show the pricing performance of our proposed model. Finally,
some conclusions are stated in Section 4.

2. Equity Option Valuation

In this section, we introduce a general class of stochastic volatility models with jumps for the
dynamics of both the market factor and individual equity prices and derive closed form solutions to
the prices of the European equity call and put options.

2.1. Model Description

Consider a filtered probability space (Ω, F , Q) with information filtration {Ft }0≤t≤T satisfying
the usual conditions (increase, right-continuous, and augmented), where Q is a risk neutral measure.
We model an equity market consisting of N firms with a single market factor, It (usually approximated
by a market index in practice). The individual stock prices are denoted by Sti , for i = 1, 2, . . . , N.
For the sake of convenience, we ignore the superscript i, and denote (St )t≥0 the pricing process of an
individual stock. Investors also have access to a risk free bond that pays a return rate of r. To start,
the market factor It evolves under a risk neutral measure Q as:
 
dIt
= rdt + VI,t dW1,t
I
+ (ey − 1) Ñy (dt, dy), (1)
It− R

dVI,t = κ I (θ I − VI,t )dt + σI I
VI,t dW2,t , (2)

where It− stands for the value of It before a possible jump occurs, y ∈ R = R \ {0}, VI,t is the variance
of market factor, θ I denotes the long run variance, κ I captures the mean reversion speed of VI,t to
θ I , σI measures the volatility of volatility, 2κ I θ I ≥ σI2 to ensure that the process VI,t remains strictly
positive2 , W1,t
I and W I are correlated standard Brownian motions, i.e., the innovations to the market
2,t  
return and volatility are correlated with correlation coefficient ρ I , Cov dW1,t 2,t = ρ I dt, and
I , dW I

Ñy (dt, dy) = Ny (dt, dy) − νy (dy)dt is a compensated jump measure, where Ny (dt, dy) is the jump
measure and the Lévy kernel (or density) νy (dy) satisfies R min(1, y2 )νy (dy) < ∞.
Furthermore, we separate the effects of the market factor on individual equities’ returns into
two types of risks: the systematic diffusive volatility and jump. More specifically, the diffusive
random variation of individual equities’ returns is dependent on the Brownian motion that drives
market returns through the coefficient β di f f . In addition, the discontinuous movements in the market
return can also trigger jumps in individual equities’ returns through the coefficient β jump . Therefore,
the individual equity prices are driven by the market factor, as well as an idiosyncratic component that
also has stochastic volatility and jump, whose process under a risk neutral measure Q follows:3

  
dSt
= rdt + β di f f I
VI,t dW1,t + (e β jump y − 1) Ñy (dt, dy) + S
VS,t dW1,t
St − R
Systematic diffusive Systematic jump Idiosyncratic diffusive
 (3)
+ (eξ − 1) Ñξ (dt, dξ ),
R

Idiosyncratic jump

dVS,t = κS (θS − VS,t )dt + σS S
VS,t dW2,t , (4)

2 One can refer to Assumption 2.1 of Cheang et al. (2013) and Cheang and Garces (2019) for a more detailed explanation.
3 Obviously, our proposed model for the dynamics of the market factor and individual equity prices is an extension of
Christoffersen et al. (2018). In fact, our model also can be regarded as a further generalization of Cheang et al. (2013) and
Cheang and Garces (2019) by taking into account the factor structure.

95
JRFM 2020, 13, 16

where St− stands for the value of St before a possible jump occurs, ξ ∈ R = R \ {0}, VS,t is the
idiosyncratic variance of individual equity, θS denotes the long run idiosyncratic variance, κS captures
the mean reversion speed of VS,t to θS , σS measures the volatility of idiosyncratic variance, 2κS θS ≥ σS2
to ensure that the process VS,t remains strictly positive4 , W1,t
S and W S are correlated standard Brownian
2,t
motions, i.e., the innovations
 to the
 idiosyncratic return and volatility are correlated with correlation
coefficient ρS , Cov dW1,tS , dW S
2,t = ρS dt, but they are independent of Brownian motions in the
 
market factor, i.e., Cov dWi,t , dWj,t = 0 for i, j = 1, 2, and Ñξ (dt, dy) = Nξ (dt, dξ ) − νξ (dξ )dt is a
S I

compensated jump measure, where Nξ (dt, dξ ) is the jump measure and the Lévy kernel (or density)
νξ (dξ ) satisfies R min(1, ξ 2 )νξ (dξ ) < ∞.

2.2. Characteristic Function

In order to be able to derive the pricing formulas for the European call and put equity options,
we are particularly interested in the characteristic function of the logarithm asset price. Given the
dynamics of the underlying asset price under the Q measure, we consider the conditional characteristic
function of log-asset price XT = ln ST given the market information up to time t, which is denoted by
ϕ( x, υ1 , υ2 , t, T; φ):
  
ϕ( x, υ1 , υ2 , t, T; φ) = EQ eiφXT  Xt = x, VI,t = υ1 , VS,t = υ2
  (5)
 EQ
t e
iφXT
,

√
where EQ
t [·] denotes the condition expectation under the Q measure, t ≤ T, and i = −1.

Lemma 1. Suppose that the market factor It and individual equity price St are driven by Equations (1) and (3),
respectively. Then, the conditional characteristic function of log-asset price XT = ln ST is given by:

ϕ( x, υ1 , υ2 , t, T; φ) = exp { A(τ ) x + B(τ )υ1 + C (τ )υ2 + D (τ )}, (6)

where:

A(τ ) = iφ,
 
κ I − iφβ di f f σI ρ I − d1
1 − e − d1 τ
B(τ ) = ,
σI2 1 − g1 e − d 1 τ
 
κ − iφσS ρS − d2 1 − e − d2 τ
C (τ ) = S ,
σS2 1 − g2 e − d 2 τ
⎡ ⎤
⎢         ⎥
D (τ ) = ⎢
⎣iφr + R e
iφβ jump y
− 1 − iφ e β jump y − 1 νy (dy) + eiφξ − 1 − iφ eξ − 1 νξ (dξ )⎥
⎦τ
R
I1 I2
 
κI θI   1 − g1 e − d 1 τ
+ 2 κ I − iφβ di f f σI ρ I − d1 τ − 2 ln
σI 1 − g1
 
κ θ −
1 − g2 e d 2 τ
+ S 2 S (κS − iφσS ρS − d2 ) τ − 2 ln ,
σS 1 − g2
κ I − iφβ di f f σI ρ I − d1
g1 = ,
κ I − iφβ di f f σI ρ I + d1

4 One can refer to the Assumption 2.1 of Cheang et al. (2013) and Cheang and Garces (2019) for a more detailed explanation.

96
JRFM 2020, 13, 16

κS − iφσS ρS − d2
g2 = ,
κS − iφσS ρS + d2

 2
d1 = iφβ di f f σI ρ I − κ I + β2di f f σI2 (iφ + φ2 ),

d2 = (iφσS ρS − κS )2 + σS2 (iφ + φ2 ),

and τ = T − t.

Proof. To obtain the conditional characteristic function of log-asset price XT = ln ST , we first take the
following transformation by using the Itô lemma for Equation (3):
       
1 1
d ln St = r − β2di f f VI,t − VS,t − e β jump y − 1 νy (dy) − eξ − 1 νξ (dξ ) dt
2 2 R R
   
+ β di f f VI,t dW1,t + β jump yNy (dt, dy) + VS,t dW1,t
I S
+ ξ Nξ (dt, dξ ).
R R

The Feynman–Kac formula states that ϕ( x, υ1 , υ2 , t, T; φ) is governed by the following partial

integro-differential equation (PIDE):
⎧       
⎪ ∂ϕ 1 1 ∂ϕ
⎪
⎪
⎪ = r − β2di f f VI,t − VS,t − e β jump y − 1 νy (dy) − eξ − 1 νξ (dξ )
⎪
⎪ ∂τ 2 2 R R ∂x
⎪
⎪
⎪
⎪
⎪ 1 2  ∂2 ϕ ∂ϕ 1 2 ∂2 ϕ
⎪
⎪ + β di f f υ1 + υ2 + κ I ( θ I − υ1 ) + σI υ1 2
⎪
⎪ 2 ∂x2 ∂υ1 2 ∂υ1
⎪
⎪
⎪
⎪
⎨ ∂ϕ 1 2 ∂ ϕ 2 ∂2 ϕ ∂2 ϕ
+ κ S ( θ S − υ2 ) + σS υ2 2 + β di f f σI ρ I υ1 + σS ρS υ2 (7)
⎪ ∂υ2 2 ∂υ2 ∂x∂υ1 ∂x∂υ2
⎪
⎪  !
⎪
⎪ "
⎪
⎪ + ϕ( x + β jump y, υ1 , υ2 , t, T; φ) − ϕ( x, υ1 , υ2 , t, T; φ) νy (dy)
⎪
⎪
⎪
⎪ R
⎪
⎪
⎪
⎪ + [ ϕ( x + ξ, υ1 , υ2 , t, T; φ) − ϕ( x, υ1 , υ2 , t, T; φ)] νξ (dξ ),
⎪
⎪
⎪
⎪ R
⎩
ϕ( x, υ1 , υ2 , t, T; φ)|t=T = eiφXT .

Due to the affine structure of our model, we postulate ϕ( x, υ1 , υ2 , t, T; φ) admitting the form of (6).
Substituting Equation (6) into the above PIDE (7) gives the following system of ordinary differential
equations (ODEs) for A(τ ), B(τ ), C (τ ), and D (τ ):
⎧
⎪ ∂A(τ )
⎪
⎪ ∂τ = 0,
⎪
⎪
⎪
⎪    
⎪ ∂B(τ )
⎪ 1 1
⎪
⎪ = σI2 B2 (τ ) + β di f f σI ρ I A(τ ) − κ I B(τ ) − β2di f f A(τ ) − A2 (τ ) ,
⎪
⎪ ∂τ 2 2
⎪
⎨
∂C (τ ) 1 2 2 1 
⎪
= σS C (τ ) + [σS ρS A(τ ) − κS ] C (τ ) − A ( τ ) − A2 ( τ ) ,
⎪
⎪ ∂τ 2 2
⎪
⎪    
⎪
⎪ ∂D (τ )
⎪
⎪ = rA(τ ) + κ I θ I B(τ ) + κS θS C (τ ) + e A(τ ) β jump y − 1 − A(τ ) e β jump y − 1 νy (dy)
⎪
⎪ ∂τ R
⎪
⎪    
⎪
⎩ A(τ )ξ ξ
+ e − 1 − A(τ ) e − 1 νξ (dξ ),
R

where the boundary conditions are given as A(0) = iφ and B(0) = C (0) = D (0) = 0.
By solving the above ODEs, we can obtain the characteristic function (6).

97
JRFM 2020, 13, 16

Lemma 2. Suppose that the market factor It is driven by Equation (1). Then, the conditional characteristic
function of log-market factor ZT = ln IT is given by:
  
ψ(z, υ1 , t, T; φ) = EQ eiφZT  Zt = z, VI,t = υ1
# $ (8)
= exp Ã(τ )z + B̃(τ )υ1 + D̃ (τ ) ,

where:

Ã(τ ) = iφ,
 
κ I − iφσI ρ I − d 1 − e−dτ
B̃(τ ) = ,
σI2 1 − ge−dτ
⎡ ⎤

 
⎢   ⎥ κI θI 1 − ge−dτ
D̃ (τ ) = ⎢
⎣ iφr + e iφy
− 1 − iφ ( e y
− 1 ) νy ( dy ) ⎥
⎦ τ + ( κ I − iφσ ρ
I I − d ) τ − 2 ln ,
R σI2 1−g
I3
κ − iφσI ρ I − d
g= I ,
κ I − iφσI ρ I + d

d = (iφσI ρ I − κ I )2 + σI2 (iφ + φ2 ),

and τ = T − t.

Proof. Similar to the proof of Lemma 1, we can easily verify the above results.

2.3. Valuation of the European Index and Equity Options

Once the characteristic function is found, it is straightforward to calculate the prices of European
options by using Fourier inversion. Let C (St , T, K ) and P(St , T, K ) be the prices of the European equity
call and put options at time t with strike price K and maturity T under the risk neutral measure Q,
respectively. Then, these option prices are determined by:

C (St , T, K ) = e−rτ EQ
t [max( ST − K, 0)]

and:
P(St , T, K ) = e−rτ EQ
t [max( K − ST , 0)]

where τ = T − t is the time to maturity.

Theorem 1. Suppose that the market factor It and the individual equity price St are driven by
Equations (1) and (3), respectively. Then, the prices of the European equity call and put options with strike price
K and maturity τ = T − t are given by:
   
C (St , T, K ) = St Π1 St , T, K; β di f f , β jump − Ke−rτ Π2 St , T, K; β di f f , β jump (9)

and:
     
P(St , T, K ) = Ke−rτ 1 − Π2 St , T, K; β di f f , β jump − St 1 − Π1 St , T, K; β di f f , β jump (10)

where the risk neutral probability distribution functions Π1 and Π2 are defined by:


 
  1 e−rτ +∞ e−iφ ln K ϕ( x, υ1 , υ2 , t, T; φ − i)
Π1 St , T, K; β di f f , β jump = + dφ
2 πSt 0 iφ

98
JRFM 2020, 13, 16

and:

 
  1 1 +∞ e−iφ ln K ϕ( x, υ1 , υ2 , t, T; φ)
Π2 St , T, K; β di f f , β jump = + dφ,
2 π 0 iφ

where ϕ( x, υ1 , υ2 , t, T; φ) is the conditional characteristic function of ln ST , which can be seen in Equation (6),
and [·] indicates the real part of a complex number.

Proof. In order to get the pricing formulas of the European equity call and put options, let us first
introduce a change of measure from Q to Q̃ by the following Radon–Nikodym derivative:

dQ̃ S
= e −r ( T − t ) T .
dQ St

We denote the conditional characteristic function of XT = ln ST under the Q̃ measure by

ϕ̃( x, υ1 , υ2 , t, T; φ). Then, ϕ̃( x, υ1 , υ2 , t, T; φ) can be expressed as:
 
ϕ̃( x, υ1 , υ2 , t, T; φ) = EQ̃
t e
iφXT

−r ( T −t) ST iφXT
= EQ
t e e
St
 
= e−r(T −t)− x EQt e
i( φ −i) X T

= e−r(T −t)− x ϕ( x, υ1 , υ2 , t, T; φ − i).

Thus, the price of a European equity call option C (St , T, K ) can be calculated by utilizing
ϕ( x, υ1 , υ2 , t, T; φ) and ϕ̃( x, υ1 , υ2 , t, T; φ):

C (St , T, K ) = e−rτ EQt [max( ST − K, 0)]

   
= e Et ST 1{ST ≥K} − Ke−rτ EQ
−rτ Q
t 1{ S T ≥ K }
   
= St EQ̃
t 1{ST ≥K } − Ke
−rτ Q
Et 1{ S T ≥ K }
   
= St EQ̃
t 1{ XT ≥ln K } − Ke
−rτ Q
Et 1{ XT ≥ln K }
   
= St Π1 St , T, K; β di f f , β jump − Ke−rτ Π2 St , T, K; β di f f , β jump .

 ϕ( x, υ1 , υ2 , t, T; φ) is
Once the conditional characteristic function  obtained, we
 can easily calculate

the probability distribution functions Π1 St , T, K; β di f f , β jump and Π2 St , T, K; β di f f , β jump
according to the Lévy inversion formula:


 
  1 1 +∞ e−iφ ln K ϕ̃( x, υ1 , υ2 , t, T; φ)
Π1 St , T, K; β di f f , β jump = + dφ
2 π 0 iφ

and:

 
  1 1 +∞ e−iφ ln K ϕ( x, υ1 , υ2 , t, T; φ)
Π2 St , T, K; β di f f , β jump = + dφ,
2 π 0 iφ

A similar approach can be used to derive the pricing formula for the European equity
put option.

In a similar way, we also can present the pricing formulas for the European index call and
put options.

99
JRFM 2020, 13, 16

Theorem 2. Suppose that the market factor It is driven by Equation (1). Then, the time t prices of the European
index call and put options with strike price K and maturity τ = T − t are given by:

C ( It , T, K ) = It Π̃1 ( It , T, K ) − Ke−rτ Π̃2 ( It , T, K ) (11)

and:
! " ! "
P( It , T, K ) = Ke−rτ 1 − Π̃2 ( It , T, K ) − It 1 − Π̃1 ( It , T, K ) (12)

where the risk neutral probability distribution functions Π1 and Π2 are defined by:

 +∞
 
1 e−rτ e−iφ ln K ψ(z, υ1 , t, T; φ − i)
Π̃1 ( It , T, K ) = + dφ
2 π It 0 iφ

and:
 +∞
 
1 1 e−iφ ln K ψ(z, υ1 , t, T; φ)
Π̃2 ( It , T, K ) = + dφ,
2 π 0 iφ

where ψ(z, υ1 , t, T; φ) is the conditional characteristic function of ln IT , which can be seen in Equation (8).

3. Empirical Studies
In this section, we empirically compare the pricing performance of our proposed model with those
of the classical two factor stochastic volatility models, such as Bates (2000) (two variance SVmodel
with price jumps, 2-SVJmodel), Christoffersen et al. (2009) (two-variance SV model, 2-SV model), and
Christoffersen et al. (2018) (two-variance SV model with a single market factor, 2-FSVmodel), being
taken as benchmark models.

3.1. Data Description

We used the S&P 500 index (SPX) to proxy for the market factor and AAPL as the individual
equity. We employed the delayed market quotes on arbitrary date 8 May 2019, which was the last
date available at the time of writing, as the in-sample data to calibrate the risk neutral parameters,
and those on 9 May 2019 were used for the out-of-sample test. We used mid-quotes to represent the
option prices. To eliminate the sample noise in raw option data, we adopted some filtering rules
commonly used within the related literature: (i) we omitted those options with fewer than seven days
and more than 365 days to maturity; (ii) all observations with zero trading volume were discarded;
(iii) all options with implied volatility equal to zero and larger than 1.0 were discarded. In addition,
for the convenience of the empirical analysis in the following, we only considered the sample data
of the index call options and individual equity call options with the same expiration date. Thus, we
focused only on ten maturities slices, namely on the maturities of 24 May 2019, 31 May 2019, 7 June
2019, 14 June 2019, 21 June 2019, 19 July 2019, 16 August 2019, 20 September 2019, 18 October 2019,
and 17 January 2020.
After these filters, we had a total of 401 observations for the S&P 500 index call option on
8 May 2019. The individual equity option sample contained 233 call options on 8 May 2019 and 264
call options on 9 May 2019, respectively. Due to the life of an option being usually less than one year,
we chose the three month U.S. Treasury Bill Rate to substitute for the risk free interest rate. All of the
data were downloaded from the Chicago Board Options Exchange ([Link]

3.2. Parameter Estimation

Our proposed model allowed a general distribution for jump components of the market factor
and individual equity price and thus could be easily introduced to the special cases such that the jump
components follow the compound Poisson process of Merton (1976) and Kou (2002), etc. For different
types of Lévy kernels, different forms of our model can be presented. In order to keep consistent with

100
JRFM 2020, 13, 16

Bates (2000) for comparative analysis, in the following, we assumed that the jump components of the
dynamics for the market factor and individual equity followed compound Poisson processes and the
jump magnitude was drawn from the log-normal distribution of Merton (1976). Thus, the Lévy kernels
for the market factor and individual equity, respectively, are given by:
% &
1 ( y − μ I )2
νy (dy) = λ I  exp − dy (13)
2πδ2I 2δ2I

and: % &
1 ( ξ − μ S )2
νξ (dξ ) = λS  exp − dξ, (14)
2πδS2 2δS2

where λ j , for j = I, S, denotes the jump intensity, μ j is the mean of the jump size, and δj is the variance
of the jump size. Then, the integrals Ii , for i = 1, 2, 3, in Lemmas 1 and 2 can be calculated as follows:
  
iφβ μ − 1 φ2 β2 δ2 β μ + 1 β2 δ2
I1 = λ I e jump I 2 jump I − 1 − iφ e jump I 2 jump I − 1 ,
  
I2 = λS eiφμS − 2 φ δS − 1 − iφ eμS + 2 δS − 1 ,
1 2 2 1 2

and
  
I3 = λ I eiφμ I − 2 φ δI − 1 − iφ eμ I + 2 δI − 1 .
1 2 2 1 2

Based on Theorems 1 and 2, we employed a two step calibration procedure (see, for example,
Wong et al. 2012; Christoffersen et al. 2018) to estimate the model parameters. First, we calibrated
the market index dynamic Θ I based on the S&P 500 index option price alone. Second, we used the
equity option price to calibrate the individual equity dynamic ΘS conditional on estimates of Θ I .
Consider the situation in which an investor wants to hedge his or her equity position with index
options and hedging horizon T. For brevity, we further suppose that the investor observes index
option prices and equity option prices both with maturity T, the same as hedging horizon. Specifically,
the dataset contains Mt index option prices C ( It , T, Ki ), for i = 1, 2, . . . , Mt , and Nt equity option prices
C (St , T, K j ), for j = 1, 2, . . . , Nt .
In the calibration process, the risk neutral model parameters were backed out by minimizing a
loss function capturing the fit between the theoretical model and market prices. We employed the
root mean squared errors (RMSE) as the objective function. The first step calibrated the risk neutral
parameters for the index process, which are calibrated by:
'
(  2
( 1 Mt
RMSE( I ) = arg min )
ΘI
ΘI Mt ∑ Ci,market ( It , T, Ki ) − Ci,model ( It , T, Ki ) , (15)
i =1

where Ci,market ( It , T, Ki ) is the market price of the index call option contract from the sample and
ΘI
Ci,model ( It , T, Ki ) represents the model price calculated using Equation (15) and the vector of model
input parameters Θ I .
The second calibrated the beta and the parameters for the idiosyncratic risk:
'
(  2
( 1 Nt
ΘS
RMSE(S) = arg min ) ∑ Cj,market (St , T, K j ) − Cj,model (St , T, K j ) , (16)
ΘS Nt j =1

101
JRFM 2020, 13, 16

where Cj,market (St , T, K j ) is the market price of the equity call option contract from the sample and
Θ S
Cj,model (St , T, K j ) represents the model price calculated using Equation (13) and the vector of model
input parameters ΘS .
On the basis of the above calibration method, Table 1 presents the risk neutral parameter estimates
across various model specifications. Note that the values of the diffusive beta β di f f and jump beta β jump
for our proposed model were 0.3891 and 0.8429, respectively. The corresponding value of β di f f for
the 2-FSV model was 0.2457. Obviously, both our proposed model and the 2-FSV model showed that
AAPL tended to have a relatively low exposure to diffusive market movements. However, the jump
exposure coefficient β jump = 0.8429 indicated that the AAPL had a strong exposure to market jumps,
which meant that the factor structure of the jumps was much stronger than the one of the diffusive
movements. The reason for this result may be related to the sample data we selected. If we can get
more sample data in the future, we will do an in-depth analysis. Moreover, we also can see that the
values of correlation ρ were strongly negative for four models, capturing the so-called leverage effect
both in the index and individual equity.

Table 1. Estimated parameters. Note: This table shows the average of the estimated parameters
obtained by minimizing the root mean squared pricing errors between the market price and the model
price for each option on 8 May 2019. Standard errors are reported in parentheses .

Our 2-FSV 2-SV 2-SVJ

Parameters
SPX AAPL SPX AAPL AAPL AAPL
VI,0 /V1,0 0.0133 0.0119 0.0239 0.0181
(0.0000) (0.0000) (0.0002) (0.0001)
VS,0 /V2,0 0.0470 0.0514 0.0197 0.0176
(0.0000) (0.0000) (0.0002) (0.0002)
κ I /κ1 0.2496 0.2929 0.3489 0.4064
(0.0212) (0.0148) (0.0118) (0.0311)
κS /κ2 0.2454 0.1504 0.4131 0.4108
(0.0288) (0.0797) (0.0729) (0.0171)
θ I /θ1 0.2820 0.3066 0.3314 0.2817
(0.0181) (0.0317) (0.0534) (0.0348)
θS /θ2 0.2303 0.3683 0.2447 0.3415
(0.0190) (0.0590) (0.0365) (0.0423)
σI /σ1 0.3472 0.3932 0.1615 0.1898
(0.0127) (0.0137) (0.0081) (0.0106)
σS /σ2 0.1496 0.1640 0.2206 0.1970
(0.0056) (0.0135) (0.0386) (0.0059)
λI 0.0450
(0.0017)
λS 0.3413 0.3065
(0.2463) (0.1194)
μI 0.1657
(0.0599)
μS 0.0889 0.0333
(0.0391) (0.0042)
δI 0.0850
(0.0113)
δS 0.0679 0.0534
(0.0078) (0.0013)
β di f f 0.3891 0.2457
(0.0381) (0.0983)
β jump 0.8429
(0.8091)
ρ I /ρ1 −0.9290 −0.8498 −0.9222 −0.7445
(0.0063) (0.0080) (0.0096) (0.0297)
ρS /ρ2 −0.9926 −0.8938 −0.7673 −0.7817
(0.0001) (0.0469) (0.1632) (0.0549)

102
JRFM 2020, 13, 16

3.3. Pricing Performance

In this subsection, we present the empirical results for the calibrated models. In order to investigate
the impacts of the systematic and idiosyncratic volatility and jump risks on equity option pricing,
we took the 2-FSV, 2-SV, and 2-SVJ models as benchmark models to evaluate the pricing performance
of our proposed model.
Figures 1–10 exhibit the predicted prices of the four model specifications and market prices listed
on 9 May 2019, with 11, 16, 21, 26, 31, 51, 71, 96, 116, and 181 trading days to expiry, respectively.
Here, the predicted prices (out-of-sample pricing) were calculated by the in-sample calibration
parameters reported in Table 1. One can clearly observe from the left panels of Figures 1–10 that the
option prices obtained by theoretical models were generally closer to the market prices for different
strike prices. To further investigate the pricing performance of the four models, the right panels of
Figures 1–10 show the relative price differences (relative errors) between the theoretical model prices
and market prices.5 For simplicity, we refer to a call option as deep out-of-the-money (DOTM) if
S/K ≤ 0.90; out-of-the-money (OTM) if 0.90 < S/K ≤ 0.97; at-the-money (ATM) if 0.97 < S/K ≤ 1.03;
in-the-money (ITM) if 0.97 < S/K ≤ 1.10; and deep in-the-money (ITM) if 1.10 < S/K. Moreover,
we considered options less than 60 days to expiration as short term; options with 60–120 days to
expiration as medium term; and options larger than 120 days to expiration as long term. For the
options with 11, 16, 21, 26, 31, and 51 trading days to expiry, the relative pricing errors produced by
our proposed model were all significantly lower than those of 2-FSV, 2-SV, and 2-SVJ models in the
case of DOTM options, while the relative errors of all models were slightly higher.
It is also worth noting that the pricing performance of the stochastic model with jump behavior
was much better than that of the model without jump in the case of deep out-of-money. For the options
with 71, 96, 116, and 181 trading days to expiry, we did not find the same conclusions as the above short
term options. In conclusion, the pricing performance of equity option valuation model considering
market and idiosyncratic volatility and jump risks was significantly improved for short term and
DOTM options.

In-sample, T=May 24, 2019 Out-of-sample, T=May 24, 2019

0.6 1
Our model Our model
2-FSV model 0.9 2-FSV model
2-SV model 2-SV model
0.5
2-SVJ model 0.8 2-SVJ model

0.7
0.4
Relative Error

Relative Error

0.6

0.3 0.5

0.4
0.2
0.3

0.2
0.1
0.1

0 0
180 200 220 240 140 160 180 200 220 240
Strike Price Strike Price

Figure 1. The comparison of predicted prices of four model specifications and market prices on
9 May 2019, with maturity T = 24 May 2019.

|C −C |
5 The relative error is defined by model market
Cmarket × 100%, where Cmodel and Cmarket denote the theoretical model option prices
and the real market prices, respectively.

103
JRFM 2020, 13, 16

In-sample, T=May 31, 2019 Out-of-sample, T=May 31, 2019

0.7 1
Our model Our model
2-FSV model 0.9 2-FSV model
0.6 2-SV model 2-SV model
2-SVJ model 0.8 2-SVJ model

0.5 0.7
Relative Error

Relative Error
0.6
0.4
0.5
0.3
0.4

0.2 0.3

0.2
0.1
0.1

0 0
180 200 220 240 140 160 180 200 220 240
Strike Price Strike Price

Figure 2. The comparison of predicted prices of four model specifications and market prices on
9 May 2019, with maturity T = 31 May 2019.

In-sample, T=June 7, 2019 Out-of-sample, T=June 7, 2019

0.45 0.9
Our model Our model
0.4 2-FSV model 0.8 2-FSV model
2-SV model 2-SV model
2-SVJ model 2-SVJ model
0.35 0.7

0.3 0.6
Relative Error

Relative Error

0.25 0.5

0.2 0.4

0.15 0.3

0.1 0.2

0.05 0.1

0 0
160 180 200 220 240 160 180 200 220 240
Strike Price Strike Price

Figure 3. The comparison of predicted prices of four model specifications and market prices on
9 May 2019, with maturity T = 7 June 2019.

104
JRFM 2020, 13, 16

In-sample, T=June 14, 2019 Out-of-sample, T=June 14, 2019

0.6 0.9
Our model Our model
2-FSV model 0.8 2-FSV model
2-SV model 2-SV model
0.5
2-SVJ model 2-SVJ model
0.7

0.4 0.6
Relative Error

Relative Error
0.5
0.3
0.4

0.2 0.3

0.2
0.1
0.1

0 0
160 180 200 220 240 260 160 180 200 220 240 260
Strike Price Strike Price

Figure 4. The comparison of predicted prices of four model specifications and market prices on
9 May 2019, with maturity T = 14 June 2019.

In-sample, T=June 21, 2019 Out-of-sample, T=June 21, 2019

1 1
Our model Our model
0.9 2-FSV model 0.9 2-FSV model
2-SV model 2-SV model
0.8 2-SVJ model 0.8 2-SVJ model

0.7 0.7
Relative Error

Relative Error

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
150 200 250 150 200 250
Strike Price Strike Price

Figure 5. The comparison of predicted prices of four model specifications and market prices on
9 May 2019, with maturity T = 21 June 2019.

105
JRFM 2020, 13, 16

In-sample, T=July 19, 2019 Out-of-sample, T=July 19, 2019

0.7 1
Our model Our model
2-FSV model 0.9 2-FSV model
0.6 2-SV model 2-SV model
2-SVJ model 0.8 2-SVJ model

0.5 0.7
Relative Error

Relative Error
0.6
0.4
0.5
0.3
0.4

0.2 0.3

0.2
0.1
0.1

0 0
140 160 180 200 220 240 160 180 200 220 240 260
Strike Price Strike Price

Figure 6. The comparison of predicted prices of four model specifications and market prices on
9 May 2019, with maturity T = 19 July 2019.

In-sample, T=August 16, 2019 Out-of-sample, T=August 16, 2019

0.5 0.7
Our model Our model
0.45 2-FSV model 2-FSV model
2-SV model 0.6 2-SV model
0.4 2-SVJ model 2-SVJ model

0.35 0.5
Relative Error

Relative Error

0.3
0.4
0.25
0.3
0.2

0.15 0.2

0.1
0.1
0.05

0 0
150 200 250 150 200 250
Strike Price Strike Price

Figure 7. The comparison of predicted prices of four model specifications and market prices on
9 May 2019, with maturity T = 16 August 2019.

106
JRFM 2020, 13, 16

In-sample, T=September 20, 2019 Out-of-sample, T=September 20, 2019

0.25 0.4
Our model Our model
2-FSV model 2-FSV model
0.35
2-SV model 2-SV model
0.2 2-SVJ model 2-SVJ model
0.3

0.25
Relative Error

Relative Error
0.15

0.2

0.1
0.15

0.1
0.05

0.05

0 0
140 160 180 200 220 240 260 160 180 200 220 240 260
Strike Price Strike Price

Figure 8. The comparison of predicted prices of four model specifications and market prices on
9 May 2019, with maturity T = 20 September 2019.

In-sample, T=October 18, 2019 Out-of-sample, T=October 18, 2019

0.7 0.8
Our model Our model
2-FSV model 2-FSV model
0.6 0.7
2-SV model 2-SV model
2-SVJ model 2-SVJ model
0.6
0.5

0.5
Relative Error

Relative Error

0.4
0.4
0.3
0.3

0.2
0.2

0.1 0.1

0 0
150 200 250 300 150 200 250 300
Strike Price Strike Price

Figure 9. The comparison of predicted prices of four model specifications and market prices on
9 May 2019, with maturity T = 18 October 2019.

107
JRFM 2020, 13, 16

In-sample, T=January 17, 2020 Out-of-sample, T=January 17, 2020

0.4 0.8
Our model Our model
2-FSV model 2-FSV model
0.35 2-SV model 2-SV model
0.7
2-SVJ model 2-SVJ model

0.3 0.6
Relative Error

Relative Error
0.25 0.5

0.2 0.4

0.15 0.3

0.1 0.2

0.05 0.1

0 0
100 150 200 250 300 350 100 150 200 250 300 350
Strike Price Strike Price

Figure 10. The comparison of predicted prices of four model specifications and market prices on
9 May 2019, with maturity T = 17 January 2019.

To summarize the model calibration results, we also adopted the RMSE as a measure of the
goodness of fit. Table 2 reports the out-of-sample pricing errors for the four models across different
maturities. Note from Table 2 that our proposed model generally outperformed the other three models
in terms of out-of-sample pricing errors. In fact, the same was true for in-sample, whose pricing errors
were generally lower than those of the out-of-sample. We will not repeat them here. To measure
the extent to which a model was better or worse than another, we defined the improvement rate
as the relative differences between the pricing errors from the benchmark model and our proposed
model, i.e.,
RMSEbenchmark − RMSEour
Improvement rate = × 100%
RMSEbenchmark
where RMSEour and RMSEbenchmark denote the RMSE implied by our model and benchmark model,
respectively. A positive (or negative) value of improvement rate meant that our model yielded lower
(or higher) pricing errors than benchmark model, implying that the pricing performance of the former
was better (or worse) than that of the latter by a percentage of that value.
From the last column of Table 2, we can see that our model was superior to the 2-SVJ model
across different maturities, which meant that it was necessary to consider the market factor structure
in equity option pricing. From the third last column of Table 2, the improvement rate indicated that
our model slightly outperformed the 2-FSV model in terms of short term options, but was worse than
that of both medium and long term. In spite of this, our empirical study presented here could at least
illustrate that the equity option pricing model considering systematic and idiosyncratic volatility and
jump risks may offer a good competitor of the models of Bates (2000), Christoffersen et al. (2009), or
Christoffersen et al. (2018) for some other equity option markets.

108
JRFM 2020, 13, 16

Table 2. Out-of-sample pricing errors. Note: This table shows the out-of-sample pricing errors across
different maturities. Pricing errors are reported as the root mean squared errors (RMSE) of option
prices for four models.

RMSE Improvement Rate

Our 2-FSV 2-SV 2-SVJ
Maturity Our vs. 2-FSV Our vs. 2-SV Our vs. 2-SVJ
24 May 2019 0.2573 0.2574 0.2596 0.2707 0.0373% 0.8803% 4.9568%
31 May 2019 0.2507 0.2508 0.2564 0.2652 0.0392% 2.2499% 5.4846%
7 June 2019 0.2343 0.2347 0.2527 0.2474 0.1764% 7.2947% 5.3044%
14 June 2019 0.1992 0.2041 0.2261 0.2099 2.4278% 11.9155% 5.0858%
21 June 2019 0.1824 0.1827 0.1873 0.1916 0.1399% 2.5963% 4.7934%
19 July 2019 0.3256 0.3301 0.3326 0.3383 1.3434% 2.0948% 3.7368%
16 August 2019 0.2856 0.2835 0.2879 0.2922 −0.7573% 0.7946% 2.2384%
20 September 2019 0.3177 0.3159 0.3162 0.3222 −0.5932% -0.4851% 1.4002%
18 October 2019 0.1185 0.1180 0.1215 0.1272 −0.4458% 2.4886% 6.8593%
17 January 2020 0.4882 0.4882 0.4893 0.4943 −0.0071% 0.2182% 1.2201%

4. Conclusions
In Christoffersen et al. (2018), the issues of the equity volatility levels, skews, and term structures
were investigated by using equity option prices and the principal component analysis method.
Their empirical results indicated that the equity options had a strong factor structure, and then,
they developed an equity option pricing model with a CAPM factor structure and stochastic volatility.
In addition, jumps in stock returns of individual firms were triggered by either systematic events or
idiosyncratic shocks. Some recent studies indicated that idiosyncratic jumps were a key important
determinant of expected stock; see, for example, Xiao and Zhou (2018), Kapadia and Zekhnini (2019)
and Bégin et al. (2020).
Motivated by these insights, we developed a novel model for pricing individual equity options
that incorporated a market factor structure, which could be seen as a generalized version of the work
by Christoffersen et al. (2018). Specifically, in our model, the individual equity prices were driven by
the market factor, as well as an idiosyncratic component that also had stochastic volatility and jump.
Due to our model belonging to the affine class, we derived the closed-form solutions for the prices of
both the market index and individual equity options by utilizing the Fourier inversion. In addition,
we provided the empirical results to test the pricing performance of our proposed factor model based
on the S&P 500 index and the AAPL stock on options. Toward this end, we empirically compared
the pricing performance of our proposed model with those of the other three classical two factor
stochastic volatility models being taken as benchmark models. The out-of-sample pricing performance
of equity option valuation model considering market and idiosyncratic volatility and jump risks
was significantly improved for short term and DOTM options. In conclusion, the empirical results
presented here at least confirmed that the equity option pricing model considering systematic and
idiosyncratic volatility and jump risks may offer as good competitor of the models of Bates (2000),
Christoffersen et al. (2009), or Christoffersen et al. (2018) for some other option markets.

Funding: This work was supported by the National Natural Science Foundation of China (Grant No. 71901124)
and the Natural Science Foundation of Jiangsu Province (Grant No. BK20190695).
Conflicts of Interest: The author declares no conflict of interest.

References
Andersen, Torben G., Nicola Fusari, and Viktor Todorov. 2015. The risk premia embedded in index options.
Journal of Financial Economics 117: 558–84. [CrossRef]
Bakshi, Gurdip, Charles Cao, and Zhiwu Chen. 1997. Empirical performance of alternative option pricing models.
Journal of Finance 52: 2003–49. [CrossRef]

109
JRFM 2020, 13, 16

Bakshi, Gurdip, Nikunj Kapadia, and Dilip Madan. 2003. Stock return characteristics, skew laws, and the
differential pricing of individual equity options. Review of Financial Studies 16: 101–43. [CrossRef]
Bakshi, Gurdip, and Nikunj Kapadia. 2003a. Delta-hedged gains and the negative market volatility risk premium.
Review of Financial Studies 16: 527–66.
Bakshi, Gurdip, and Nikunj Kapadia. 2003b. Volatility risk premiums embedded in individual equity options:
Some new insights. Journal of Derivatives 11: 45–54. [CrossRef]
Bardgett, Chris, Elise Gourier, and Markus Leippold. 2019. Inferring volatility dynamics and risk premia from the
S&P 500 and VIX markets. Journal of Financial Economics 131: 593–618. [CrossRef]
Bates, David S. 1996. Jumps and stochastic volatility: Exchange rate processes implicit in Deutsche mark options.
Review of Financial Studies 9: 69–107. [CrossRef]
Bates, David S. 2000. Post-’87 crash fears in the S&P 500 futures option market. Journal of Econometrics 94: 181–238.
[CrossRef]
Broadie, Mark, Mikhail Chernov, and Michael Johannes. 2007. Model specification and risk premia: Evidence
from futures options. Journal of Finance 62: 1453–90. [CrossRef]
Bégin, Jean-François, Christian Dorion, and Geneviève Gauthier. 2020. Idiosyncratic jump risk matters: Evidence
from equity returns and options. Review of Financial Studies 33: 155–211.
Carr, Peter, and Dilip B. Madan. 2012. Factor models for option pricing. Asia-Pacific Financial Markets 19: 319–29.
[CrossRef]
Cheang, Gerald H. L., Carl Chiarella, and Andrew Ziogas. 2013. The representation of American options prices
under stochastic volatility and jump-diffusion dynamics. Quantitative Finance 13: 241–53. [CrossRef]
Cheang, Gerald H. L., and Len Patrick Dominic M. Garces. 2019. Representation of exchange option prices under
stochastic volatility jump-diffusion dynamics. Quantitative Finance. [CrossRef]
Christoffersen, Peter, Kris Jacobs, and Chayawat Ornthanalai. 2012. Dynamic jump intensities and risk premiums:
Evidence from S&P 500 returns and options. Journal of Financial Economics 106: 447–72.
Christoffersen, Peter, Mathieu Fournier, and Kris Jacobs. 2018. The factor structure in equity options. Review of
Financial Studies 31: 595–637. [CrossRef]
Christoffersen, Peter, Steven Heston, and Kris Jacobs. 2009. The shape and term structure of the index option smirk:
Why multifactor stochastic volatility models work so well. Management Science 55: 1914–32. [CrossRef]
Duffie, Darrell, Jun Pan, and Kenneth Singleton. 2000. Transform analysis and asset pricing for affine jump
diffusions. Econometrica 68: 1343–76. [CrossRef]
Eraker, Biørn, Miichael Johannes, and Nicholas Polson. 2003. The Impact of Jumps in Volatility and Returns.
Journal of Finance 58: 1269–1300. [CrossRef]
Fouque, Jean-Pierre, and Adam P. Tashman. 2012. Option pricing under a stressed-beta model. Annals of Finance 8:
183–203. [CrossRef]
Fouque, Jean-Pierre, and Eli Kollman. 2011. Calibration of stock betas from skews of implied volatilities. Applied
Mathematical Finance 18: 119–37. [CrossRef]
Kapadia, Nishad, and Morad Zekhnini. 2019. Do idiosyncratic jumps matter? Journal of Financial Economics 131:
666–92. [CrossRef]
Kou, Steven G. 2002. A jump-diffusion model for option pricing. Management Science 48: 1086–101. [CrossRef]
Merton, Robert C. 1976. Option pricing when underlying stock returns are discontinuous. Journal of Financial
Economics 3: 125–44. [CrossRef]
Wong, Hoi Ying, Edwin Kwan Hung Cheung, and Shiu Fung Wong. 2012. Lévy betas: Static hedging with index
futures. Journal of Futures Markets 32:1034–59. [CrossRef]
Xiao, Xiao, and Chen Zhou. 2018. The decomposition of jump risks in individual stock returns. Journal of Empirical
Finance 47: 207–28.

c 2020 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

110
 Journal of
Risk and Financial
Management

Article
CVaR Regression Based on the Relation between
CVaR and Mixed-Quantile Quadrangles
Alex Golodnikov 1 , Viktor Kuzmenko 1 and Stan Uryasev 2, *
1 V.M. Glushkov Institute of Cybernetics, 40, pr. Akademika Glushkova, 03187 Kyiv, Ukraine
2 Applied Mathematics & Statistics, Stony Brook University, B-148 Math Tower, Stony Brook, NY 11794, USA
* Correspondence: [Link]@[Link]

Received: 16 May 2019; Accepted: 20 June 2019; Published: 26 June 2019

Abstract: A popular risk measure, conditional value-at-risk (CVaR), is called expected shortfall (ES) in
financial applications. The research presented involved developing algorithms for the implementation
of linear regression for estimating CVaR as a function of some factors. Such regression is called
CVaR (superquantile) regression. The main statement of this paper is: CVaR linear regression can
be reduced to minimizing the Rockafellar error function with linear programming. The theoretical
basis for the analysis is established with the quadrangle theory of risk functions. We derived
relationships between elements of CVaR quadrangle and mixed-quantile quadrangle for discrete
distributions with equally probable atoms. The deviation in the CVaR quadrangle is an integral.
We present two equivalent variants of discretization of this integral, which resulted in two sets of
parameters for the mixed-quantile quadrangle. For the first set of parameters, the minimization
of error from the CVaR quadrangle is equivalent to the minimization of the Rockafellar error from
the mixed-quantile quadrangle. Alternatively, a two-stage procedure based on the decomposition
theorem can be used for CVaR linear regression with both sets of parameters. This procedure is valid
because the deviation in the mixed-quantile quadrangle (called mixed CVaR deviation) coincides
with the deviation in the CVaR quadrangle for both sets of parameters. We illustrated theoretical
results with a case study demonstrating the numerical efficiency of the suggested approach. The case
study codes, data, and results are posted on the website. The case study was done with the Portfolio
Safeguard (PSG) optimization package, which has precoded risk, deviation, and error functions for
the considered quadrangles.

Keywords: quantile; VaR; quadrangle; CVaR; conditional value-at-risk; expected shortfall; ES;
superquantile; deviation; risk; error; regret; minimization; CVaR estimation; regression; linear
regression; linear programming; portfolio safeguard; PSG

1. Introduction
We start the introduction with a quick outline of the main result of this paper. The conditional
value-at-risk (CVaR) is a popular risk measure. It is called expected shortfall (ES) in financial applications
and it is included in financial regulations. This paper provides algorithms for the estimation of CVaR
with linear regression as a function of factors. This task is of critical importance in practical applications
involving low probability events.
By definition, CVaR is an integral of the value-at-risk (VaR) in the tail of a distribution. VaR can
be estimated with the quantile regression by minimizing the Koenker–Bassett error function. This
paper shows that CVaR can be estimated by minimizing a mixture of the Koenker–Bassett errors with
an additional constraint. This mixture is called the Rockafellar error and it has been earlier used for
CVaR estimation without a rigorous mathematical justification. One more equivalent variant of CVaR
regression can be done by minimizing a mixture of CVaR deviations for finding all coefficients, except

JRFM 2019, 12, 107; doi:10.3390/jrfm12030107 111 [Link]/journal/jrfm

JRFM 2019, 12, 107

the intercept. In this case, the intercept is calculated using an analytical expression, which is the CVaR
of the optimal residual without an intercept. The new mathematical result links quantile and CVaR
regressions and shows that convex and linear programming methods can be straightforwardly used
for CVaR estimation. Mathematical justification of the results involves a risk quadrangle concept
combining regret, error, risk, deviation, and statistic notions.
Quantiles evaluating different parts of a distribution of a random value are quite popular in
various applications. In particular, quantiles are used to estimate tail of a distribution (e.g., 90%, 95%,
and 99% quantiles). This paper is motivated by finance applications, where a quantile is called VaR.
Risk measure VaR is included in finance regulations for the estimation of market risk. VaR has several
attractive properties, such as the simplicity of calculation, stability of estimation, and availability
of quantile regression, for the estimation of VaR as a function of explanatory factors. The quantile
regression (see Koenker and Bassett (1978), Koenker (2005)) is an important factor supporting the
popularity of VaR. For instance, a quantile regression was used by Adrian and Brunnermeier (2016) to
estimate institution’s contribution to systemic risk.
However, VaR also has some undesirable properties:

• Lack of convexity: portfolio diversification may increase VaR.

• VaR is not sensitive to outcomes exceeding VaR, which allows for stretching of the distribution
without an increasing of the risk measured by VaR.
• VaR has poor mathematical properties, such as discontinuity with respect to (w.r.t.) portfolio
positions for discrete distributions based on historical data.

Shortcomings of VaR led financial regulators to use an alternative measure of risk, which
is called conditional value-at-risk (CVaR) in this paper. This risk measure was introduced in
Rockafellar and Uryasev (2000) and further studied in Rockafellar and Uryasev (2002) and many other
papers. CVaR for continuous distributions equals the conditional expectation of losses exceeding VaR.
An important mathematical fact is that CVaR is a coherent risk measure (see Acerbi and Tasche (2002),
Rockafellar and Uryasev (2002)). Ziegel (2014) shows that CVaR is elicitable in a week sense.
Fissler and Ziegel (2015) proved that (VaR, CVaR) is jointly elicitable, meaning elic(CVaR) ≤ 2, and more
generally, that spectral risk measures have a low elicitation complexity. These results clarify the
regression procedure of Rockafellar et al. (2014); their algorithm implicitly tracks the quantiles
suggested by elicitation complexity.
Rockafellar and Uryasev (2000, 2002) have shown that CVaR of a convex function of variables is
also a convex function. Due to this property, CVaR optimization problems can be reduced to convex
and linear optimization problems.
This paper is based on risk quadrangle theory, which defines quadrangles (i.e., groups) of stochastic
functionals Rockafellar and Uryasev (2013). Every quadrangle contains risk, deviation, error, and regret
(negative utility). These elements of the quadrangle are linked by the statistic function.
The relation of quantile regression and CVaR optimization was explained using a quantile
quadrangle (see Rockafellar and Uryasev (2013)). It was shown that the Koenker–Bassett error
function and CVaR belong to the same quantile quadrangle. By minimizing the Koenker–Bassett error
function with respect to one parameter, we obtain the CVaR deviation (which is the CVaR for the
centered random value). The optimal value of the parameter, which is called statistic, equals VaR.
Therefore, the linear regression with the Koenker–Bassett error estimates VaR as a function of factors.
The fact that statistic equals VaR and is also used for building the optimization approach for CVaR
(see Rockafellar and Uryasev (2000, 2002)).
Another important contribution that takes advantage of quadrangle theory is the regression
decomposition theorem proved in Rockafellar et al. (2008). With this decomposition theorem,
the regression problem is decomposed in two steps: (1) minimization of deviation from the
corresponding quadrangle, and (2) calculation of the intercept by using statistic from this quadrangle.
For instance, by applying the decomposition theorem to the quantile quadrangle, we can do quantile

112
JRFM 2019, 12, 107

regression by minimizing CVaR deviation for finding all regression coefficients, except the intercept.
Then, the intercept is calculated by using VaR statistic.
CVaR can be approximated using the weighted average of VaRs with different confidence levels,
which is called the mixed VaR method. Rockafellar and Uryasev (2013) demonstrated that mixed VaR
is a statistic in the mixed-quantile quadrangle. The error function, corresponding to this quadrangle
(called the Rockafellar error) can be minimized for the estimation of the mixed VaR with linear regression.
The Rockafellar error is a solution of a minimization problem with one linear constraint. Linear
regression for estimating mixed VaR can be done by minimizing the Rockafellar error with convex and
linear programming (Appendix A contains these formulations). Alternatively, this regression can be
done in two steps with the decomposition theorem. The deviation in the mixed-quantile quadrangle
is the mixed CVaR deviation, therefore all regression coefficients, except the intercept, can be found
by minimizing this deviation. Further, the intercept can be found by using statistic, which is the
mixed VaR.
Rockafellar et al. (2014) developed the CVaR quadrangle with the statistic equal to CVaR.
Risk envelopes and identifiers for this quadrangle were calculated in Rockafellar and Royset (2018).
This CVaR quadrangle is a theoretical basis for constructing the regression for estimating CVaR.
Rockafellar et al. (2014) called the linear regression for estimation of CVaR using superquantile (CVaR)
regression. The superquantile is an equivalent term for CVaR. Here we use the term “CVaR regression”.
The CVaR regression plays a major role in various engineering areas, especially in financial applications.
For instance, Huang and Uryasev (2018) used CVaR regression for the estimation of risk contributions
of financial institutions and Beraldi et al. (2019) used CVaR for solving portfolio optimization problems
with transaction costs.
This paper considers only discrete random values with a finite number of equally probable atoms.
This special case is considered because it is needed for the implementation of the linear regression
for the CVaR estimation. We have explained with an example how parameters of the optimization
problems are calculated.
The equal probabilities property was used for calculating parameters of optimization problems.
It is possible to calculate parameters with non-equal probabilities of atoms, but this is beyond the scope
of the paper, which is focused on the linear regression.
We suggested two sets (Sets 1 and 2) of parameters for the mixed-quantile quadrangle. Set 1
corresponds to the two-step implementation of the CVaR regression in Rockafellar et al. (2014),
and Set 2 is a new set of parameters. We proved that with Set 1, the statistic, risk, and deviation of the
mixed-quantile and CVaR quadrangles coincide. Therefore, CVaR regression can be done by minimizing
the Rockafellar error with convex and linear programming. For Set 2, the mixed-quantile and CVaR
quadrangle share risk and deviation parameters. Also, the statistic of this mixed-quantile quadrangle
(which may not be unique) includes statistic of the CVaR quadrangle. Therefore, minimizing the
Rockafellar error correctly calculates all regression coefficients, but may provide an incorrect intercept.
This is actually not a big concern because we know that the intercept is equal to the CVaR of an optimal
residual without intercept.
Also, we demonstrated that the CVaR regression can be done in two steps with the decomposition
theorem by using parameters from Sets 1 and 2 in the mixed-quantile deviation. A similar two-step
procedure was used for CVaR regression in Rockafellar et al. (2014). Here we justify this two-step
procedure through the equivalence of deviations in CVaR and mixed-quantile quadrangles with
parameters from Sets 1 and 2.
This paper is organized as follows. Section 2 provides general results about quadrangles.
In particular, we considered quantile, mixed-quantile, and CVaR quadrangles. Sections 3 and 4
introduced and investigated the parameters from Sets 1 and 2, respectively. Section 5 provided
optimization problem statements based on CVaR and mixed-quantile quadrangles and described the
linear regression for CVaR estimation. Section 6 presented a case study and applied CVaR regression
to the financial style classification problem. The case study is posted on the web with codes, data,

113
JRFM 2019, 12, 107

and solutions. Appendix A provides convex and linear programming problems for minimization of
the Rockafellar error; Appendix B provides Portfolio Safeguard (PSG) codes implementing regression
optimization problems.

2. Quantile, Mixed-Quantile, and CVaR Quadrangles

Rockafellar and Uryasev (2013) developed a new paradigm called the risk quadrangle, which
linked risk management, reliability, statistics, and stochastic optimization theories. The risk quadrangle
methodology united risk functions for a random value X in groups (quadrangles) consisting of
five elements:

• Risk R(X), which provides a numerical surrogate for the overall hazard in X.
• Deviation D(X), which measures the “nonconstancy” in X as its uncertainty.
• Error E(X), which measures the “nonzeroness” in X.
• Regret V(X), which measures the “regret” in facing the mix of outcomes of X.
• Statistic S(X) associated with X through E and V.

These elements of a risk quadrangle are related as follows:

V (X ) = E(X ) + E(X )

R (X ) = D(X ) + E(X )
 
R(X) = min C + V(X − C)
C
 
D(X) = min E(X − C)
C
   
argmin C + V(X − C) = S(X) = argmin E(X − C)
C C

where E(X) denotes the mean of X and the statistic, S(X), can be a set, if the minimum is achieved for
multiple points.
Further, we use the following notations. The cumulative distribution function is denoted by
FX (x) = prob{X ≤ x}. The positive and negative part of a number are denoted using:
 
t, for t > 0 −t, for t < 0
[t] + = and [t]− =
0, for t ≤ 0 0, for t ≥ 0

The lower and upper VaR (quantile) are defined as follows:

lower VaR:   
sup x, FX (x) < α f or 0 < α ≤ 1
VaR−α ( X ) =  
in f x, FX (x) ≥ α f or α = 0
upper VaR:
  
in f x, FX (x) > α f or 0 ≤ α < 1
VaRα+ (X) =  
sup x, FX (x) ≤ α f or α = 1
VaR (quantile) is a set if the lower and upper quantiles do not coincide:
 
+
VaRα (X) = VaR−
α (X ), VaRα (X )

+
otherwise VaR is a singleton VaRα (X) = VaR− α (X ) = VaRα (X ).
Conditional value-at-risk (CVaR) with the confidence level α ∈ (0, 1) can be defined in many ways.
We prefer the following constructive definition:
 
1
CVaRα (X) = min C + E[X − C] +
C 1−α

114
JRFM 2019, 12, 107

In financial applications, however, the most popular definition of CVaR is

 1
1
CVaRα (X) = VaR−
β (X )dβ.
1−α α

For α = 0, CVaR0 (X) is defined as CVaR0 (X) = lim CVaRε (X) = E(X).
ε→0
For α = 1, CVaR1 (X) is defined as CVaR1 (X) = VaR− 1
(X) if a finite value of VaR−1 (X) exists.
Quadrangles are named after statistic functions. The most famous quadrangle is the quantile
quadrangle (see Rockafellar and Uryasev (2013)), named after the VaR (quantile) statistic. This
quadrangle establishes relations between the CVaR optimization technique described in Rockafellar
and Uryasev (2000, 2002) and quantile regression (see Koenker and Bassett (1978), Koenker (2005)).
In particular, it was shown that CVaR minimization and the quantile regression are similar procedures
based on the VaR statistic in the regret and error representation of risk and deviation.
Here is the definition of the quantile quadrangle for α ∈ (0, 1):
Quantile Quadrangle (Rockafellar and Uryasev (2013))

Statistic: Sα (X) = VaRα (X) = VaR (quantile) statistic.

 
Risk: Rα (X) = CVaRα (X) = min C + Vα (X − C) = CVaR risk.
C
 
Deviation: Dα (X) = CVaRα (X) − E[X] = min Eα (X − C) = CVaR deviation.
C
Regret: Vα (X) = 1
E[X]+ = average absolute loss, scaled.
1−α 
Error: Eα (X) = α
E 1−α [X]+ + [X]− = Vα (X) − E[X] = normalized Koenker–Bassett error.

The quantile quadrangle sets an example for development of more advances quadrangles. The
following mixed-quantile quadrangle includes statistic, which is equal to the weighted average of
VaRs (quantiles) with specified positive weights. Therefore, the error in this quadrangle can be used to
build a regression for the weighted average of VaRs (quantiles). Since CVaR can be approximated by a
weighted average of VaRs, the error function in this quadrangle can be used to build linear regression
for the estimation of CVaR.
Mixed-Quantile Quadrangle (Rockafellar and Uryasev (2013)).

Confidence levels αk ∈ (0, 1), k = 1, . . . , r, and weights λk > 0, rk=1 λk = 1. The error in this
quadrangle is called the Rockafellar Error.

Statistic: S(X) = rk=1 λk VaRαk (X) = mixed VaR (quantile).
r
Risk: R(X) = k=1 λk CVaRαk (X) = mixed CVaR.

Deviation: D(X) = rk=1 λk CVaRαk (X − E[X]) = mixed CVaR deviation.
  
Regret: V(X) = min r
 r
k=1 λk Vαk (X − Bk ) k=1 λk Bk = 0 = the minimal weighted average of regrets
B1 ,...,Br
+
Vαk (X − Bk ) = 1−αk E[X − Bk ] satisfying the linear constraint on B1 , . . . , Br .
1
  
Error: E(X) = min r
 r
k=1 λk Eαk (X − Bk ) k=1 λk Bk = 0 = Rockafellar error = the minimal weighted
B1 ,...,Br
 α 
average of errors Eαk (X − Bk ) = E 1−αk [X − Bk ]+ + [X − Bk ]− satisfying the linear constraint on
k
B1 , . . . , Br .

The following CVaR quadrangle can be considered as the limiting case of the mixed-quantile
quadrangle when the number of terms in this quadrangle tends to infinity. The statistic in this
quadrangle is CVaR; therefore, the error in this quadrangle can be used for the estimation of CVaR
with linear regression.

115
JRFM 2019, 12, 107

CVaR Quadrangle (Rockafellar et al. (2014)) for α ∈ (0, 1).

Statistic: Sα (X) = CVaRα (X) = CVaR.

1
Risk: Rα (X) = 1−α1
α
CVaRβ (X)dβ = CVaR2 risk.
1
Deviation: Dα (X) = Rα (X) − E[X] = 1−α 1
α
CVaRβ (X)dβ − E[X] = CVaR2 deviation.
1   +
Regret: Vα (X) = 1−α1
0
CVaRβ ( X ) dβ = CVaR2 regret.
Error: Eα (X) = Vα (X) − E[X] = CVaR2 error.

The following section proves that for a discretely distributed random value with equally probable
atoms, the CVaR quadrangle is “equivalent” to a mixed-quantile quadrangle with some parameters in
the sense that statistic, risk, and deviation in these quadrangles coincide. This fact was proved for a set
of random values with equal probabilities and variable locations of atoms.
The set of parameters considered in the following section is used in two-step CVaR regression in
Rockafellar et al. (2014).

3. Set 1 of Parameters for Mixed-Quantile Quadrangle

Set 1 of parameters for the mixed-quantile quadrangle for a discrete uniformly distributed random

value X consists of confidence levels αk ∈ (0, 1), k = 1, . . . , r and weights λk > 0 such that rk=1 λk = 1.
Parameter r depends only on the number of atoms in X and the confidence level α of the CVaR
quadrangle. We proved that statistic, risk, and deviation of the mixed-quantile quadrangle with the
Set 1 of parameters coincide with the statistic, risk, and deviation of the CVaR quadrangle.
Let X be a discrete random value with support xi and Prob X = xi = 1/ν for i = 1, 2, . . . , ν, where ν
is the number of atoms. Denote xmax = max xi . For this random value, CVaR1 (X) = VaR− 1
(X) = xmax .
i=1,...,ν

Set 1 of parameters:

• partition of the interval [α, 1]: βνα −1 = α, and βi = iδ, for i = να , να + 1, . . . , ν, where δ = 1/ν,
να = να + 1, with z being the largest integer less than or equal to z; δα = βνα − α.
δα δ
• weights: pνα = 1−α , pi = 1−α , i = να + 1, . . . , ν.
βi −βi−1
• confidence levels: γi = 1 −  , i
1−βi−1
= να , . . . , ν − 1; γν = 1.
ln 1−βi

Lemma 1. Let X be a discrete random value with ν equally probable atoms. Then, statistic, risk, and deviation
of the CVaR quadrangle for X are given by the following expressions with parameters specified by Set 1:

1. CVaR statistic:
ν
Sα (X) = CVaRα (X) = pi VaRγi (X) (1)
i = να

2. CVaR2 risk:
 1 ν
1
Rα ( X ) = CVaRβ (X)dβ = pi CVaRγi (X) (2)
1−α α i=να

3. CVaR2 deviation:
 1 ν
1
Dα (X) = CVaRβ (X)dβ − E[X] = pi CVaRγi (X) − E[X] (3)
1−α α i = να

116
JRFM 2019, 12, 107

Proof. Appendix C contains proof of the lemma. 

Note. Expression (1) is valid for arbitrary γi ∈ (βi−1 , βi ), i = να , . . . , ν. Equations (2) and (3) are valid
for arbitrary γν ∈ [βν−1 , 1].
We want to emphasize that the statement of Lemma 1 is valid for any discrete random value with
equally probable atoms. The statement does not depend upon atom locations.

Corollary 1. For the random value X defined in Lemma 1, statistic, risk, and deviation of the CVaR quadrangle
coincide with statistic, risk, and deviation of the mixed-quantile quadrangle with r = ν − να + 1, λk = pνα −1+k ,
αk = γνα −1+k , k = 1, . . . , r.

Proof. Right hand sides in Equations (1)–(3) define statistic, risk, and deviation of the mixed-quantile

quadrangle because pi > 0, i = να , . . . , ν, νi=να pi = 1 and VaRγi (X), i = να , . . . , ν, are singletons. 

Example 1. Let X be a discrete random value with five atoms (−40; −10; 20; 60 100) and equal probabilities = 0.2.

Figure 1 explains how to calculate statistic in the CVaR quadrangle and mixed-quantile quadrangle
with Set 1 of parameters for α = 0.5. Bold lines show VaR− α (X ) as a function of α. CVaR0.5 (X ) equals
the dark area under the VaRα (X) divided by 1 − α. CVaR can be calculated as integral of VaRα (X) or
as the sum of areas of rectangles. Figure 2 explains how to calculate risk in the CVaR quadrangle and
mixed-quantile quadrangle with Set 1 of parameters for α = 0.5. The bold continuous curve shows
CVaRα (X) as a function of α. Risk R0.5 (X) is equal to the area under the CVaR curve divided by 1 − α.
This area can be calculated as the integral of CVaR or as the sum of areas of rectangles. The area of
every rectangle is equal to the area under CVaR in the appropriate range of α. The equality of areas
defines values of γi . Parameters pi , γi do not depend on the values of atoms.

Figure 1. Five equally probable atoms. VaR and CVaR with Set 1 of parameters for α = 0.5.

117
JRFM 2019, 12, 107

Figure 2. Five equally probable atoms. Risk in the CVaR quadrangle and mixed-quantile quadrangle
with Set 1 of parameters for α = 0.5.

4. Set 2 of Parameters for the Mixed-Quantile Quadrangle

This section gives an alternative expression for the risk Rα (X) and deviation Dα (X) in the CVaR
quadrangle for a discrete uniformly distributed random variable. This expression is based on the
following Set 2 of parameters. This set of parameters has the same number of parameters as Set 1 but
different values of weights and confidence levels. Similar
 to Section 3, let X be a discrete random value
with support, xi , i = 1, 2, . . . , ν, and Prob X = xi = 1/ν for i = 1, 2, . . . , ν. Denote xmax = max xi .
i=1,...,ν
For this random value CVaR1 (X) = VaR−
1
(X) = xmax .
Set 2 of parameters:

• partition of the interval [βνα −1 , 1]: δ = 1/ν, βi = iδ, for i = να − 1, να , . . . , ν, where να = να + 1,

with z being the largest integer less than or equal to z; δα = βνα − α.
• confidence levels: βi , i = να − 1, να , . . . , ν.
• weights:
qν = 0;
δ δ
qν−1 = 1−α × [2 ln(2)] ≈ 1−α × 1.386294361, (if ν − 1 > να )
    
δ δ
qν−2 = 1−α × 2 3 ln 2 + ln 12 ≈ 1−α
3
× 1.046496288, (if ν − 2 > να )
  j+1   j−1 
δ
qν−j = 1−α × j ( j + 1) ln j + ( j − 1) ln j , (if j > 2, ν − j > να )
    j−1 
δ
qνα = 1−α × j δ − δα + ( j + 1) ln 1−α δj + ( j − 1) ln j , (if να < ν − 1, j = ν − να )
  
δ δ( j−1)
qνα −1 = 1−α × j δα + ( j − 1) ln 1−α , (if να − 1 < ν − 1, j = ν − να + 1)
      
δ δ δ
if να = ν − 1, then qνα = 1−α × 2 1 + ln 1−α δ − 1, qνα −1 = 1−α × 2 ln 1−α −1 +2
if να = ν, then qνα −1 = 1.

Lemma 2. Let X be a discrete random value with ν equally probable atoms. Then, risk and deviation of the
CVaR quadrangle for X are given by the following expressions with parameters from Set 2.

1. CVaR2 Risk:
 1 ν−1
1
Rα ( X ) = CVaRβ (X)dβ = qi CVaRβi (X) (4)
1−α α i=να −1

118
JRFM 2019, 12, 107

2. CVaR2 Deviation:
 1 ν−1
1
Dα (X) = CVaRβ (X)dβ − E[X] = qi CVaRβi (X) − E[X] (5)
1−α α i=να −1

Proof. Appendix D contains proof of the lemma. 

Note. Equations (4) and (5) are valid if CVaRβν−1 is replaced by CVaRγ with an arbitrary γ ∈ [βν−1 , 1].

Corollary 2. For the random value X defined in Lemma 2, risk and deviation of the CVaR quadrangle coincide
with risk and deviation of the mixed-quantile quadrangle with r = ν − να + 1, λk = qνα −2+k , αk = βνα −2+k ,
k = 1, . . . , r.

Proof. Right hand sides in Equations (4) and (5) define risk and deviation of the mixed-quantile

quadrangle because qi > 0, i = να − 1, . . . , ν − 1, and ν−1
i=να −1 qi = 1. 
 
Lemma 3. Let X be a discrete random value with equally probable atoms xi , i = 1, 2, . . . , ν, Prob X = xi = 1/ν.
Then, statistic of the mixed-quantile quadrangle defined by the Set 2 of parameters is a range containing statistic
of the CVaR quadrangle.

Proof. Appendix E contains proof of the lemma. 

5. On the Estimation of CVaR with Mixed-Quantile Linear Regression

This section formulates regression problems using the CVaR quadrangle and mixed-quantile
quadrangle. For discrete final distributions with equally probable atoms, we prove some equivalence
statements for the CVaR and mixed-quantile quadrangles. Further, we demonstrate how to estimate
CVaR by using the linear regression with error and deviation from the mixed-quantile quadrangle.
We want to estimate variable V using a linear function f (Y) = C0 + CT Y of the explanatory factors
Y = (Y1 , . . . , Yn ). Let 
E be an error from some quadrangle (further we consider, mixed-quantile and
CVaR quadrangles), and D  and S be a deviation and a statistic, respectively, corresponding to this
quadrangle. Below we consider optimization statements for solving regression problems.
General Optimization Problem 1
Minimize error 
E and find optimal C∗0 , C∗

min 
E (Z(C0 , C))
C0 ∈R, C∈Rm

where Z(C0 , C) = V − C0 − CT Y.
General Optimization Problem 2

• Step 1. Find an optimal vector C∗ by minimizing deviation:

 (Z0 (C))
min D
C∈Rm

where Z0 (C) = V − CT Y.
• Step 2. Assign C∗0 :
(Z0 (C∗ ))
C∗0 ∈ S

119
JRFM 2019, 12, 107

Error 
E (X) is called nondegenerate if:

in f 
E (X) > 0 for constants D  0.
X:EX=D

Rockafellar et al. (2008), p. 722, proved the following decomposition theorem.

Theorem 1. (Error-Shaping Decomposition of Regression). Let  E be a nondegenerate error and D  = min

  C
  be the associated statistic. Point (C∗ , C∗ ) is a solution
E(X − C) be the corresponding deviation, and let S 0
of the General Optimization Problem 1 if and only if C∗ is a solution of the General Optimization Problem 2,
(Z0 (C∗ )) with Step 2.
Step 1 and C∗0 ∈ S

According to the decomposition theorem, when   and S

E, D,  are elements of the CVaR quadrangle,
the following Optimization Problems 1 and 2 are equivalent.
Optimization Problem 1
Minimize error from the CVaR quadrangle:

min Eα (Z(C0 , C))

C0 ∈R, C∈Rm

where Z(C0 , C) = V − C0 − CT Y.
Optimization Problem 2

• Step 1. Find an optimal vector C∗ by minimizing deviation from the CVaR quadrangle:

min Dα (Z0 (C))

C∈Rm

where Z0 (C) = V − CT Y.
• Step 2. Calculate:
C∗0 = CVaRα (Z0 (C∗ ))

In Optimization Problem 2 in Step 2 the statistic equals S(Z0 (C∗ )) = CVaRα (Z0 (C∗ )), which is the
specification of the inclusion operation in the Optimization Problem General 2 in Step 2.
Optimization Problem 2 is used in Rockafellar et al. (2014) for the construction of the linear
regression algorithms for estimating CVaR.
According to the decomposition theorem, when   and S
E, D,  are elements of a mixed-quantile
quadrangle, the following Optimization Problems 3 and 4 are equivalent.
Optimization Problem 3
Minimize error from the mixed-quantile quadrangle:

min E(Z(C0 , C))

C0 ∈R, C∈Rm

Optimization Problem 4

• Step 1. Find an optimal vector C∗ by minimizing deviation from the mixed-quantile quadrangle:

min D(Z0 (C))

C∈Rm

120
JRFM 2019, 12, 107

• Step 2. Assign:
C∗0 ∈ S(Z0 (C∗ ))

Corollaries 1 and 2 can be used for constructing the linear regression for estimating CVaR. Let Y be
a random vector of factors for estimating the random value V. We consider that the linear regression
function f (Y) = C0 + CT Y approximates CVaR of V, where  C0 ∈ R , C ∈ R are variables in the linear
m

regression. The residual is denoted by Z(C0 , C) = V − C0 + C Y and Z0 (C) = V − CT Y.

T

Further we provide a lemma about linear regression problems based on Corollary 1 with the
Set 1 of parameters. The main statement here is that the Optimization Problems 3 and 4 for the
mixed-quantile quadrangle can be used to solve linear regression problems for estimating CVaR. This
is the case because CVaR and mixed-quantile quadrangles have the same Statistic and Deviation.
 
Lemma 4. Let the residual random value Z(C0 , C) = V − C0 + CT Y be discretely distributed with ν
equally probable atoms. Let us consider the CVaR quadrangle with error Eα (Z(C0 , C)), deviation Dα (Z0 (C)),
and statistic Sα (Z0 (C∗ )). Let us also consider the mixed-quantile quadrangle with the error E(Z(C0 , C)),
deviation D(Z0 (C)), and statistic S(Z0 (C∗ )) with parameters defined by Set 1 and r = ν − να + 1, λk = pνα −1+k ,
αk = γνα −1+k , k = 1, . . . , r. Then, the Optimization Problems 1–4 are
 equivalent, i.e., the sets of optimal
vectors of these optimization problems coincide. Moreover, let C∗0 , C∗ be a solution vector of the equivalent
Optimization Problems 1–4. Then:
C∗0 = CVaRα (Z0 (C∗ ))
   
Eα C∗0 , C∗ = E C∗0 , C∗ = D(Z0 (C∗ ))= Dα (Z0 (C∗ ))

Proof. This lemma is a direct corollary of the decomposition Theorem 1 and Corollary 1 of
Lemma 1. Indeed, Corollary 1 implies that the Optimization Problems 2 and 4 are equivalent.
Further, the decomposition theorem implies that Optimization Problems 1 and 2 and the Optimization
Problems 3 and 4 are equivalent. 

Further we provide a lemma about linear regression problems based on Corollary 2 with the Set 2
of parameters. The main statement is that the Optimization Problem 4, Step 1 for the Mixed-Quantile
Quadrangle can be used to solve linear regression problem for estimating CVaR. This is the case
because CVaR and Mixed-Quantile Quadrangles have the same Deviation. After obtaining vector of
coefficients C∗ , intercept is calculated, C∗0 = CVaRα (Z0 (C∗ )).
 
Lemma 5. Let the residual random value Z(C0 , C) = V − C0 + CT Y be discretely distributed with ν equally
probable atoms. Let the mixed-quantile quadrangle with deviation D(Z0(C)) be defined by parameters of Set 2
and r = ν − να + 1, λk = qνα −2+k , αk = βνα −2+k , k = 1, . . . , r. Then, C∗0 , C∗ is a solution of Optimization
 
Problem 1, if and only if, C∗0 , C∗ is a solution of the following two-step procedure:

• Step 1. Find an optimal vector C∗ by minimizing deviation from the mixed-quantile quadrangle:

min D(Z0 (C))

C∈Rm

• Step 2. Calculate C∗0 = CVaRα (Z0 (C∗ )).

Proof. This lemma is a direct corollary of the decomposition Theorem 1 and Corollary 2 or Lemma 2.
Indeed, the Corollary 2 implies that the Optimization Problems 2 and 4 are equivalent. Further, since
deviations of the CVaR and mixed-quantile quadrangles coincide, we can use Step 1 to calculate
optimal coefficients C∗ . Further, the intercept is calculated with C∗0 = CVaRα (Z0 (C∗ )) because CVaR is
the statistic in the CVaR quadrangle. 

121
JRFM 2019, 12, 107

For the Set 2 of parameters, the deviations in the CVaR and mixed-quintile quadrangles coincide.
The two-step procedure in Optimization Problem 4 can be used to solve linear regression problems
with Set 2 parameters for the mixed-quantile deviation. Also, the minimization of the Rockafellar error
with the Set 2 of parameters may result in a correct C∗ . The statistic of CVaR belongs to the statistic of
the mixed-quantile quadrangle. Therefore, the optimization of the Rockafellar error with the Set 2 of
parameters may lead to a wrong value of intercept C∗0 . This potential incorrectness can be fixed by
assigning C∗0 = CVaRα (Z0 (C∗ )).

6. Case Study: Estimation of CVaR with Linear Regression and Style Classification of Funds
The case study described in this section is posted online (see Case Study (2016)). The codes and
data are available for downloading and verification. Every optimization problem is presented in three
formats: Text, MATLAB, and R. Calculations were done with a PC with a 3.14 GHz processor.
We have applied CVaR regression to the return-based style classification of a mutual fund.
We regress a fund return by several indices as explanatory factors. The estimated coefficients represent
the fund’s style with respect to each of the indices.
A similar problem with a standard regression based on the mean squared error was considered
by Carhart (1997) and Sharpe (1992). They estimated the conditional expectation of a fund
return distribution (under the condition that a realization of explanatory factors is observed).
Basset and Chen (2001) extended this approach and conducted the style analyses of quantiles of
the return distribution. This extension is based on the quantile regression suggested by Koenker and
Bassett (1978). The Case Study (2014), “Style Classification with Quantile Regression” implemented
this approach and applied quantile regression to the return-based style classification of a mutual fund.
For the numerical implementation of CVaR linear regression, we used the Portfolio Safeguard (2018)
package. Portfolio Safeguard (PSG) can solve nonlinear and mixed-integer nonlinear optimization
problems. A special feature of PSG is that it includes precoded nonlinear functions: CVaR2 error
(cvar2_err) and CVaR2 deviation (cvar2_dev) from the CVaR quadrangle, Rockafellar error (ro_err)
from the mixed-quantile quadrangle, and CVaR deviation (cvar_dev) from the quantile quadrangle.
We implemented the following equivalent variants of CVaR regression:

• Minimization of the CVaR2 error (PSG function cvar2_err).

• Two-step procedure with CVaR2 deviation (PSG function cvar2_dev).
• Minimization of the Rockafellar error (PSG function ro_err) with the Set 1 of parameters.
• The two-step procedure using mixed CVaR deviation with the Set 1 and Set 2 of parameters.
This is calculated as a weighted sum of CVaR deviations (PSG function cvar_dev) from the
quantile quadrangle.

PSG automatically converts the analytic problem formulations to the mathematical programming
codes and solves them. We included in Appendix A convex and linear programming problems for the
minimization of the Rockafellar error with the Set 1 of parameters. These formulations are provided
for verification purposes. They can be implemented with standard commercial software. For instance,
the linear programming formulation can be implemented with the Gurobi optimization package.
If Gurobi is installed in the computer, PSG can use Gurobi code as a subsolver. With the CARGRB
solver in PSG, by setting the linearize option to 1, it is possible to solve the linear programming problem
with Gurobi. However, this conversion will deteriorate the performance, compared to the default PSG
solver VAN. For small problems it will not be noticeable. However, for problems with a large number
of scenarios (e.g., with 108 observations), the standard PSG solver VAN will dramatically outperform
the Gurobi linear programming implementation. In this case, Gurobi may not even start on a small
PC because of a shortage of memory. Nevertheless, if the number of observations is small (e.g., 103 )
and the number of factors is very large (e.g., 107 ), it is recommended that the linear programming
formulation is used.

122
JRFM 2019, 12, 107

We regressed the CVaR of the return distribution of the Fidelity Magellan Fund on the explanatory
variables: Russell 1000 Growth Index (RLG), Russell 1000 Value Index (RLV), Russell Value Index
(RUJ), and Russell 2000 Growth Index (RUO). The dataset includes 1264 historical daily returns of the
Magellan Fund and the indices, which were downloaded from the Yahoo Finance website. The data
(design matrix for the regression) is posted on the Case Study (2016) website.
The CVaR regression was done with the confidence levels α = 0.75 and α = 0.9. Calculation
results are in Tables 1 and 2, respectively. Here is the description of the columns of the tables:

• Optimization Problem #: Optimization Problem number, as denoted in Section 5; also it is the

problem number in the case study posted online, see, Case Study (2016).
• Set #: Set of parameters for the mixed-quantile quadrangle.
• Objective: Optimal value of the objective function.
• RLG: coefficient for the Russell Value Index.
• RLV: coefficient for the Russell 1000 Value Index.
• RUJ: coefficient for the Russell Value Index.
• RUO: coefficient for the Russell 2000 Growth Index.
• Intercept: regression intercept.
• Solving Time: solver optimization time.

Table 1. Optimization outputs: estimating CVaR with the linear regression, α = 0.75.

Optimization Solving
Set # Objective RLG RLV RUJ RUO Intercept
Problem # Time (s)
1 N/A 0.01248 0.486 0.581 −0.0753 −6.22 × 10−3 6.98 × 10−3 0.02
2 N/A 0.01248 0.486 0.582 −0.0753 −6.22 × 10−3 6.98 × 10−3 0.02
3 Set 1 0.01247 0.486 0.582 −0.0753 −6.22 × 10−3 6.96 × 10−3 0.03
4 Set 1 0.01251 0.486 0.582 −0.0752 −6.22 × 10−3 6.98 × 10−3 0.11
4 Set 2 0.01248 0.486 0.582 −0.0753 −6.23 × 10−3 6.98 × 10−3 0.03

Table 2. Optimization outputs: estimating CVaR with the linear regression, α = 0.9.

Optimization Solving
Set # Objective RLG RLV RUJ RUO Intercept
Problem # Time (s)
1 N/A 0.016656 0.472 0.606 −0.078 −7.052 × 10−3 1.05 × 10−2 0.02
2 N/A 0.016656 0.472 0.606 −0.078 −7.052 × 10−3 1.05 × 10−2 0.02
3 Set 1 0.016656 0.472 0.606 −0.078 −7.052 × 10−3 1.05 × 10−2 0.02
4 Set 1 0.016656 0.472 0.606 −0.078 −7.052 × 10−3 1.05 × 10−2 0.11
4 Set 2 0.016656 0.472 0.606 −0.078 −7.052 × 10−3 1.05 × 10−2 0.04

Tables 1 and 2 show calculation results for the considered equivalent problems. We observe that
regression coefficients coincide for all problems in Tables 1 and 2. This confirms the correctness of
theoretical results and the numerical implementation. Also, we want to point out that the regression
coefficients are quite similar for α = 0.75 (Table 1) and α = 0.9 (Table 2).
The calculation time in majority of cases was around 0.02–0.04 s, except for the case with the
mixed CVaR deviation for Set 1, which took 0.11 s. The PSG calculation times were quite low because
the solver “knows” analytical expressions for the functions and can take advantage of this knowledge.

7. Conclusions
The quadrangle risk theory Rockafellar and Uryasev (2013) and the decomposition theorem
Rockafellar et al. (2008) provided a framework for building a regression with relevant deviations.
Solution of a regression problem is split in two steps: (1) minimization of deviation from the
corresponding quadrangle, and (2) determining of intercept by using statistic from this quadrangle.
For CVaR regression, Rockafellar et al. (2014) reduced the optimization problem at Step 1 to

123
JRFM 2019, 12, 107

a high-dimension linear programming problem. We suggested two sets of parameters for the
mixed-quantile quadrangle and investigated its relationship with the CVaR quadrangle. The Set 1 of
parameters corresponds to CVaR regression in Rockafellar et al. (2014), where the Set 2 is a new set
of parameters.
For the Set 1 of parameters, the minimization of error from CVaR Quadrangle was reduced to the
minimization of the Rockafellar error from the mixed-quantile quadrangle. For both sets of parameters,
the minimization of deviation in CVaR quadrangle is equivalent to the minimization of deviation in
mixed-quantile quadrangle.
We presented optimization problem statements for CVaR regression problems using CVaR
and mixed-quantile quadrangles. Linear regression problem for estimating CVaR were efficiently
implemented in Portfolio Safeguard (2018) with convex and linear programming. We have done a case
study for the return-based style classification of a mutual fund with CVaR regression. We regressed
the fund return by several indices as explanatory factors. Numerical results validating the theoretical
statements are placed to the web (see Case Study (2016)).

Supplementary Materials: Data and codes used in the case study can be downloaded: 1. Case Study (2016):
Estimation of CVaR through Explanatory Factors with CVaR (Superquantile) Regression. [Link]
edu/uryasev/research/testproblems/financial_engineering/on-implementation-of-cvar-regression/. 2. Case Study
(2014): Style Classification with Quantile Regression. [Link]
financial_engineering/style-classification-with-quantile-regression/.
Author Contributions: Conceptualization, S.U.; Formal analysis, V.K.; Investigation, A.G.; Methodology, S.U.;
Software, V.K.; Supervision, S.U.; Writing—original draft, A.G.
Funding: Research of Stan Uryasev was partially funded by the AFOSR grant FA9550-18-1-0391 on Massively
Parallel Approaches for Buffered Probability Optimization and Applications.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A CVaR Regression with Rockafellar Error: Convex and Linear Programming
The value of the Rockafellar error with given set of parameters λk , αk (αk ∈ (0, 1), k =

1, . . . r, rk=1 λk = 1) for a random value X is a minimum w.r.t. a set of variables B1 , . . . , Br of a
mixture of Koenker–Bassett Error functions with one linear constraint on these variables:
⎧  ⎫
⎪ ⎪
⎨ 
r r
⎪ ⎪
⎬
Rocka f ellar_Error (X)λ1 ,α1 ,...,λr ,αr = min ⎪ λk Eαk (X − Bk ) λk B k = 0 ⎪
B1 ,...,Br ⎪
⎩  ⎪
⎭
k =1 k =1
 α 
where Eαk (X − Bk ) = E 1−αk [X − Bk ]+ + [X − Bk ]− is the normalized Koenker–Bassett error.
k
By using regret from the mixed-quantile quadrangle, we express the Rockafellar error as follows:
⎧  ⎫
⎪ ⎪
⎨ λk 
r r
⎪ + ⎪
⎬
Rocka f ellar_Error (X)λ1 ,α1 ,...,λr ,αr = min ⎪ E[X − Bk ]  λk Bk = 0⎪ − E[X ].
B1 ,...,Br ⎪
⎩ 1 − αk  k =1 ⎪
⎭
k =1

For the linear regression problem, the random variable X is defined by a set of differences
between observed values Vi and linear functions C0 + CT Yi , where Yi is a vector of explanatory factors,
i = 1, 2, . . . , ν. Vectors C and Y have m components, C = (C1 , . . . , Cm ), Y = (Y1 , . . . , Ym ), and C0
is a scalar. Residuals Xi = Vi − C0 − CT Yi are values (scenarios) of atoms of the random value X.
We consider that all atoms have equal probabilities. The estimation of V with factors Y is done by
minimizing the error w.r.t. variables C, C0 . Further we use the Set 1 of parameters. Let us denote:

1 1
v v
E[V ] = Vi , E [ Y ] = Yi .
v v
i=1 i=1

124
JRFM 2019, 12, 107

Appendix A.1 Convex Programming Formulation for CVaR Regression

Minimize the Rockafellar error:
⎧ ⎫
⎪ ⎪
⎨ 
r v  +
⎪ λk ⎪
⎬
min ⎪ Vi − C0 − C Yi − Bk − E[V ] + C0 + C E[Y]⎪
T T
(A1)
B1 ,...,Br ,C0 ,C1 ,..., Cm ,⎪
⎩ ( 1 − αk ) v ⎪
⎭
k =1 i=1

subject to the constraint:


r
λk Bk = 0. (A2)
k =1

This optimization problem has a convex objective and one linear constraint.

Appendix A.2 Linear Programming Formulation for CVaR Regression

Equations (A1) and (A2) are reduced to linear programming with additional variables
and constraints:
⎧ ⎫
⎪ ⎪
⎨ 
r v
⎪ λk ⎪
⎬
min ⎪
⎪ A ki − E [ V ] + C0 + CT
E [ Y ] ⎪
⎪ (A3)
A ,...,A ⎩ ( 1 − αk ) v ⎭
11 rv k =1 i=1
B1 , . . . , Br , C0 , C1 , . . . , Cm ,

subject to constraints:

r
λk Bk = 0 (A4)
k =1

Aki ≥ Vi − C0 − CT Yi − Bk , k = 1, . . . , r, i = 1, . . . , v (A5)

Aki ≥ 0, k = 1, . . . , r, i = 1, . . . , v (A6)
∗T
The linear function C∗0
+ C Y estimates CVaRα (V ) as a function of explanatory factors Y, where
C∗0 and C∗ are optimal values of variables for Equations (A1) and (A2) or (A3)–(A6).

Appendix B Codes Implementing Regression Optimization Problems

This appendix contains codes implementing Optimization Problems 1–4 described in Section 5.
Codes and solution results are posted at internet link: Case Study (2016). Codes are written in Portfolio
Safeguard (PSG) Text, MATLAB, and R environments. Here are codes in the Text environment.
Optimization Problem 1
Code in PSG Text format:

minimize
cvar2_err(0.75,matrix_s)

The keyword “minimize” indicates that the objective function is minimized. The objective function
cvar2_err(0.75,matrix_s) calculates error Eα (Z(C0 , C)) in CVaR quadrangle with confidence level
α = 0.75. The matrix_s contains scenarios of residual of the regression Z(C0 , C) = V − C0 − CT Y.
Optimization Problem 2
Code in PSG Text format:

minimize
cvar2_dev(0.75,matrix_s)
value:
cvar_risk(0.75,matrix_s)

125
JRFM 2019, 12, 107

The code includes two parts. The first part begins with the keyword “minimize” indicating that
the objective function is minimized. It implements Step 1 of the Optimization Problem 2, which
minimizes deviation from the CVaR quadrangle (for determining the optimal vector C∗ of regression
coefficients without an intercept). The PSG function cvar2_dev(0.75,matrix_s) calculates deviation
Dα (Z0 (C)) with α = 0.75. The matrix_s contains scenarios of the residual of the regression without an
intercept Z0 (C) = V − CT Y.
The second part of the code begins with the keyword “value.” This part implements Step 2
of the Optimization Problem 2 for calculating the optimal value of intercept C∗0 . The PSG function
cvar_risk(0.75,matrix_s) calculates CVaRα (Z0 (C∗ )) with α = 0.75 at the optimal point C∗ .
Optimization Problem 3
Code in PSG Text format with parameters from Set 1.

minimize
ro_err(matrix_s, matrix_coeff)

The keyword “minimize” indicates that the objective function is minimized. The objective function
ro_err(matrix_s, matrix_coeff) calculates the Rockafellar error E(Z(C0 , C)) in the mixed-quantile
quadrangle. The matrix_s contains scenarios of the residual of the regression Z(C0 , C) = V − C0 − CT Y.
The matrix_coeff includes vectors of weights and confidence levels for Set 1 with α = 0.75.
Optimization Problem 4
Code in PSG Text format with parameters from Set 1 for α = 0.75.

minimize
vector_c*cvar_dev(vector_a, matrix_s)
value:
cvar_risk(0.75,matrix_s)

The code includes two parts. The first part begins with the keyword “minimize” indicating
that the objective function is minimized. It implements Step 1 of the Optimization Problem 4,
which minimizes deviation from the mixed-quantile quadrangle (for determining optimal vector
C∗ of regression coefficients without intercept). The inner product vector_c*cvar_dev(vector_a,
matrix_s) calculates the mixed CVaR deviation D(Z0 (C)). The function cvar_dev corresponds to the
CVaR deviation from the quantile quadrangle. Vector vector_c contains weights for CVaR Deviation
mix corresponding to Set 1. Vector vector_a contains confidence levels defined by Set 1. The matrix_s
contains scenarios of the residual of the regression Z(C0 , C) = V − C0 − CT Y.
The second part of the code begins with the keyword “value.” This part implements Step 2
of the Optimization Problem 4, calculating the optimal value of intercept C∗0 . The PSG function
cvar_risk(0.75,matrix_s) calculates CVaRα (Z0 (C∗ )) with α = 0.75 at the optimal point C∗ .

Appendix C Proof of the Lemma 1

Proof. According to the definition, βv = 1. However, while proving this lemma, we consider βv a bit
smaller than 1, i.e., βv = 1 − ε > βv−1 , ε > 0, to avoid division by 0. Then, we will consider limit ε → 0
to finish the proof. Note that for α < 1 and for the considered partition, βi−1 < βi for all i = να , . . . , ν.
First, let us prove that γi ∈ (βi−1 , βi ) for βi < 1. Consider three functions of σ, δ for σ > 0, δ ≥ 0:
 
δ δ δ
f1 ( δ ) = , f2 (δ) = ln 1 + , f3 (δ) =
σ+δ σ σ

When δ = 0, all functions equal to 0 and have equal derivatives. When δ > 0, it is true for
derivatives that:
f  1 (δ) < f  2 (δ) < f  3 (δ)

126
JRFM 2019, 12, 107

Hence, when > 0, similar inequalities are valid for functions:

 
δ δ δ
< ln 1 + < .
σ+δ σ σ
 
δ
There exist some δγ such that 0 < δγ < δ and = ln 1 + σδ . If δ = βi − βi−1 , σ = 1 − βi ,
σ + δγ
γ = 1 − σ − δγ , then γ = γi and βi−1 < γ < βi . Therefore, γi ∈ (βi−1 , βi ).
 1−ε
Further, we show how to calculate the integral α CVaRβ (X)dβ as a sum of integrals over
the partition:
 ν 
1 1−ε
1  βi
Rα (X) = lim CVaRβ (X)dβ = lim CVaRβ (X)dβ.
ε→0 1 − α α βv →1 1 − α βi−1
i=να

Let us denote Cβi = CVaRβi (X) and Vi = VaRγi (X). Note that VaRγ (X) is a singleton for every
γ ∈ (βi−1 , βi ) and it equals Vi , because γi ∈ (βi−1 , βi ).
Below, we use value Vi for the closed interval [βi−1 , βi ] while calculating the integral over this
interval because the value of the integral does not depend on the finite values of VaRγ (X) at the
boundary points βi−1 , βi .
Using the definition of CVaR (Rockafellar and Uryasev (2002), Proposition 8 CVaR for scenario
models) we write:
 βi  βi  
βi−1
CVaRβ (X)dβ = Cβi (1 − βi ) + Vi (βi − β) dβ
1
βi−1 1−β
β  β β −β
= Cβi (1 − βi ) β i 1−β1
dβ + Vi β i 1−β
i
dβ (A7)
 1−β 
i−1  i−1  1−β 
= Cβi (1 − βi ) ln 1−β + Vi (βi − βi−1 ) − (1 − βi ) ln 1−βi−1
i−1
i i

Let us make the transformation of Equation (A7) using the expression for γi in the Set 1 definition:
 βi   1−β   
βi−1
CVaRβ (X)dβ = (βi − βi−1 ) Vi + βi −β1 i−1 ln 1−βi−1i Cβi − Vi (1 − βi )
     
= (βi − βi−1 ) Vi + 1−γ1
i
Cβi − Vi (1 − βi ) = (βi − βi−1 ) 1−γ
1
i
Cβi (1 − βi ) + Vi (βi − γi )
= (βi − βi−1 )CVaRγi (X)

The last equality is valid because γi ∈ (βi−1 , βi ).

Taking into account that lim γν = 1, CVaR1−ε (X) = CVaR1 (X), and VaR1−ε (X) = VaR1 (X),
ε→0
we obtain:

ν
βi − βi−1 
ν
Rα (X) = lim CVaRγi (X) = pi CVaRγi (X)
ε→0 1−α
i = να i = να

Deviation is calculated as follows:


ν
Dα (X) = pi CVaRγi (X) − E[X]
i = να

By the definition of CVaR (Rockafellar and Uryasev (2002)):


ν
CVaRα (X) = pi VaRγi (X) = Sα (X)
i=να

Lemma 1 is proved. 

127
JRFM 2019, 12, 107

Appendix D Proof of Lemma 2

Proof. Similar to proof of Lemma 1, we consider βv as a bit smaller than 1, i.e., βv = 1 − ε > βv−1 , ε > 0,
to avoid division by 0. Then, we consider limit ε → 0 to finish the proof. We denote Cβi = CVaRβi (X)
and Vi = VaRγ (X) for any γ ∈ (βi−1 , βi ) because VaRγ (X) does not change on this interval.
Additionally we denote δi = βi − βi−1 , i = να , . . . , ν, and σi = 1 − βi , i = να − 1, . . . , ν. Note that
all δi > 0, all σi > 0, and δi = δ j = δ for all i, j < ν.
 ν 
1 1−ε
1  βi
Rα (X) = lim CVaRβ (X)dβ = lim CVaRβ (X)dβ. (A8)
ε→0 1 − α α βv →1 1 − α βi−1
i=να

Equation (A7) in Lemma 1 is valid for any interval [θ, βi ] such that βi−1 ≤ θ ≤ βi , therefore:
 βi " #
1−θ 1−θ
CVaRβ (X)dβ = Cβi (1 − βi ) ln + Vi (βi − θ) − (1 − βi ) ln (A9)
θ 1 − βi 1 − βi

Let us express Vi from CVaRβi−1 (X) using the definition of CVaR from
Rockafellar and Uryasev (2002) (Equation (25)), then insert this expression into Equation (A9):
   
Cβi−1 = 1
Cβi (1 − βi ) + Vi (βi − βi−1 ) = 1
σi−1 Cβi σi + Vi δi ,
1−βi−1  
Vi = δ1 Cβi−1 σi−1 − Cβi σi
i

By substituting Vi into Equation (A9), we obtain:

 " #
βi
1−θ 1  1−θ
CVaRβ (X)dβ = Cβi σi ln + Cβi−1 σi−1 − Cβi σi (βi − θ) − σi ln (A10)
θ σi δi σi

The last equation contains two CVaRs, Cβi and Cβi−1 . Let us express the coefficients of these CVaRs.
Cβi in Equation (A10) has the following coefficient:
" #
1−θ σ βi − θ
q1i = σi ln 1+ i − (A11)
σi δi δi

and Cβi−1 has the coefficient:

" #
σi−1 1−θ
q2i−1 = (βi − θ) − σi ln (A12)
δi σi
Coefficient q2i−1 ≥ 0 because the value in square brackets is the same as in Equations (A7) and
(A9), and is a result of the integration of a non-negative function over the interval [θ, βi ]. The coefficient

σ δ −βi +β

q1i ≥ 0 because it is an integral of the non-negative function δi i 1−β over the same interval.
 1−ε i

When summing up integrals to obtain 1−α α CVaRβ (X)dβ, every Cβi , aside from Cβνα −1 and
1

Cβν , enters the sum two times with coefficients depending on ν, α, βi−1 , βi , and βi+1 . Once in Equation
(A11) for i, and the second time in Equation (A12) for i + 1. All coefficients are non-negative.
Let us explain this in more detail.
If i is such that να < i < ν, then βνα < βi < βν and θ = βi−1 for i in Equation (A11) and θ = βi for
i + 1 in Equation (A12). Then, the coefficient for Cβi in Equation (A8) equals:
  
σ σ
   σ
  σi

qi = 1−α (q1i + q2i ) = ln σi−1 δi−1 − 1 + δ i δi+1
1 1
1−α σi i σ  i σ  σi+1
− σi+1 ln σi + 1
σi σi−1 +
= 1−α δi ln σi−1i − δi+1 ln σi+i 1
i 1

128
JRFM 2019, 12, 107

If i = να < ν, then θ = α in Equation (A11) and θ = βi in Equation (A12). Then, the coefficient for
Cβi in Equation (A8) equals:
    σ  δ  σi
  σ 
qi = 1−α (q1i + q2i ) = δi − δi + δi+1 δi+
1 1
σi ln 1−α
σi
i−1 α
1 − σi+1 ln σi+1
i

σi

1−α
δα σi−1
 σi + 1 σi
= 1−α 1 − δ + δ ln σ − δ ln σ
i i
1−α
i i+1 i+1

If i = να − 1, then Cβi enters in the sum only in Equation (A12) with θ = α. Then:
" #
1 σi δα σi + 1 1−α
qi = q2i = − ln
1−α 1 − α δi+1 δi+1 σi+1

If i = ν > να , then Cβi enters in the sum only in Equation (A11) with θ = βi−1 . Then:
" #
1 σi σ σi−1
qi = q1i = ln i−1 −1
1−α 1−α σi δi

Also, in the case when i = ν = να , Cβi enters in the sum only in Equation (A11) with θ = α. Then:
" #
1 σi 1−α σ δα
qi = q1i = ln 1+ i −
1−α 1−α σi δi δi

Then, the risk in the CVaR quadrangle equals:

ν
Rα (X) = lim qi CVaRβi (X).
ε→0 i=να −1

Deviation is calculated as follows: Dα (X) = Rα (X) − E[X].

Taking into account that all βi are fixed for i < ν, the limit operation affects only coefficients
for i = ν. Namely, lim βν = 1, lim σν = 0, lim σν ln(σν ) = 0, lim δν = δ. Then, the limit values of
ε→0 ε→0 ε→0 ε→0
coefficients qi are equal to:

qν = 0 in both cases, when ν > να and ν = να .

σ
  σ 
qi = 1−α
1
× δi σi−1 ln σi−1 for i such that να < i = ν − 1.
σ
 σ i   σ 
qi = 1−α
1
× δi σi−1 ln σi−1 + σi+1 ln σi+1 for i such that να < i < ν − 1.
σ
 i   i  σ 
i+1
qi = 1−α
1
× δi δ − δα + σi−1 ln 1−α σi + σ i+1 ln σi for i = να < ν − 1.
σ
  σ

+
qi = 1−α
1
× δi δα + σi+1 ln 1−α for i = να − 1 < ν − 1.
i 1

If the number of atoms used in calculating Rα (X) is 3 or 2, then the coefficients have values:
σ
  
qi = 1−α
1
× δi δ − δα + σi−1 ln 1−α
σ for i = να = ν − 1.
i
σi
qi = δ [δα ]
= 1 for i = να − 1 = ν − 1.
1
1−α ×

It can be shown that νi=να −1 qi = 1 by sequentially summing up coefficients.
By recalling that σi = δ(v − i), we can rewrite the equations for coefficients qi using δ and j = v − i.
Lemma 2 is proved. 

Appendix E Proof of the Lemma 3

Proof. Distribution of random value X defines the partition of the interval [0, 1]: δ = 1/ν, βi = iδ,
for i = 0, 1, . . . , ν.

129
JRFM 2019, 12, 107

CDF FX (x) = prob{X ≤ x} is a non-decreasing and right-continuous function and it is

constant for every right-open interval [βi−1 , βi ), i = 1, . . . , ν. Therefore, from the definitions of
+
VaR−γ (X ) and VaRγ (X ), we have:

+
VaR−
γ (X ) = VaRγ (X ) = VaRγ (X ) = const for γ ∈ (βi−1 , βi ).

VaRβ+ (X) = VaRγ (X) for γ ∈ (βi , βi+1 ) and i = 0, . . . , ν − 1.

i

VaR−
β (X ) = VaRγ (X ) for γ ∈ (βi−1 , βi ) and i = 1, . . . , ν.
i

For i = 0 we have βi = 0 and VaR−0

(X) = −∞.
Thus, according to the definition of statistic for the mixed-quantile quadrangle, for Set 2:
ν−1
α (X ) =
SII
i=να −1
qi VaRβi (X)

where VaRβi (X) are intervals (that may have zero lengths for some i), and qi and να are defined
above for the Set 2. Therefore, SII α (X ) is also an interval and it has a non-zero length if not all
VaR−
β
( X ) , i = να − 1, . . . , ν − 1 are equal.
i
If να − 1 = 0, then SII α (X ) is a left-open interval with the lower bound −∞.
Let να − 1 > 0. For the Set 1, we defined confidence levels as internal points in intervals
γi ∈ (βi−1 , βi ). Using these definitions of γi , we can express SIIα (X ) as the interval:
ν−1 ν−1 
α (X ) =
SII
i=να −1
qi VaRγi (X),
i=να −1
qi VaRγi+1 (X)

To simplify notations, let us denote: Vi = VaRγi (X); Cγ = CVaRγ (X); and L, U are bounds for
α (X ) such that Sα (X ) = [L, U ].
SII II

Statistic Sα (X) of the CVaR quadrangle is defined in Lemma 1 (see Equation (1)) as Sα (X) =
ν
i=να pi Vi .
Therefore, we wish to prove that:
ν−1 ν−1
L= qi Vi ≤ Sα (X) ≤ q i Vi + 1 = U
i=να −1 i=να −1

Let us prove the right inequality for the upper bound.

According to Lemmas 1 and 2, it is valid that:
ν ν−1
Rα ( X ) = pi Cγi = qi Cβi
i = να i=να −1

Let di = qi−1 − pi . Then, from the last equality:

ν ν ν ν
pi Cγi = qi−1 Cβi−1 = pi Cβi−1 + di Cβi−1
i = να i = να i = να i=να


ν 
ν  
di Cβi−1 = pi Cγi − Cβi−1 ≥ 0
i = να i=να

The last inequality is valid because pi > 0, γi > βi−1 , and Cγi ≥ Cβi−1 .
Because Cβi−1 may have arbitrary but ordered values (Cβi−1 ≤ Cβi due to ordered βi ), i = να , . . . , ν,
the last inequality is valid for any ordered sequence of values zi−1 ≤ zi . Therefore, it is valid for VaRs
Vi that: ν ν
di Vi = (qi−1 − pi )Vi ≥ 0
i = να i = να

130
JRFM 2019, 12, 107

This leads to the inequality for the upper bound:

ν ν
U= qi−1 Vi ≥ pi Vi = Sα (X).
i = να i = να

Let us prove the inequality for the lower bound L:

ν ν ν ν
L= qi−1 Vi−1 = (pi + di )(Vi − ΔVi ) = Sα (X) − pi ΔVi + di Vi−1 (A13)
i = να i = να i = να i = να

where ΔVi = Vi − Vi−1 .

Let us calculate the upper estimate of L. Let us set Vνα−1 = Vνα , therefore ΔVνα = 0. This increases
the right hand side of Equation (A13). By recalling that pi = δ/(1 − α) = ν(1−α 1
)
, i = να + 1, . . . , ν,
for the Set 1, we have: ν
pi ΔVi = δ(Vν − Vνα )/(1 − α)
i=να
    
Because νi=να pi = νi=να qi−1 , then νi=να di = 0 and νi=να di Vi−1 = νi=να di (Vi−1 − D) for any D.
ν
Because i=να di zi ≥ 0 for any increasing ordered zi , and we can set zνα = 0, zi = 1, i =
να + 1, . . . , ν, then: ν ν
di zi = di = −dνα = pνα − qνα −1 ≥ 0
i = να i = να + 1

δ ν δ
Taking into account that pνα , qνα −1 ≥ 0 and pνα ≤ 1−α , we have i = να + 1 di ≤ 1−α . Then:

ν ν   ν     δ
di Vi−1 = di Vi−1 − Vνα−1 ≤ di Vν−1 − Vνα−1 ≤ Vν−1 − Vνα−1
i = να i = να i = να + 1 1−α
Let us return to estimation L by taking into account that Vνα−1 = Vνα :

ν ν δ(Vν −Vνα ) δ(Vν−1 −Vνα−1 )

L ≤ Sα (X) − i = να pi ΔVi + i = να di Vi−1 ≤ Sα (X) − 1−α + 1−α
δ(Vν −Vν−1 )
= Sα (X) − 1−α ≤ Sα (X)

Therefore, we have proved that L ≤ Sα (X) ≤ U.

Lemma 3 is proved. 

References
Acerbi, Carlo, and Dirk Tasche. 2002. On the Coherence of Expected Shortfall. Journal of Banking and Finance 26:
1487–503. [CrossRef]
Adrian, Tobias, and Markus K. Brunnermeier. 2016. CoVaR. American Economic Review 106: 1705–41. [CrossRef]
Basset, Gilbert W., and Hsiu-Lang Chen. 2001. Portfolio Style: Return-based Attribution Using Quantile Regression.
Empirical Economics 26: 293–305. [CrossRef]
Beraldi, Patrizia, Antonio Violi, Massimiliano Ferrara, Claudio Ciancio, and Bruno Antonio Pansera. 2019. Dealing
with complex transaction costs in portfolio management. Annals of Operations Research, 1–16. [CrossRef]
Carhart, Mark M. 1997. On Persistence in Mutual Fund Performance. Journal of Finance 52: 57–82. [CrossRef]
Case Study. 2014. Style Classification with Quantile Regression. Available online: [Link]
uryasev/research/testproblems/financial_engineering/style-classification-with-quantile-regression/ (accessed
on 24 June 2019).
Case Study. 2016. Estimation of CVaR through Explanatory Factors with CVaR (Superquantile)
Regression. Available online: [Link]
on-implementation-of-cvar-regression/ (accessed on 24 June 2019).
Fissler, Tobias, and Johanna F. Ziegel. 2015. Higher order elicitability and Osband’s principle. arXiv. [CrossRef]
Huang, Wei-Qiang, and Stan Uryasev. 2018. The CoCVaR Approach: Systemic Risk Contribution Measurement.
Journal of Risk 20: 75–93. [CrossRef]
Koenker, R. 2005. Quantile Regression. Cambridge: Cambridge University Press.
Koenker, Roger, and Gilbert Bassett. 1978. Regression Quantiles. Econometrica 46: 33–50. [CrossRef]

131
JRFM 2019, 12, 107

Portfolio Safeguard. 2018. American Optimal Decisions, USA. Available online: [Link] (accessed
on 24 June 2019).
Rockafellar, R. Tyrrell, and Johannes O. Royset. 2018. Superquantile/CVaR Risk Measures: Second-order Theory.
Annals of Operations Research 262: 3–29. [CrossRef]
Rockafellar, R. Tyrrell, and Stan Uryasev. 2000. Optimization of Conditional Value-At-Risk. Journal of Risk 2: 21–41.
[CrossRef]
Rockafellar, R. Tyrrell, and Stan Uryasev. 2002. Conditional Value-at-Risk for General Loss Distributions. Journal
of Banking and Finance 26: 1443–71. [CrossRef]
Rockafellar, R. Tyrrell, and Stan Uryasev. 2013. The Fundamental Risk Quadrangle in Risk Management,
Optimization and Statistical Estimation. Surveys in Operations Research and Management Science 18: 33–53.
[CrossRef]
Rockafellar, R. Tyrrell, Stan Uryasev, and Michael Zabarankin. 2008. Risk Tuning with Generalized Linear
Regression. Mathematics of Operations Research 33: 712–29. [CrossRef]
Rockafellar, R. Terry, Johannes O. Royset, and Sofia I. Miranda. 2014. Superquantile Regression with Applications
to Buffered Reliability, Uncertainty Quantification and Conditional Value-at-Risk. European Journal Operations
Research 234: 140–54. [CrossRef]
Sharpe, William F. 1992. Asset Allocation: Management Style and Performance Measurement. Journal of Portfolio
Management (Winter) 18: 7–19. [CrossRef]
Ziegel, Johanna F. 2014. Coherence and elicitability. arXiv. [CrossRef]

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

132
 Journal of
Risk and Financial
Management

Article
Determining Distribution for the Quotients of
Dependent and Independent Random Variables by
Using Copulas
Sel Ly 1 , Kim-Hung Pho 1, *, Sal Ly 1 and Wing-Keung Wong 2,3,4
1 Faculty of Mathematics and Statistics, Ton Duc Thang University, Ho Chi Minh City 756636, Vietnam;
lysel@[Link] (S.L.); shanlee5611@[Link] (S.L.)
2 Department of Finance, Fintech Center, and Big Data Research Center, Asia University, Taichung 41354,
Taiwan; wong@[Link]
3 Department of Medical Research, China Medical University Hospital, Taichung 40402, Taiwan
4 Department of Economics and Finance, Hang Seng University of Hong Kong, Shatin,
Hong Kong 999077, China
* Correspondence: phokimhung@[Link]

Received: 21 February 2019; Accepted: 8 March 2019; Published: 12 March 2019

Abstract: Determining distributions of the functions of random variables is a very important problem
with a wide range of applications in Risk Management, Finance, Economics, Science, and many other
areas. This paper develops the theory on both density and distribution functions for the quotient
Y= X X1
X2 and the ratio of one variable over the sum of two variables Z = X1 + X2 of two dependent or
1

independent random variables X1 and X2 by using copulas to capture the structures between X1 and
X2 . Thereafter, we extend the theory by establishing the density and distribution functions for the
quotients Y = X X1
X2 and Z = X1 + X2 of two dependent normal random variables X1 and X2 in the case
1

of Gaussian copulas. We then develop the theory on the median for the ratios of both Y and Z on
two normal random variables X1 and X2 . Furthermore, we extend the result of median for Z to a
larger family of symmetric distributions and symmetric copulas of X1 and X2 . Our results are the
foundation of any further study that relies on the density and cumulative probability functions of
ratios for two dependent or independent random variables. Since the densities and distributions
of the ratios of both Y and Z are in terms of integrals and are very complicated, their exact forms
cannot be obtained. To circumvent the difficulty, this paper introduces the Monte Carlo algorithm,
numerical analysis, and graphical approach to efficiently compute the complicated integrals and
study the behaviors of density and distribution. We illustrate our proposed approaches by using
a simulation study with ratios of normal random variables on several different copulas, including
Gaussian, Student-t, Clayton, Gumbel, Frank, and Joe Copulas. We find that copulas make big
impacts from different Copulas on behavior of distributions, especially on median, spread, scale and
skewness effects. In addition, we also discuss the behaviors via all copulas above with the same
Kendall’s coefficient. The approaches developed in this paper are flexible and have a wide range
of applications for both symmetric and non-symmetric distributions and also for both skewed and
non-skewed copulas with absolutely continuous random variables that could contain a negative
range, for instance, generalized skewed-t distribution and skewed-t Copulas. Thus, our findings are
useful for academics, practitioners, and policy makers.

Keywords: copulas; dependence structures; quotient of random variables; density functions;

distribution functions

JRFM 2019, 12, 42; doi:10.3390/jrfm12010042 133 [Link]/journal/jrfm

JRFM 2019, 12, 42

1. Introduction
Determining distributions of the functions of random variables is a very crucial task and this
problem has been attracted a number of researchers because there are numerous applications in Risk
Management, Finance, Economics, Science, and, many other areas, see, for example, (Donahue 1964;
Ly et al. 2016; Nadarajah and Espejo 2006; Springer 1979). Basically, the distributions of an algebraic
combination of random variables including the sum, product, and quotient are focused on some
common distributions along with the assumptions of independence or correlated through Pearson’s
coefficient or dependence via multivariate normal joint distributions (Arnold and Brockett 1992;
Bithas et al. 2007; Cedilnik et al. 2004; Hinkley 1969; Macalos and Arcede 2015; Marsaglia 1965;
Matović et al. 2013; Mekićet al. 2012; Nadarajah and Espejo 2006; Nadarajah and Kotz 2006a,
2006b; Pham-Gia et al. 2006; Pham-Gia 2000; Rathie et al. 2016; Sakamoto 1943). Regarding ratio,
it often appears in the problems of constructing statistics used in hypothesis testing and estimating
issues. Some well-known distributions are results of such quotients. For example, the quotient of
a Gaussian random variable divided by a square root of an independent chi-distributed random
variable follows the t-distribution while the F-distribution is derived via the ratio of two independent
chi-squared distributed random variables. To relax independence assumption, it is necessary to develop
a framework for modeling dependence structures of random vectors in more general sense. To do so,
Dolati et al. (2017) develop the distribution for X/Y in which both X and Y are positive.
In our paper, we first extend the theory developed by Dolati et al. (2017) to relax the positive
assumption for the variables by developing the theory on both density and distribution function (CDF)
for the quotient Y = X 1
X2 of two dependent or independent continuous random variables X1 and X2
in which X1 and X2 could be any real number. Thereafter, we develop a theory on both density and
distribution function for the ratio of one variable over the sum of two variables Z = X X+1X2 of two
1
dependent or independent continuous random variables X1 and X2 by using copulas to capture the
structures between X1 and X2 .
Since the density and the CDF formula of the ratios of both Y and Z are in terms of integrals and
are very complicated, we cannot obtain the exact forms of the densities and the CDFs. To circumvent the
difficulty, in this paper, we propose to use a Monte Carlo algorithm, numerical analysis, and graphical
approach to study behavior of density and distribution. We illustrate our proposed approaches by
using a simulation study with ratios of standard normal random variables on several different copulas,
including Gaussian, Student-t, Clayton, Gumbel, Frank, and Joe Copulas and we find that copulas
make big impacts from different Copulas on behavior of distributions, especially on median, spread,
skewness and scale effects. For instance, when X1 and X2 tend to be more co-monotonic indicated by
increasing the parameters of copulas, then the median of Y is shifted to be higher and its shape tends
to be more symmetric. In the meantime, the median of Z is equally unchanged one-half and the shape
always has symmetry. We note that the approaches developed in this paper are flexible and have a
wide range of applications for both symmetric and non-symmetric distributions and also for both
skewed and non-skewed copulas with absolutely continuous random variables that could contain a
negative range, for instance, generalized skewed-t distribution and skewed-t Copulas.1 Thus, our
findings are useful for academics, practitioners, and policy makers.
The rest of the paper is organized as follows. In Sections 2 and 3, we will briefly discuss the
background theory and copula theory related to the theory developed in our paper. In Section 4,
we provide main results on the quotients of dependent and independent random variables. Section 5
proposes using the Monte Carlo to deal with complex integrals and estimate some percentiles by using
some special copulas, and investigate their effects on the behavior of ratios of two standard normal
random variables. The last section provides the conclusions.

1 We would like to thank the anonymous reviewer for giving us helpful comments so that we could draw this conclusion.

134
JRFM 2019, 12, 42

2. Background Theory
We first review some previous work on the weighted sum, for example, in constructing portfolio
Y1 that is composed of two dependent assets defined by

Y1 = w1 X1 + w2 X2 , (1)

in which the random variables Xi,t (i = 1, 2) denotes the rate of return at time t for the asset defined in
terms of the following random quotient:

Pi,t − Pi,t−1
Xi,t = ,
Pi,t−1 ,

where Pi,t denotes the price of the ith asset at time t. Note that Xi is assumed to be absolutely continuous
with the cumulative distribution functions (CDF) Fi . Suppose that ( X1 , X2 ) follows copula C, then the
CDF, FY1 (y), of Y1 defined in (1) satisfies:

 1
* * ++
∂ y − w1 F1−1 (u)
FY1 (y) = 1{w2 <0} + sgn(w2 ) C u, F2 du, (2)
0 ∂u w2

where sgn(·) denotes the sign function such that

%
1, if x > 0,
sgn( x ) =
−1, if x < 0.

Then, the CDF, FY1 , can be used to estimate the distortion risk measure of the portfolio defined by
 ∞  0
R g [Y1 ] = g( FY1 (y))dy + [ g( FY1 (y)) − 1]dy,
0 −∞

where g is a distortion function and FY1 (y) = 1 − FY1 (y) is a survival function of Y1 . Readers may refer
to Ly et al. (2016) for more detailed information.
In the credit model, the total loss is defined as the aggregation of the product of risk factors. Thus,
it is necessary to find the distribution for the product case, for instance, Y2 given by

Y2 = X1 X2 . (3)

Ly et al. (2019) show that the CDF of Y2 can be determined by

 1
* * ++
  ∂ y
FY2 (y) = F1 (0) + sgn F1−1 (u) C u, F2 du. (4)
0 ∂u F1−1 (u)

3. Copulas
In this section, we will briefly discuss the copula theory related to the theory developed in our
paper. Readers may refer to (Cherubini et al. 2004; Joe 1997; Nelsen 2007; Tran et al. 2015, 2017) for
more information. Let I = [0, 1] be the closed unit interval and I2 = [0, 1] × [0, 1] be the closed unit
square interval. We first state the most basic definition of copula in two dimensions in the following:

Definition 1. (Copula) A 2-copula (two-dimensional copula) is a function C: I2 → I satisfying the

following conditions:

(i) C (u, 0) = C (0, v) = 0 for any u, v ∈ I;

(ii) C (u, 1) = u and C (1, v) = v for any u, v ∈ I; and

135
JRFM 2019, 12, 42

(iii) for any u1 , u2 , v1 , v2 ∈ I with u1 ≤ u2 and v1 ≤ v2 ,

C (u2 , v2 ) + C (u1 , v1 ) − C (u2 , v1 ) − C (u1 , v2 ) ≥ 0.

In copula theory, Sklar proposed a very important theorem in 1959 called Sklar’s Theorem
(Cherubini et al. (2004); Joe (1997); Nelsen (2007)), which plays the most important role in this theory.
It tells us that given a random vector ( X1 , X2 ) with absolutely continuous marginal distribution
functions FX1 and FX2 , respectively, and its joint distribution function denoted by H, and then there
exists a unique copula C such that
, -
H ( x1 , x2 ) = C FX1 ( x1 ), FX2 ( x2 ) ,
∂2 , -
h ( x1 , x2 ) = H ( x1 , x2 ) = c FX1 ( x1 ), FX2 ( x2 ) f X1 ( x1 ) f X2 ( x2 ) , (5)
∂x1 ∂x2

∂ 2
where c(u, v) := ∂u∂v C (u, v) denotes density of copula C, f Xi is probability density function (PDF) of
Xi , i = 1, 2, and h( x1 , x2 ) is the joint density function of X1 and X2 . Copula is used to combine several
univariate distributions together into bivariate [multivariate] settings so as the copula C can capture
the dependence structure of ( X1 , X2 ) [( X1 , · · · , Xn )]. For any copula C, we have the bounds

W (u, v) ≤ C (u, v) ≤ M(u, v),

where the copula W (u, v) := max(u + v − 1, 0) captures counter-monotonicity structure; that is,
X2 = f ( X1 ) a.s., where f is strictly decreasing, while the copula M (u, v) := min(u, v) is used to
capture comonotonicity; that is, X2 = f ( X1 ) a.s., where f is strictly increasing. In case X1 and X2
are independent, they follow copula denoted by Π(u, v) := uv. Copulas can be used not only to
model the dependence structure of the variables, but also capture the correlation between the variables.
The Kendall’s coefficient τ can be expressed in terms of copulas as shown in the following:
 
τ ( X1 , X2 ) = τ ( C ) = 4 C (u, v)dC (u, v) − 1. (6)
I2

In the next section, we will derive the two main propositions regarding formulas that can be
used to determine the probability density and probability distribution of the quotient of dependent
random variables by using copulas. In addition, we will apply the results to derive some corollaries on
PDFs, CDFs, and median of the ratios in case X1 and X2 are normal distributed and they follow the
Gaussian copulas.

4. Theory
We now develop two propositions on both density and distribution functions for the quotient
Y := X X1
X2 and the ratio of one variable over the sum of two variables Z : = X1 + X2 of two dependent
1

random variables X1 and X2 by using copulas. We first develop the proposition on the density and
distribution functions for the quotient Y = X 1
X2 as stated in the following:

Proposition 1. Supposing that ( X1 , X2 ) is a vector of two absolutely continuous random variables X1 and
X2 with the marginal distributions F1 and F2 , respectively, let C be an absolutely continuous copula modeling
dependence structure of the random vector ( X1 , X2 ), and define Y as

X1
Y := . (7)
X2

136
JRFM 2019, 12, 42

Then, the density f Y (y) and distribution functions FY (y) of Y are

 1       
fY (y) =  F −1 (v)c F1 yF −1 (v) , v f 1 yF −1 (v) dv, (8)
2 2 2
0
 1   ∂    
FY (y) = F2 (0) + sgn F2−1 (v) C F1 yF2−1 (v) , v dv, (9)
0 ∂v

respectively, where F2−1 denotes the inverse function of F2 , c is the density of copula C, and sgn(·) stands for a
sign function such that
%
1, if x > 0,
sgn( x ) =
−1, if x < 0.

Proof. Letting
⎧
⎨ Y : = X1 ,
1
X2
⎩
Y2 := X2 .

We note that since X2 is absolutely continuous, P( X2 = 0) = 0; that is, X2 = 0 almost surely.

Hence, the transformation Y1 = X 1
X2 always exists with probability 1 and we can obtain its inverse
transformation by using
%
X1 = Y1 Y2 ,
X2 = Y2 ,
and their corresponding Jacobian  
Y Y1 

J= 2  = Y2 .
0 1

Then, we obtain the joint density of Y1 and Y2 such that

h ( y1 , y2 ) = f ( y1 y2 , y2 ) | y2 |
= |y2 | c ( F1 (y1 y2 ), F2 (y2 )) f 1 (y1 y2 ) f 2 (y2 ) .

This yields the density of Y1 :

∞
f Y1 (y1 ) = |y2 | c ( F1 (y1 y2 ), F2 (y2 )) f 1 (y1 y2 ) f 2 (y2 ) dy2 (10)
−∞
 1       
 −1 
= F2 (v) c F1 y1 F2−1 (v) , v f 1 y1 F2−1 (v) dv. (11)
0

As a result, the CDF of Y1 is determined by

 1 t        
 −1 
FY1 (t) = F2 (v) c F1 y1 F2−1 (v) , v f 1 y1 F2−1 (v) dy1 dv. (12)
0 −∞
   
By changing variable, u = F1 y1 F2−1 (v) , we get du = F2−1 (v) f 1 y1 F2−1 (v) dy1 and we
note that
F2−1 (v) ≥ 0 ⇐⇒ v ∈ [0, F2 (0)], and F2−1 (v) ≤ 0 ⇐⇒ v ∈ [ F2 (0), 1].

137
JRFM 2019, 12, 42

This yields
 F2 (0)  F1 (tF−1 (v))  1  F2 (tF−1 (v))
2 ∂2 1 ∂2
FY1 (t) = − C (u, v) dudv + C (u, v) dvdu
0 1 ∂u∂v F2 (0) 0 ∂u∂v
 F2 (0)       1    
∂ ∂ ∂
= − C F1 tF2−1 (v) , v − C (1, v) dv + C F1 tF2−1 (v) , v dv (13)
0 ∂v ∂v F2 (0) ∂v
 1   ∂    
= F2 (0) + sgn F2−1 (v) C F1 tF2−1 (v) , v dv.
0 ∂v

Thus, the assertions of Proposition 1 hold.

From Proposition 1 and applying Equation (8), we obtain the following corollary on both density
and distribution functions for the quotient Y = X 1
X2 of two independent random variables X1 and X2
by using copulas:

Corollary 1. When X1 and X2 are independent, then its copula C (u, v) = uv has the density c(u, v) = 1,
∀u, v ∈ I and the density f Y (y) of the ratio Y := X 1
X2 of two independent random variables becomes
 ∞
fY (y) = | x | f 1 ( xy) f 2 ( x )dx.
−∞

This result is well known in the literature.

Next, we apply Equation (9) to derive both density and distribution function for Y = X 1
X2 in case
X1 and X2 are normal random variables and their dependence structure is captured by Gaussian
Copulas. We first obtain the following corollary:

Corollary 2. Assume that X1 ∼ N (μ1 , σ12 ), X2 ∼ N (μ2 , σ22 ) and ( X1 , X2 ) follows Gaussian Copulas
Cr (u, v), |r | < 1, given in (46). Then, the density f Y (y) and distribution function FY (y) of Y = X 1
X2 have
the forms
 1    (yσ − rσ )Φ−1 (v) + yμ − μ  σ Φ−1 (v) + μ
sgn σ2 Φ−1 (v) + μ2 ϕ
2 1 2 1 2 2
fY (y) = √ √ dv, (14)
0 σ1 1 − r2 σ1 1 − r2
 μ   1    (yσ − rσ )Φ−1 (v) + yμ − μ 
sgn σ2 Φ−1 (v) + μ2 Φ
2 2 1 2 1
FY (y) = Φ − + √ dv, (15)
σ2 0 σ1 1 − r2

respectively, where ϕ( x ) and Φ( x ) are PDF and CDF of the standard normal distribution, respectively,
and Φ−1 ( x ) denotes for the inverse function of Φ( x ).

Proof. Let X1 ∼ N (μ1 , σ12 ), and X2 ∼ N (μ2 , σ22 ); then, their CDFs and inverse functions can be
expressed in the following form:
x − μ 
Fi ( x ) = Φ i
, Fi−1 (v) = σi Φ−1 (v) + μi , i = 1, 2.
σi
∂Cr
Given Gaussian Copulas Cr (u, v) with |r | < 1, one can obtain its derivative ∂v ( u, v ),
see Meyer (2013), as shown in the following:

∂Cr  Φ−1 (u) − rΦ−1 (v) 

(u, v) = Φ √ .
∂v 1 − r2

138
JRFM 2019, 12, 42

Now, applying Equation (9), we can simplify it to be

  −1

 μ2 
 1    Φ−1 Φ yσ2 Φ (vσ)+yμ2 −μ1 − rΦ−1 (v) 
sgn σ2 Φ−1 (v) + μ2 Φ
1
FY (y) = Φ − + √
σ2 0 1 − r2
 μ   1    (yσ − rσ )Φ−1 (v) + yμ − μ 
sgn σ2 Φ−1 (v) + μ2 Φ
2 2 1 2 1
= Φ − + √ dv.
σ2 0 σ1 1 − r2

Taking derivative of Fy (y) with respect to y, one gets the density f Y (y) defined as in (14).
The assertions of Corollary 2 hold.

We note that the probability exhibited in (15) can be easily computed by using the following
Monte Carlo algorithm: For each y ∈ R, we generate V from the uniform distribution on the unit
interval [0, 1] with sample size N, say N = 10,000, and then the estimated probability is given by

 μ  1 N    (yσ − rσ )Φ−1 (v ) + yμ − μ 
F.y (y) ≈ Φ − ∑ sgn σ2 Φ−1 (vi ) + μ2 Φ
2 2 1 2 1
+ √ i . (16)
σ2 N i =1 σ1 1 − r2

Using the result, we obtain the following corollary:

Corollary 3. If X1 ∼ N (μ1 , σ12 ), X2 ∼ N (0, σ22 ), and ( X1 , X2 ) follows Gaussian Copulas Cr (u, v), |r | < 1,
given in (46), then the median of Y = X 1
X2 satisfies

σ1
median(Y ) = r , for all μ1 ∈ R. (17)
σ2
μ
Proof. Since X2 ∼ N (0, σ22 ), Φ(− σ22 ) = Φ(0) = 0.5. Hence, it is sufficient to prove that the integral
term given in (15) is equal to zero. In fact, we find that
 σ   1  
−μ
FY r 1 = 0.5 + sgn(σ2 Φ−1 (v))Φ √ 1 dv
σ2 0 σ1 1 − r2
   1/2  1 
−μ
= 0.5 + Φ √ 1 − dv + dv
σ1 1 − r2 0 1/2
= 0.5.

Hence, the quantity r σσ12 is the median of Y. The proof is complete.

We turn to develop the proposition on density and distribution functions for the ratio of one
variable over the sum of two variables Z := X X+1X2 of two dependent random variables X1 and X2 by
1
using copulas as stated in the following:

Proposition 2. Suppose that ( X1 , X2 ) is a vector of two absolutely continuous random variables X1 and X2
with the marginal distributions F1 and F2 , respectively, and let C be an absolutely continuous copula modeling
dependence structure of the random vector ( X1 , X2 ), and define Z as

X1
Z := . (18)
X1 + X2

Then, the density f Z (z) and distribution function FZ (z) of Z are

⎧     
1 | F1 (u)|
−1
⎪ 1− z −1 1− z −1
⎪
⎨ 0 z2
c u, F2 z F1 ( u ) f2 z F1 ( u ) du, if z = 0,
f Z (z) = (19)
⎪
⎪
⎩ f (0)  −1   
1
1  F (v)c F1 (0), v dv, if z = 0,
0 2

139
JRFM 2019, 12, 42

       
F1−1 (u) ∂
u, F2 − F1−1 (u) − ∂ 1− z −1
1
FZ (z) = 1{z≥0} + 0 sgn ∂u C ∂u C u, F2 z F1 ( u ) du, (20)

respectively, where 1{·} denotes an indicator function, Fi−1 denotes the inverse function of Fi for i = 1, 2, c is the
density of copula C, and sgn(·) is the sign function such that
% %
1, if z ≥ 0, 1, if x > 0,
1 { z ≥0} = and sgn( x ) =
0, if z < 0, −1, if x < 0.

Proof. By defining
⎧
⎨Z = X1
1 ,
X1 + X2
⎩
Z2 = X1 + X2 .

Here, we note that, since X1 and X2 are absolutely continuous, P( X1 + X2 = 0) = 0; that is,
X1 + X2 = 0 almost surely. Hence, the transformation Z1 = X X+1X2 always exists with probability 1
1
and we obtain the following inverse transformation:
%
X1 = Z1 Z2 ,
X2 = Z2 − Z1 Z2 ,

and the Jacobian  

 Z Z1 

J= 2  = Z2 .
− Z2 1 − Z1 

Thus, the joint density of Z1 and Z2 becomes

h ( z1 , z2 ) = f ( z1 z2 , z2 − z1 z2 ) | z2 |
= |z2 | c ( F1 (z1 z2 ), F2 (z2 − z1 z2 )) f 1 (z1 z2 ) f 2 (z2 − z1 z2 ) ,

which leads us to get the density of Z1 such that

∞
f Z1 (z1 ) = |z2 | c ( F1 (z1 z2 ), F2 (z2 − z1 z2 )) f 1 (z1 z2 ) f 2 (z2 − z1 z2 ) dz2 . (21)
−∞

If z1 = 0, then by taking v := F2 (z2 ), we get

 1   
 −1 
f Z1 (0) = f 1 (0) F2 (v) c F1 (0), v dv.
0

If z1 > 0, then by taking u := F1 (z1 z2 ), we obtain

 
 1 F −1 (u)     
1 1 − z1 −1 1 − z1 −1
f Z1 (z1 ) = c u, F2 F1 (u) f2 F1 (u) du.
0 z21 z1 z1

If z1 < 0, then also by taking u := F1 (z1 z2 ), we yield

 
 0 F −1 (u)     
1 1 − z1 −1 1 − z1 −1
f Z1 (z1 ) = c u, F2 F1 (u) f2 F1 (u) du. (22)
1 −z21 z1 z1

140
JRFM 2019, 12, 42

Hence, for z1 = 0, we obtain

 
 1 F −1 (u)     
1 1 − z1 −1 1 − z1 −1
f Z1 (z1 ) = c u, F2 F1 ( u ) f 2 F1 ( u ) du for z1 = 0.
0 z21 z1 z1

As a consequence, the distribution of Z1 becomes

 
 1 t  −1     
F1 (u)  1 − z1 −1 1 − z1 −1
FY1 (t) = 2
c u, F2 F1 (u) f2 F1 (u) dz1 du. (23)
0 −∞ z1 z1 z1
   
1− z1 −1 F1−1 (u) 1− z1 −1
Setting v = F2 z1 F1 (u) =⇒ dv = − z21
f2 z1 F1 (u) dz1 , and note that

F1−1 (u) ≥ 0 ⇐⇒ u ≥ F1 (0), and F1−1 (u) ≤ 0 ⇐⇒ u ≤ F1 (0),

we consider two cases as follows:

(i) Case 1: For t < 0, we have

F1 (0) F2 ( t F1 (u)) ∂2 F2 ( 1−
t F1 ( u ) ) ∂2
1− t −1 t −1
1
FZ1 (t) = −1 ∂u∂v C ( u, v ) dvdu − F1 (0) F2 (− F −1 (u)) ∂u∂v C ( u, v ) dvdu
 F2 (− F1 (u)) 
0
   1 
F1 (0) ∂ 1− t −1 ∂ −1
= ∂u C u, F2 F (u) − ∂u C u, F2 − F1 (u) du
0
  t 1     (24)
∂
C u, F2 1−t t F1−1 (u) − ∂u ∂
C u, F2 − F1−1 (u)
1
− F (0) ∂u du
1        
1 −1 ∂ −1 ∂ 1− t −1
= 0 sgn F1 ( u ) ∂u C u, F2 − F1 ( u ) − ∂u C u, F2 t F1 (u) du.

(ii) Case 2: For t ≥ 0, we first split the integrals

 
 1  0 F −1 (u)     
1 1 − z1 −1 1 − z1 −1
FZ1 (t) = c u, F2 F1 (u) f2 F1 (u) dz1 du
0 −∞ z2 z1 z1
 1 
 1 t  −1     
F1 (u)  1 − z1 −1 1 − z1 −1
+ 2
c u, F2 F1 (u) f2 F1 (u) dz1 du (25)
0 0 z1 z1 z1
=: I1 + I2 .

We then apply (24) to obtain the following expression for the integral I1 :
 
 
   
F1−1 (u) u, F2 − F1−1 (u) −
∂ ∂ 1− t −1
1
I1 = FZ1 (0) = 0 sgn ∂u C ∂u C u, lim F2 t F1 ( u ) du
     t → 0−
1 −1 ∂ −1
= 0 sgn F1 ( u ) ∂u C u, F2 − F1 ( u ) du
(26)
F1 (0) ∂ 1 ∂
+ 0  ∂u C ( u,1) du  − F1 (0) ∂u C (u, 0) du
 
1 −1 ∂ −1
= 0 sgn F1 ( u ) ∂u C u, F2 − F1 (u) du + F1 (0),

and obtain the following expression for the integral I2 :

 F (0)  F2 ( 1−t F−1 (u))  1  F2 ( 1−t F−1 (u))
1 t 1 ∂2 t 1 ∂2
I2 = C (u, v) dvdu − C (u, v) dvdu
0 0 ∂u∂v F1 (0) 1 ∂u∂v
 F (0)     
1 ∂ 1 − t −1 ∂
= C u, F2 F1 (u) − C u, 0) du
0 ∂u t ∂u
 1    
∂ 1 − t −1 ∂  
− C u, F2 F1 (u) − C u, 1 du
F1 (0) ∂u t ∂u
 1   ∂   
−1
= 1 − F1 (0) − sgn F1 (u) C u, F2 − F1−1 (u) du. (27)
0 ∂u

141
JRFM 2019, 12, 42

From (26) and (27), we get the following for t ≥ 0,

 1   ∂    ∂
 
1 − t −1

FZ1 (t) = 1 + sgn F1−1 (u) C u, F2 − F1−1 (u) − C u, F2 F1 (u) du. (28)
0 ∂u ∂u t

Combining (24) and (28) imply (20), we complete the proof.

In the situation X1 and X2 are independent, applying Proposition 2, we obtain the

following corollary:

Corollary 4. When X1 and X2 are independent, then its copula C (u, v) = uv has the density c(u, v) = 1,
∀u, v ∈ I and the density f Z (z) and distribution function FZ (z) for the ratio of one variable over the sum of two
variables Z := X X+1X2 of two independent random variables X1 and X2 become
1

∞
f Z (z) = | x | f 1 ( xz) f 2 ((1 − x )z) dx
−∞

and
 1     1 − z 
FZ (z) = 1 { z ≥0} + sgn F1−1 (u) F2 − F1−1 (u) − F2 F1−1 (u) du,
0 z
 ∞  1 − z 
= 1 { z ≥0} + sgn( x ) F2 (− x ) − F2 x f 1 ( x )dx,
−∞ z

respectively.

Next, we apply Equation (20) to derive the distribution function of Z := X X+1X2 in the situation
1
that both X1 and X2 are normal distributed such that their dependence structure can be captured by
Gaussian Copulas as shown in the following corollary:

Corollary 5. Assume that X1 ∼ N (μ1 , σ12 ), X2 ∼ N (μ2 , σ22 ), and ( X1 , X2 ) follows Gaussian Copulas
Cr (u, v), |r | < 1, given in (46). Then, distribution function FZ (z) of Z := X X+1X2 has the form
1

μ   1   (σ + rσ )Φ−1 (u) + μ − μ 
sgn σ1 Φ−1 (u) + μ1 Φ
1 1 2 1 2
FZ (z) = = 1{z≥0} + 2Φ −1− √ du
σ1 0 σ2 1 − r2
 1    [(1 − z)σ − zrσ ]Φ−1 (u) − z(μ + μ ) + μ 
sgn σ1 Φ−1 (u) + μ1 Φ 1 2 1 2 1
− √ du, (29)
0 zσ2 1 − r2

where Φ( x ) and Φ−1 ( x ) are CDF and its inverse of the standard normal random variable, respectively.

Proof. Let X1 ∼ N (μ1 , σ12 ) and X2 ∼ N (μ2 , σ22 ), the CDFs and their inverse functions can be written
in the form x − μ 
Fi ( x ) = Φ i
, Fi−1 (v) = σi Φ−1 (v) + μi , i = 1, 2.
σi
Given Gaussian Copulas Cr (u, v), |r | < 1, we apply the results from Meyer (2013) to obtain its
derivative ∂C
∂u ( u, v ) as shown in the following:
r

∂Cr  Φ−1 (v) − rΦ−1 (u) 

(u, v) = Φ √ .
∂u 1 − r2

142
JRFM 2019, 12, 42

Applying Equation (20), one can simplify it to be

  −1

 1    Φ−1 Φ − σ1 Φ (uσ)+μ1 −μ2 − rΦ−1 (u) 
sgn σ1 Φ−1 (u) + μ1 Φ
2
FZ (z) = 1 { z ≥0} + √ du
0 1 − r2
! "
  1−z σ Φ−1 (u)+μ −μ 
 1   Φ −1 Φ z 1 σ − rΦ−1 (u) 
2
 1

sgn σ1 Φ−1 (u) + μ1 Φ

2
− √ du
0 1 − r2
 1    (σ + rσ )Φ−1 (u) + μ − μ 
sgn σ1 Φ−1 (u) + μ1 Φ − 1
2 1 2
= 1 { z ≥0} + √ du
0 σ2 1 − r2
 1    [(1 − z)σ − zrσ ]Φ−1 (u) − z(μ + μ ) + μ 
sgn σ1 Φ−1 (u) + μ1 Φ 1 2 1 2 1
− √ du
0 zσ2 1 − r2
μ   1    (σ + rσ )Φ−1 (u) + μ − μ 
sgn σ1 Φ−1 (u) + μ1 Φ
2 2
= 1{z≥0} + 2Φ 1 − 1 − 1
√ 1
du
σ1 0 σ2 1 − r2
 1    [(1 − z)σ − zrσ ]Φ−1 (u) − z(μ + μ ) + μ 
sgn σ1 Φ−1 (u) + μ1 Φ 1 2 1 2 1
− √ du.
0 zσ2 1 − r2

In the last step of the above, we use the property Φ(− x ) = 1 − Φ( x ) and
 
 1    Φ − μ1  1    
sgn σ1 Φ −1
(u) + μ1 du = −
σ 1
du +   du = 1 − 2Φ − μ1 = 2Φ μ1 − 1.
0 0 Φ
μ
− σ1 σ1 σ1
1

The proof is complete.

We note that the probability given in (29) can also be easily computed by using the following
Monte Carlo algorithm: For each z ∈ R, we first generate U from the uniform distribution on the unit
interval [0, 1] with sample size N, say N = 10, 000. Then, we obtain the following estimated probability:

μ  1 N   (σ + rσ )Φ−1 (u ) + μ − μ 

F.Z (z) σ1 Φ−1 (ui ) + μ1 Φ
∑ sgn
1 1 2 1 2
≈ 1{z≥0} + 2Φ −1− √ i
σ1 i =1
N σ2 1 − r2
1 N    [(1 − z)σ − zrσ ]Φ−1 (u ) − z(μ + μ ) + μ 
− ∑ sgn σ1 Φ−1 (ui ) + μ1 Φ 1 2
√ i 1 2 1
. (30)
N i =1 zσ2 1 − r2

Using the above results, we obtain the following corollary:

Corollary 6. Assume that X1 ∼ N (0, σ2 ), X2 ∼ N (0, σ2 ) and ( X1 , X2 ) follows Gaussian Copulas Cr (u, v),
|r | < 1, given in (46). Then, the median of Z := X X+1X2 is equal to 12 .
1

   
Proof. Because X1 ∼ N (0, σ2 ), X2 ∼ N (0, σ2 ) and sgn σΦ−1 (u) = sgn Φ−1 (u) , we obtain CDF of
X1
the ratio Z := X1 + X2 from (29) as shown in the following:

   1    (1 + r )Φ−1 (u)   [1 − z − zr ]Φ−1 (u) 

FZ z = 1{z≥0} − sgn Φ−1 (u) Φ √ +Φ √ du. (31)
0 1 − r2 z 1 − r2

We get
1  1    (1 + r )Φ−1 (u)   [1 − r ]Φ−1 (u) 
FZ = 1− sgn Φ−1 (u) Φ √ +Φ √ du. (32)
2 0 1 − r2 1 − r2

143
JRFM 2019, 12, 42

Hence, it is sufficient to prove that the integral term given in (32) is equal to 12 . We let
 1    (1 + r ) Φ −1 ( u )   1    (1 − r ) Φ −1 ( u ) 
I1 := sgn Φ−1 (u) Φ √ du, I2 := sgn Φ−1 (u) Φ √ du,
0 1 − r2 0 1 − r2

and denote ∂i Cr (u, v), i = 1, 2 to be the partial derivative of Cr (u, v) with respect to the ith variable,
∂ ∂
That is, ∂1 Cr (u, v) = ∂u Cr (u, v) and ∂2 Cr (u, v) = ∂v Cr (u, v). One can observe that

 (1 − r ) Φ −1 ( u ) 
∂1 Cr (u, u) = Φ √ .
1 − r2

Since Gaussian Copulas is symmetric; that is, Cr (u, v) = Cr (v, u), we have ∂1 Cr (u, v) = ∂2 Cr (v, u),
and, thus, for u = v, we can derive ∂1 Cr (u, u) = ∂2 Cr (u, u). Thereafter, the differentiation of Cr (u, u)
can be obtained:
! "
dCr (u, u) = ∂1 Cr (u, u) + ∂2 Cr (u, u) du = 2∂1 Cr (u, u),

and we get
 1    
1 1 1/2 1 1
I2 = sgn Φ−1 (u) dCr (u, u) = − dCr (u, u) + dCr (u, u)
2 0 2 0 2 1/2
1  1 1  1   1 1 
= − Cr , + 1 − Cr , (33)
2 2 2 2 2 2
1 1 1
= − Cr ( , ).
2 2 2

Similarly, for I1 , since Φ−1 (1 − u) = −Φ−1 (u), we obtain

 Φ−1 (1 − u) − rΦ−1 (u)   (1 + r ) Φ −1 ( u )   (1 + r ) Φ −1 ( u ) 
∂1 Cr (u, 1 − u) = Φ √ =Φ − √ = 1−Φ √ ,
1−r 2 1−r 2 1 − r2

and get
 1    1    1  
I1 = sgn Φ−1 (u) du − sgn Φ−1 (u) ∂1 Cr (u, 1 − u)du = − sgn Φ−1 (u) ∂1 Cr (u, 1 − u)du,
0 0 0
 
sgn Φ−1 (u) du = 0. From symmetry of the Gaussian Copulas, we also have
1
in which we apply 0
∂1 Cr (1 − u, u) = ∂2 Cr (u, 1 − u), obtain the differentiation of Cr (u, 1 − u) given by
! " ! "
dCr (u, 1 − u) = ∂1 Cr (u, 1 − u) − ∂2 Cr (u, 1 − u) du = ∂1 Cr (u, 1 − u) − ∂1 Cr (1 − u, u) du,

and get
 1/2  1
I1 = ∂1 Cr (u, 1 − u)du − ∂1 Cr (u, 1 − u)du,
0 1/2
 1/2  0
= ∂1 Cr (u, 1 − u)du + ∂1 Cr (1 − u, u)du
0 1/2
 1/2  1/2
= ∂1 Cr (u, 1 − u)du − ∂1 Cr (1 − u, u)du
0 0
 1/2
= dCr (u, 1 − u)
0
1 1
= Cr , . (34)
2 2

144
JRFM 2019, 12, 42

 
Combining (32)–(34), we find that FZ 1
2 = 12 . Hence, the quantity 12 is the median of Z. The proof
is complete.

Applying the proof of Corollary 6, we extend the result to obtain the following corollary for a
larger family of symmetric distribution and symmetric copulas:

Corollary 7. Assume that X1 and X2 are identically and symmetrically distributed with distribution F that
has zero median and the dependence structure of ( X1 , X2 ) is modelled by a family of symmetric copulas C (u, v),
i.e., C (u, v) = C (v, u) for all u, v ∈ I. Then, the median of Z := X X+1X2 is equal to 12 .
1

Proof. Since X1 and X2 are identically and symmetrically distributed with distribution F and zero
median, we have F (− x ) = 1 − F ( x ), for all x ∈ R. By applying Equation (20), we obtain the CDF of
the ratio Z, which is defined by
 1   ∂    ∂
 
1 − z −1

FZ (z) = 1 { z ≥0} + sgn F −1 (u) C u, F − F −1 (u) − C u, F F (u) du,
0 ∂u ∂u z
 1  
 ∂     
∂ 1 − z −1
= 1 { z ≥0} + sgn F −1 (u) C u, 1 − u − C u, F F (u) du. (35)
0 ∂u ∂u z

Since copulas C (u, v) are symmetric, i.e., C (v, u) = C (u, v), we have ∂1 C (v, u) = ∂2 C (u, v), and,
thus, for v = 1 − u, one can easily obtain ∂1 C (1 − u, u) = ∂2 C (u, 1 − u) and find that the differentiation
of C (u, 1 − u) with respect to u satisfies
! " ! "
dC (u, 1 − u) = ∂1 C (u, 1 − u) − ∂2 C (u, 1 − u) du = ∂1 C (u, 1 − u) − ∂1 C (1 − u, u) du.

In addition, because the distribution F has zero median; that is, F (0) = 0.5, we have
 1   ∂    1/2  1
sgn F −1 (u) C u, 1 − u = − ∂1 C (u, 1 − u)du + ∂1 C (u, 1 − u)du,
0 ∂u 0 1/2
 1/2  0
= − ∂1 C (u, 1 − u)du − ∂1 C (1 − u, u)du
0 1/2
 1/2  1/2
= − ∂1 C (u, 1 − u)du + ∂1 C (1 − u, u)du
0 0
 1/2 1 1
= − dC (u, 1 − u) = −C , .
0 2 2

Therefore, we get
1 1  1   ∂  
1 − z −1

FZ (z) = 1 { z ≥0} − C , − sgn F −1 (u) C u, F F (u) du, (36)
2 2 0 ∂u z

and thus,
1 1  1   ∂
FZ (0.5) = 1−C , − sgn F −1 (u) C (u, u) du. (37)
2 2 0 ∂u

Similarly, since copulas C (u, v) are symmetric, i.e., C (v, u) = C (u, v), we have ∂1 C (v, u) =
∂2 C (u, v), and, thus, for v = u, we get ∂1 C (u, u) = ∂2 C (u, u). Thus, the differentiation of C (u, u) with
respect to u satisfies
! "
dC (u, u) = ∂1 C (u, u) + ∂2 C (u, u) du = 2∂1 C (u, u)du.

145
JRFM 2019, 12, 42

Applying this relation, we find

 1   ∂    1 1
1 1 1
sgn F −1 (u) C (u, u) du = sgn F −1 (u) dC (u, u) = − C , .
0 ∂u 2 0 2 2 2

Hence, FZ (0.5) = 0.5, i.e., 1

2 is median of Z. The proof is complete.

Remark: In the literature, they are many symmetric distributions with zero median, for
example, normal N (0, σ), Student-t tν , Cauchy distribution with location parameter α = 0, uniform
U (− a, a), a ∈ R+ , and logistic distribution with zero location. In addition, Elliptical copulas (Gaussian,
Student-t copulas) and Archimedian copulas (Clayton, Gumbel, Frank, Joe,...) are classes of symmetric
copulas. Thus, if we apply ( X1 , X2 ) with these distributions and these copulas, Proposition 7 tells
us that the random variable Z := X X+1X2 always gets the median one-half. This theoretical result is
1
consistent with our simulation results displayed in the next section.
We note that the formulas given in (8), (9), (19) and (20) may not have closed forms. However,
they are easily computed by using the Monte Carlo (MC) simulation method or any techniques of
numerical integration. We provide simulation studies in the next section.

5. A Simulation Study
Since the density and the CDF formula of the ratio Y = X X1
X2 [Z = X1 + X2 ] expressed in (8) and (9)
1

((19) and (20)) are in terms of integrals and are very complicated, we cannot obtain the exact forms of
the density and the CDF. To circumvent the difficulty, in this paper, we propose to use the Monte Carlo
algorithm, numerical analysis and graphical approach to study behavior of density and distribution
and the changes of their shapes when parameters are changing.
Suppose that X1 and X2 are normally distributed and denoted by Xi ∼ N (μi , σi2 ) for i = 1, 2 with
PDF given by
1 ( x − μ i )2
f Xi ( x ) = √ exp(− ).
2πσi 2σi2
Without loss of generality, we consider μ1 = μ2 = 0 and σ1 = σ2 = 1. We note that, if X1 and X2
X1
are independent and standard normal distributed, then it is well known that Y = follows standard
X2
Cauchy distribution. The Cauchy distribution is a type of distribution that has no mean and also does
not exist any higher moments. To circumvent this problem, one may use median to measure the central
tendency and use range or interquartile range to measure the spread of the distribution. The general
Cauchy distribution has the following PDF:

β
f ( x ) := , α ∈ R, β > 0.
π ( β2 + ( x − α )2 )

X1
Thus, location and scale parameter of Y = are α = 0 and β = 1, respectively, if X1 and X2 are
X2
identically independent standard normal distributed.
We now investigate different dependence structures of X1 and X2 through several families of
copulas and observe the shapes of the corresponding distributions for both Y and Z as well as estimate
their percentiles at different levels including 0.05, 0.25, 0.5, 0.75, 0.95 and denote the corresponding
percentiles to be Q0.05 , Q0.25 , Q0.50 , Q0.75 , and Q0.95 , respectively. In risk analysis, the percentile Q0.05 is
often used as the Value-at-Risk (VaR) 5% while Q0.25 and Q0.75 are, respectively, called the first and
third quartile of the random variable and the interquartile range (IQR) is defined by their difference;
that is, IQR = Q0.75 − Q0.25 . The median measures the center of the distribution, which is equal
to Q0.5 .
For each copula Cθ (u, v), the PDF and CDF of Y and Z can be plotted on the interval [−4, 4] by
using the following steps:

146
JRFM 2019, 12, 42

(i) For each pair of y and z ∈ {−4, −3.9, −3.8, . . . , 4}, generate the uniform random variables U and
V on the unit interval; that is, U, V ∼ Uni f orm[0, 1] with the sample size N, say N = 10,000;
(ii) estimate the values for f Y (y), FY (y), f Z (z), and FZ (z) by using

1 N        
f.Y (y) ≈ ∑  F2−1 (vi )c F1 yF2−1 (vi ) , vi f 1 yF2−1 (vi ) , (38)
N i =1

1 N   ∂    
F.Y (y) ≈ F2 (0) + ∑ sgn F2−1 (vi ) C F1 yF2−1 (vi ) , vi , (39)
N i =1
∂v
⎧     
⎪ N | F1−1 (ui )|
⎪
⎪
⎪
1
N ∑ z2
c ui , F2 1−z z F1−1 (ui ) f 2 1−z z F1−1 (ui ) , if z = 0,
⎨ i =1
f.Z (z) ≈ (40)
⎪
⎪    
⎪
⎪ f 1 (0) N  −1 
⎩ N ∑ F2 ( vi )  c F1 (0), vi , if z = 0,
i =1

1 N   ! ∂C , , -- ∂C , , 1 − z −1 --"
F.Z (z) ≈ 1{z≥0} + sgn F1−1 (ui ) ui , F2 − F1−1 (ui ) −
N i∑
ui , F2 F1 (ui ) , (41)
=1
∂u ∂u z

∂ ∂
in which the density copula cθ (ui , vi ), the derivatives ∂u Cθ (ui , vi ) and ∂v Cθ ( ui , vi ) can be obtained
by using the packages of VineCopula in R language; and
(iii) plot f.Y (y), F.Y (y), f.Z (z), and F.Z (z) with y, z ∈ {−4, −3.9, −3.8, · · · , 3.9, 4}.

To estimate percentiles Qα ’s of Y and Z, we first construct the joint distribution of ( X1 , X2 ) by

using Sklar’s Theorem as shown in the following: For each copula Cθ (u, v), we first obtain the joint
CDF of ( X1 , X2 ) such that  
Hθ ( x1 , x2 ) = Cθ F1 ( x1 ), F2 ( x2 ) .

We then repeat 5000 times, k = 1, 2, . . . , 5000 for the following steps in the computation:

(1) For each repetition k = 1, 2, . . . , 5000,

(i) generate ( X1 , X2 ) from Hθ ( x1 , x2 ) of sample size 104 by using the package copula in R
language and define

(k)
(k) x1i
yi = (k)
, i = 1, 2, · · · , 104 ; (42)
x2i
(k)
(k) x1i
zi = (k) (k)
, i = 1, 2, · · · , 104 ; (43)
x1i + x2i

(ii) estimate the percentiles Qα with α = 0.05, 0.25, 0.5, 0.75, 0.95 for both Y and Z by using the
following formula
   
. α (Y ( k ) ) (k) (k) (k)
Q = y(h) + (h − h) y(h+1) − y(h) , with h = 104 − 1 α + 1, (44)
 
. α ( Z (k) ) = z(k) + (h −  h) z(k)
Q − z
(k)
, (45)
( h ) ( h + 1 ) ( h )

(k) (k) (k) (k) (k) (k)

where y(1) ≤ y(2) ≤ . . . ≤ y( N ) and z(1) ≤ z(2) ≤ . . . ≤ z( N ) denote the order statistics of
both y(k) and z(k) , respectively, and  h denotes integer part of h.

147
JRFM 2019, 12, 42

(2) Finally, we take average for each of the above quantities by using the following formula:

. α (Y ) 1 5000 .
Q α (Y ( k ) ) ,
5000 k∑
Q =
=1

.α (Z) 1 5000 .
Q α ( Z ( k ) ),
5000 k∑
Q =
=1

to obtain the estimates of the percentiles for Y and Z.

We note that the algorithms discussed in the above can be applied to any non-symmetric marginal
distribution and skewed copulas family with absolutely continuous random variable that could contain
negative range, for example, generalized skewed-t distribution and skewed-t Copulas. 2

5.1. Gaussian Copulas

We first investigate dependence structures of X1 and X2 through Gaussian Copulas Cr (u, v) and
observe the shapes of the corresponding distributions for both Y and Z.
 Φ −1 ( u )  Φ −1 ( v )  2 
1 s − 2rst + t2
Cr (u, v) = √ exp − dsdt, (46)
2π 1 − r2 −∞ −∞ 2(1 − r 2 )

where Φ−1 ( x ) is the inverse of standard normal CDF and r is Pearson’s correlation coefficient between
X1 and X2 , |r | < 1. We now consider the cases with r = −0.9, −0.5, 0, 0.5, 0.9. When r = 0, it is
corresponding to the independence situation, and we get PDFs and CDFs of Y and Z shown in
Figures 1 and 2, respectively. As can be seen from the Figures and Tables 1 and 2, when the parameter
r varies from negative to positive, the median is totally equal to the Pearson’s correlation coefficient r.
The more correlated, that is, the higher |r | between X1 and X2 is, the smaller the spread, that is, IQR of
Y, becomes. In contrast to Y, the center of Z is definitely unchanged (0.5), but the scale parameter of Z
is smaller indicated by the higher height of the density. The shapes of both Y and Z are symmetric.

Table 1. Some percentiles of Y = X1 /X2 , where ( X1 , X2 ) follows Gaussian Copulas.

r τ (Cr ) Q0.05 Q0.25 Median Q0.75 Q0.95

−0.9 −0.71 −3.65 −1.34 −0.9 −0.46 1.85
−0.5 −0.33 −5.97 −1.37 −0.5 0.37 4.97
0 0 −6.31 −1.00 0.0 1.00 6.31
0.5 0.33 −4.97 −0.37 0.5 1.37 5.97
0.9 0.71 −1.85 0.46 0.9 1.34 3.65

Table 2. Some percentiles of Z = X1 /( X1 + X2 ), where ( X1 , X2 ) follows Gaussian Copulas.

r τ (Cr ) Q0.05 Q0.25 Median Q0.75 Q0.95

−0.9 −0.71 −13.27 −1.68 0.5 2.68 14.26
−0.5 −0.33 −4.97 −0.37 0.5 1.37 5.97
0 0 −2.66 0.00 0.5 1.00 3.65
0.5 0.33 −1.33 0.21 0.5 0.79 2.32
0.9 0.71 −0.23 0.39 0.5 0.61 1.22

2 We would like to show our appreciation to the anonymous reviewer to giving us helpful comments so that we could draw
this conclusion.

148
JRFM 2019, 12, 42

X1
Figure 1. PDFs and CDFs of the ratio Y = X2 , where ( X1 , X2 ) follows Gaussian Copulas.

X1
Figure 2. PDFs and CDFs of the ratio Z = X1 + X2 , where ( X1 , X2 ) follows Gaussian Copulas.

5.2. Student-t Copulas

We then investigate dependence structures of X1 and X2 through Student-t Copulas Cr,ν (u, v)
and observe the shapes of the corresponding distributions of both Y and Z:

 t −1 ( u )  t −1 ( v )  (v+2)/2
1 ν ν s2 − 2rst + t2
Cr,ν (u, v) = √ 1+ dsdt,
2π 1 − r2 −∞ −∞ v (1 − r 2 )

where tν−1 ( x ) is the inverse of Student CDF with degrees of freedom ν, r denotes Pearson’s correlation
coefficient between X1 and X2 , and |r | < 1, and ν > 2 is the degrees of freedom. We also consider
r = −0.9, −0.5, 0, 0.5, 0.9 with three degrees of freedom (ν = 3), where r = 0 is corresponding to no
linear correlation. The PDFs and CDFs of Y of Z are, respectively, represented in Figures 3 and 4.
Some percentiles are estimated and displayed in Tables 3 and 4. Similarly to Gaussian copula, in this
case, the center and spread of Y and Z are also varying in the same way. However, one can see a
representation of skewness and tailedness, that is, right skewed if r < 0 and left skewed if r > 0,
since the fact that Student-t Copulas can capture tail dependence between X1 and X2 .

149
JRFM 2019, 12, 42

Table 3. Some percentiles of Y = X1 /X2 , where ( X1 , X2 ) follows Student-t Copulas, ν = 3.

r τ (Cr ) Q0.05 Q0.25 Median Q0.75 Q0.95

−0.9 −0.71 −3.42 −1.26 −0.92 −0.50 1.82
−0.5 −0.33 −5.18 −1.28 −0.56 0.40 4.42
0 0 −5.40 −1.00 0.00 1.00 5.41
0.5 0.33 −4.42 −0.40 0.56 1.28 5.18
0.9 0.71 −1.81 0.50 0.92 1.26 3.42

Table 4. Some percentiles of Y = X1 /( X1 + X2 ), where ( X1 , X2 ) follows Student-t Copulas, ν = 3.

r τ (Cr ) Q0.05 Q0.25 Median Q0.75 Q0.95

−0.9 −0.71 −18.20 −2.05 0.5 3.05 19.21
−0.5 −0.33 −6.62 −0.45 0.5 1.45 7.61
0 0 −3.35 0.00 0.5 1.00 4.35
0.5 0.33 −1.54 0.24 0.5 0.76 2.54
0.9 0.71 −0.24 0.40 0.5 0.60 1.24

X1
Figure 3. PDFs and CDFs of the ratio Y = X2 , where ( X1 , X2 ) follows Student-t Copulas, ν = 3.

X1
Figure 4. PDFs and CDFs of the ratio Z = X1 + X2 , where ( X1 , X2 ) follows Student-t Copulas, ν = 3.

5.3. Clayton Copulas

We turn to investigate dependence structures of X1 and X2 through the following Clayton Copulas
Cθ (u, v) and observe the shapes of the corresponding distributions of both Y and Z:

150
JRFM 2019, 12, 42

/ 0− 1
Cθ (u, v) = max u−θ + v−θ − 1 θ , θ ∈ [−1; +∞) \0.

In practice, we use θ > 0 that leads to

 − 1
Cθ (u, v) = u−θ + v−θ − 1 θ , θ > 0.

For θ = 1, 2, 3, 4, we obtain CDFs and PDFs of Y and Z and exhibit them in Figures 5 and 6,
respectively, and their percentiles as shown in Tables 5 and 6. Clearly, Clayton Copulas affect Y to get
heavier left tail and the more positive dependence is; that is, θ → ∞, the greater the median and the
smaller the IQR of Y tend to be. On the other hand, the shape of distribution of Z is still symmetric
with unchanged median, less spread, and more spike.

Table 5. Some percentiles of Y = X1 /X2 , where ( X1 , X2 ) follows Clayton Copulas.

θ τ (Cθ ) Q0.05 Q0.25 Median Q0.75 Q0.95

1 0.33 −4.97 −0.36 0.53 1.31 5.83
2 0.5 −3.83 0.03 0.76 1.34 5.24
3 0.60 −2.87 0.25 0.86 1.34 4.71
4 0.67 −2.12 0.39 0.91 1.32 4.25

Table 6. Some percentiles of Y = X1 /( X1 + X2 ), where ( X1 , X2 ) follows Clayton Copulas.

θ τ (Cθ ) Q0.05 Q0.25 Median Q0.75 Q0.95

1 0.33 −1.31 0.22 0.5 0.78 2.31
2 0.5 −0.73 0.30 0.5 0.70 1.73
3 0.60 −0.42 0.35 0.5 0.65 1.42
4 0.67 −0.24 0.37 0.5 0.63 1.24

X1
Figure 5. PDFs and CDFs of the ratio Y = X2 , where ( X1 , X2 ) follows Clayton Copulas.

151
JRFM 2019, 12, 42

X1
Figure 6. PDFs and CDFs of the ratio Z = X1 + X2 , where ( X1 , X2 ) follows Clayton Copulas.

5.4. Gumbel Copulas

We now investigate dependence structures of X1 and X2 through the following Gumbel Copulas
Cθ (u, v) and observe the shapes of the corresponding distributions of both Y and Z:
⎛ ⎞
 1
⎝ θ
Cθ (u, v) = exp − (− ln u) + (− ln v) θ θ⎠
, θ > 0.

The parameter θ = 1 implies uncorrelated ( X1 , X2 ). Figures 7 and 8 show the behavior of Y

and Z via the copulas. Tables 7 and 8 represent some estimated percentiles for both Y and Z. In the
comparison with Clayton, Gumbel Copula gets left skewness, higher median, and less spread for Y,
but gets a symmetric shape, unchanged median, less spread, and smaller scale (higher spike) for Z.
However, the shape of Y tends to be more symmetric when one increases the parameter θ.

Table 7. Some percentiles of Y = X1 /X2 , where ( X1 , X2 ) follows Gumbel Copulas.

θ τ (Cθ ) Q0.05 Q0.25 Median Q0.75 Q0.95

1 0 −6.31 −1.00 0.00 1.00 6.32
2 0.5 −3.70 0.00 0.74 1.36 5.01
3 0.67 −2.26 0.38 0.89 1.32 3.95
4 0.75 −1.45 0.56 0.94 1.27 3.30

Table 8. Some percentiles of Y = X1 /( X1 + X2 ), where ( X1 , X2 ) follows Gumbel Copulas.

θ τ (Cθ ) Q0.05 Q0.25 Median Q0.75 Q0.95

1 0 −2.66 0.00 0.5 1.00 3.66
2 0.5 −0.82 0.30 0.5 0.70 1.82
3 0.67 −0.34 0.37 0.5 0.63 1.34
4 0.75 −0.12 0.41 0.5 0.59 1.12

152
JRFM 2019, 12, 42

X1
Figure 7. PDFs and CDFs of the ratio Y = X2 , where ( X1 , X2 ) follows Gumbel Copulas.

X1
Figure 8. PDFs and CDFs of the ratio Z = X1 + X2 , where ( X1 , X2 ) follows Gumbel Copulas.

5.5. Frank Copulas

In addition, we investigate dependence structures of X1 and X2 through the following Frank
Copulas Cθ (u, v) and observe the shapes of the corresponding distributions of both Y and Z:
* , −θu -, -+
1 e − 1 e−θv − 1
Cθ (u, v) = − ln 1 + , θ ∈ R\{0}.
θ e−θ − 1

For Frank Copulas, the parameter θ represents two independent random variables when it tends
to zero, becomes more monotonic structure when θ → ∞, and becomes more counter-monotonic when
θ → −∞. For θ = 1, 2, 3, 4, we have Tables 9 and 10, and Figures 9 and 10. In contrast to both Clayton
and Gumbel Copulas, the density of Y via Frank Copulas is more symmetric and the density of Z is
not scaling too much, but for the median and spread of Y, it behaves like the Gumbel Copulas and
Clayton Copulas; that is, if we increase the parameter θ, the median also increases, whereas the spread
(via IQR) decreases.

Table 9. Some percentiles of Y = X1 /X2 , where ( X1 , X2 ) follows Frank Copulas.

θ τ (Cθ ) Q0.05 Q0.25 Median Q0.75 Q0.95

1 0.11 −6.03 −0.79 0.19 1.18 6.42
2 0.21 −5.59 −0.56 0.36 1.31 6.37
3 0.31 −5.05 −0.33 0.50 1.40 6.19
4 0.39 −4.46 −0.14 0.60 1.46 5.90

153
JRFM 2019, 12, 42

Table 10. Some percentiles of Y = X1 /( X1 + X2 ), where ( X1 , X2 ) follows Frank Copulas.

θ τ (Cθ ) Q0.05 Q0.25 Median Q0.75 Q0.95

1 0.11 −2.09 0.09 0.5 0.91 3.09
2 0.21 −1.62 0.16 0.5 0.84 2.62
3 0.31 −1.25 0.21 0.5 0.79 2.25
4 0.39 −0.96 0.25 0.5 0.75 1.96

X1
Figure 9. PDF and CDF of quotient of the ratio Y = X2 , where ( X1 , X2 ) follows Frank Copulas.

X1
Figure 10. PDF and CDF of the ratio Z = X1 + X2 , where ( X1 , X2 ) follows Frank Copulas.

5.6. Joe Copulas

Finally, we investigate dependence structures of X1 and X2 through the following Joe Copulas
Cθ (u, v) and observe the shapes of the corresponding distributions of both Y and Z:

Cθ (u, v) = 1 − [(1 − u)θ + (1 − v)θ − (1 − u)θ (1 − v)θ ]1/θ , θ ∈ [1, ∞).

Similar to Gumbel Copulas, Joe Copula shows independence when θ = 1 and becomes more
monotonicity if θ → ∞. With assistance of the tables and graphs with θ = 1, 2, 3, 4, Tables 11 and 12
and Figures 11 and 12 tell us that the distribution behaviors of Y are also affected with higher median,
less IQR, smaller scale (higher spike), and the shape is more asymmetric if one increases the parameter
θ. On the other hand, Z still gets median unchanged with median = 0.5 and less spread with the sum
of the first quartile and third quartile is always equal to 1, i.e., (Q0.25 + Q0.75 = 1).

154
JRFM 2019, 12, 42

Table 11. Some percentiles of Y = X1 /X2 , where ( X1 , X2 ) follows Joe Copulas.

θ τ (Cθ ) Q0.05 Q0.25 Median Q0.75 Q0.95

1 0 −6.31 −1.00 0.00 1.00 6.32
2 0.36 −4.82 −0.32 0.57 1.30 5.67
3 0.52 −3.66 0.07 0.79 1.33 5.13
4 0.61 −2.71 0.29 0.88 1.33 4.62

Table 12. Some percentiles of Y = X1 /( X1 + X2 ), where ( X1 , X2 ) follows Joe Copulas.

θ τ (Cθ ) Q0.05 Q0.25 Median Q0.75 Q0.95

1 0 −2.66 0.00 0.5 1.00 3.66
2 0.36 −1.25 0.23 0.5 0.77 2.25
3 0.52 −0.66 0.31 0.5 0.69 1.66
4 0.61 −0.37 0.35 0.5 0.65 1.37

X1
Figure 11. PDFs and CDFs of the ratio Y = X2 , where ( X1 , X2 ) follows Joe Copulas.

X1
Figure 12. PDFs and CDFs of the ratio Z = X1 + X2 , where ( X1 , X2 ) follows Joe Copulas.

5.7. Comparison of Copulas with the Same Measure of Dependence

In this section, we investigate the effects of the six copulas families as discussed above on the
shapes of different distributions for the random variables Y := X1 /X2 and Z := X1 /( X1 + X2 ) when
they have the same measure of dependence—the Kendall’s coefficient τ. Here, the parameters are
chosen to each copula to correspond to Kendall τ = 0.49. We exhibit the corresponding CDFs and PDFs
of both Y and Z in Figures 13 and 14, estimate the percentiles Qα for some α = 0.05, 0.25, 0.5, 0.75, 0.95,
and display the values in Tables 13 and 14. As can be seen from the tables and the figures, Y attains the

155
JRFM 2019, 12, 42

greatest median (0.76) and the smallest IQR (1.32) with left skewed shape when both X1 and X2 follow
Joe Copula and Student-t Copula. In contrast, Gaussian Copula produces the smallest median (0.70)
and the largest IQR (1.42) with symmetric shape. Using Clayton Copula, Y gets the second biggest
median (0.74). The ratio random variable Y has the same median (0.72) for both Gumbel and Frank
Copula, but the Frank makes Y get higher IQR than the Gumbel (1.41 > 1.33). On the other hand,
the random variable Z has symmetric shape with unchanged median (0.5) among all investigated
copulas. It only changes the scale, where the Joe Copula affects Z with the smallest scale (i.e., the tallest
height of density) whilst the Frank Copula causes Z to get the greatest scale (i.e., the shortest height
of density).

Table 13. Some percentiles of Y = X1 /X2 , where ( X1 , X2 ) modelled with six copulas having the same
Kendall coefficient τ = 0.49 .

Copulas Parameters τ (C ) Q0.05 Q0.25 Median Q0.75 Q0.95

Gaussian 0.7 0.49 −3.81 −0.01 0.70 1.41 5.21
Student-t 0.7, ν = 3 0.49 −3.52 −0.02 0.76 1.31 4.63
Clayton 1.90 0.49 −3.93 0.00 0.74 1.34 5.30
Gumbel 1.95 0.49 −3.80 −0.03 0.72 1.36 5.07
Frank 5.5 0.49 −3.60 0.08 0.72 1.49 5.41
Joe 2.8 0.49 −3.88 0.01 0.76 1.33 5.24

Table 14. Some percentiles of Y = X1 /( X1 + X2 ), where ( X1 , X2 ) modelled with six copulas having
the same Kendall coefficient τ = 0.49 .

Copulas Parameters τ (C ) Q0.05 Q0.25 Median Q0.75 Q0.95

Gaussian 0.7 0.49 −0.83 0.29 0.5 0.71 1.83
Student-t 0.7, ν = 3 0.49 −0.92 0.31 0.5 0.69 1.92
Clayton 1.90 0.49 −0.77 0.30 0.5 0.70 1.77
Gumbel 1.95 0.49 −0.86 0.30 0.5 0.70 1.86
Frank 5.5 0.49 −0.65 0.29 0.5 0.71 1.65
Joe 2.8 0.49 −0.74 0.30 0.5 0.70 1.74

X1
Figure 13. PDFs and CDFs of the ratio Y = X2 , where ( X1 , X2 ) modeled with six Copulas having the
same τ = 0.49.

156
JRFM 2019, 12, 42

X1
Figure 14. PDF and CDF of the ratio Z = X1 + X2 , where ( X1 , X2 ) modeled with six Copulas having the
same τ = 0.49.

6. Conclusions
Determining distributions of the functions of random variables is a very crucial task and this
problem has attracted a number of researchers because there are numerous applications in Economics,
Science, and many other areas, especially in the areas of finance including risk management and option
pricing. However, to the best of our knowledge, the problem of determining distribution functions
of quotient of dependent random variables using copulas has not been widely studied and, as far as
we know, no published paper or working paper has done the work we are doing in this paper. Thus,
to bridge the gap in the literature, in this paper, we first develop two general propositions on both
density and distribution functions for the quotient Y = X 1
X2 and the ratio of one variable over the sum
X1
of two variables Z := X1 + X2 of two dependent random variables X1 and X2 by using copulas. We then
X1
derive two corollaries on both density and distribution functions for the two quotients Y = X2 and
X1
Z= X1 + X2 of two dependent normal random variables X1 and X2 in case of Gaussian Copulas by
applying the two main general propositions developed in our paper. From the results, we derive the
corollaries on the median for the ratios of both Y and Z of two normal random variables X1 and X2 .
Furthermore, the result of median for Z is also extended to a larger family of symmetric distributions
and symmetric copulas of X1 and X2 .
Since the density and the CDF formula of the ratios of both Y and Z are in terms of integrals and
are very complicated, we cannot obtain the exact forms of the density and the CDF. To circumvent the
difficulty, in this paper, we propose to use the Monte Carlo algorithm, numerical analysis, and graphical
approach that can efficiently compute complicated integrals and study the behaviors of both density and
distribution and the changes of their shapes when parameters are changing. We illustrate our proposed
approaches by using a simulation study with ratios of normal random variables on several different
copulas, including Gaussian, Student-t, Clayton, Gumbel, Frank, and Joe Copulas. We find that copulas
make big impacts on behavior of distributions, and since Gaussian and Student-t Copulas belong to an
elliptical family, they similarly act on shapes of Y and Z in the same fashion. We also document the
effects when using Archimedean copulas including Clayton, Gumbel, Frank, and Joe Copulas. However,
there are also some differences, especially on location and scale effects. For example, distribution of Z
does not change the median and its shape is always symmetric for all investigated copulas while the
random variable Y is affected in skewness, median and spread. Our findings are useful for academics in
their study of the shapes, center and spread of both density and distribution functions for the ratios by

157
JRFM 2019, 12, 42

using different copulas. Since the ratios in different copulas are widely used in many important empirical
studies in Finance, Economics, and many other areas, our findings are useful to practitioners in Finance,
Economics, and many other areas who need to study the shapes, center and spread of the ratios by using
different copulas in their analysis and useful to policy makers if they need to consider the shapes, center
and spread of both density and distribution functions for the ratios by using different copulas in their
policy decision-making. Finally, we note that, although all of the propositions and corollaries developed
in our paper are relatively easy to derive, all the results developed in our paper are useful to both
academics and practitioners because there is a wide range of applications with variables that have a
negative range. Readers may refer to (Chang et al. 2016, 2018a, 2018b, 2018c; Chang et al. 2015; Wong
2016) for more information on the applications in different areas.

Author Contributions: Writing—original draft preparation, S.L. (Sel Ly) and K.-H.P.; writing—review and
editing,W.-K.W.; visualization, S.L. (Sel Ly) and S.L. (Sal Ly).
Funding: This research has been supported by Ton Duc Thang University, Asia University, China Medical
University Hospital, Hang Seng University of Hong Kong, Research Grants Council (RGC) of Hong Kong (project
number 12500915), and Ministry of Science and Technology (MOST, Project Numbers 106-2410-H-468-002 and
107-2410-H-468 -002-MY3), Taiwan.
Acknowledgments: The authors are grateful to Michael McAleer, the Editor-in-Chief, and anonymous referees
for substantive comments that have significantly improved this manuscript. The fourth author would like to
thank Robert B. Miller and Howard E. Thompson for their continuous guidance and encouragement.
Conflicts of Interest: The authors declare no conflict of interest.

References
Arnold, Barry C., and Patrick L. Brockett. 1992. On distributions whose component ratios are cauchy.
The American Statistician 46: 25–26.
Bithas, Petros S., Nikos C. Sagias, Theodoros A. Tsiftsis, and George K. Karagiannidis. 2007. Products and
ratios of two gaussian class correlated weibull random variables. Paper presented at the 12th International
Conference on Applied Stochastic Models and Data Analysis (ASMDA 2007), Chania, Crete, May 29–June 1.
Cedilnik, Anton, Katarina Kosmelj, and Andrej Blejec. 2004. The distribution of the ratio of jointly normal
variables. Metodoloski zvezki 1: 99.
Chang, Chia-Lin, Michael McAleer, and Wing-Keung Wong. 2015. Informatics, Data Mining, Econometrics and
Financial Economics: A Connection. Technical Report 1. Available online: [Link]
(accessed on 8 March 2019).
Chang, Chia-Lin, Michael McAleer, and Wing-Keung Wong. 2016. Management Science, Economics and Finance:
A Connection. Available online: [Link] (accessed on 8 March 2019).
Chang, Chia-Lin, Michael McAleer, and Wing-Keung Wong. 2018a. Big data, computational science, economics,
finance, marketing, management, and psychology: connections. Journal of Risk and Financial Management
11: 15. [CrossRef]
Chang, Chia-Lin, Michael McAleer, and Wing-Keung Wong. 2018b. Decision Sciences, Economics, Finance,
Business, Computing, and Big Data: Connections. Available online: [Link]
cfm?abstract_id=3140371 (accessed on 8 March 2019).
Chang, Chia-Lin, Michael McAleer, and Wing-Keung Wong. 2018c. Management Information, Decision Sciences,
and Financial Economics: A Connection. Decision Sciences, and Financial Economics: A Connection (January 17,
2018). Tinbergen Institute Discussion Paper, p. 4. Available online: [Link]
abstract_id=3103807 (accessed on 8 March 2019).
Cherubini, Umberto, Elisa Luciano, and Walter Vecchiato. 2004. Copula Methods in Finance. New York:
John Wiley & Sons.
Dolati, Ali, Rasool Roozegar, Najmeh Ahmadi, and Zohreh Shishebor. 2017. The effect of dependence
on distribution of the functions of random variables. Communications in Statistics-Theory and Methods
46: 10704–717. [CrossRef]
Donahue, James D. 1964. Products and Quotients of Random Variables and Their Applications. Technical Repor. Fort
Collins: MARTIN CO DENVER CO.

158
JRFM 2019, 12, 42

Hinkley, David V. 1969. On the ratio of two correlated normal random variables. Biometrika 56: 635–39. [CrossRef]
Joe, Harry. 1997. Multivariate Models and Multivariate Dependence Concepts. Boca Raton: Chapman and Hall/CRC.
Ly, Sel, H. Uyen Pham, and Radim Bris. 2016. On the distortion risk measure using copulas. In Applied Mathematics
in Engineering and Reliability: Proceedings of the 1st International Conference on Applied Mathematics in Engineering
and Reliability (Ho Chi Minh City, Vietnam, 4–6 May 2016). Boca Raton: CRC Press, p. 309.
Ly, Sel, Kim-Hung Pho, Sal Ly, and Wing-Keung Wong. 2019. Determining distribution for the product of random
variables by using copulas. Risks 7: 23. [CrossRef]
Macalos, Milburn O., and Jayrold P Arcede. 2015. On The Distribution of the Sums, Products and
Quotient Of Singh-Maddala Distributed Random Variables Based on Fgm Copula. Available online:
[Link]
the_Sums_Products_and_Quotient_of_Singh-Maddala_Distributed_Random_Variables_Based_on_FGM_
Copula/links/5836ba3a08ae503ddbb54a46/On-the-Distribution-of-the-Sums-Products-and-Quotient-of-
[Link] (accessed on 8 March 2019).
Marsaglia, George. 1965. Ratios of normal variables and ratios of sums of uniform variables. Journal of the
American Statistical Association 60: 193–204. [CrossRef]
Matović, Ana, Edis Mekić, Nikola Sekulović, Mihajlo Stefanović, Marija Matović, and Časlav Stefanović. 2013. The
distribution of the ratio of the products of two independent-variates and its application in the performance
analysis of relaying communication systems. Mathematical Problems in Engineering 2013. [CrossRef]
Mekić, Edis, Mihajlo Stefanović, Petar Spalević, Nikola Sekulović, and Ana Stanković. 2012. Statistical analysis of
ratio of random variables and its application in performance analysis of multihop wireless transmissions.
Mathematical Problems in Engineering 2012. [CrossRef]
Meyer, Christian. 2013. The bivariate normal copula. Communications in Statistics-Theory and Methods 42: 2402–22.
[CrossRef]
Nadarajah, Saralees, and Mariano Ruiz Espejo. 2006. Sums, products, and ratios for the generalized bivariate
pareto distribution. Kodai Mathematical Journal 29: 72–83. [CrossRef]
Nadarajah, Saralees, and Samuel Kotz. 2006a. On the product and ratio of gamma and weibull random variables.
Econometric Theory 22: 338–44. [CrossRef]
Nadarajah, Saralees, and Samuel Kotz. 2006b. Sums, products, and ratios for downton’s bivariate exponential
distribution. Stochastic Environmental Research and Risk Assessment 20: 164–70. [CrossRef]
Nelsen, Roger B. 2007. An Introduction to Copulas. New York: Springer Science & Business Media.
Pham-Gia, Thu, Noyan Turkkan, and E. Marchand. 2006. Density of the ratio of two normal random variables
and applications. Communications in Statistics—Theory and Methods 35: 1569–91. [CrossRef]
Pham-Gia, Thu. 2000. Distributions of the ratios of independent beta variables and applications. Communications
in Statistics-Theory and Methods 29: 2693–715. [CrossRef]
Rathie, Pushpa Narayan, Luan Carlos de SM Ozelim, and Cira E. Guevara Otiniano. 2016. Exact distribution of
the product and the quotient of two stable lévy random variables. Communications in Nonlinear Science and
Numerical Simulation 36: 204–18. [CrossRef]
Sakamoto, H. 1943. On the distributions of the product and the quotient of the independent and uniformly
distributed random variables. Tohoku Mathematical Journal, First Series 49: 243–60.
Springer, Melvin Dale. 1979. The Algebra of Random Variables (1. print. in the USA). Hoboken: Wiley.
Tran, D. Hien, Uyen H. Pham, Sel Ly, and Trung Vo-Duy. 2015. A new measure of monotone dependence by using
sobolev norms for copula. In International Symposium on Integrated Uncertainty in Knowledge Modelling and
Decision Making. New York: Springer, pp. 126–37.
Tran, Hien D., Uyen H. Pham, Sel Ly, and Trung Vo-Duy. 2017. Extraction dependence structure of distorted
copulas via a measure of dependence. Annals of Operations Research 256: 221–36. [CrossRef]
Wong, Wing-Keung. 2016. Behavioural, financial, and health and medical economics: A connection. Journal of
Health & Medical Economics 2: 1.

c 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

159
 Journal of
Risk and Financial
Management

Article
Friendship of Stock Market Indices: A Cluster-Based
Investigation of Stock Markets
László Nagy 1 and Mihály Ormos 2, *
1 Department of Finance, Budapest University of Technology and Economics, Magyar tudosok krt. 2.,
1117 Budapest, Hungary; nagyl@fi[Link]
2 Department of Economics, Janos Selye University, Hradná ul. 21., 94501 Komarno, Slovakia
* Correspondence: ormosm@[Link]

Received: 3 November 2018; Accepted: 11 December 2018; Published: 13 December 2018

Abstract: This paper introduces a spectral clustering-based method to show that stock prices contain
not only firm but also network-level information. We cluster different stock indices and reconstruct
the equity index graph from historical daily closing prices. We show that tail events have a minor
effect on the equity index structure. Moreover, covariance and Shannon entropy do not provide
enough information about the network. However, Gaussian clusters can explain a substantial part of
the total variance. In addition, cluster-wise regressions provide significant and stationer results.

Keywords: cluster analysis; equity index networks; machine learning

1. Introduction
The global stock market structure has to be well understood to diversify risk and manage
cross-border equity portfolios. Appropriate portfolio construction is rather complicated. The linear
dependence structure of the network is not stable (Erdős et al. 2011; Song et al. 2011; Maldonado
and Anthony 1981). Moreover, exogenous shocks have major impact on the correlation structure;
hence, uncorrelated assets could start moving together (Heiberger 2014). Therefore, correlation-based
techniques could cause unwanted variance peaks.
Institutional economic surveys (like MSCI 2018) provide qualitatively identified network
structures e.g., emerging markets and developed markets to stabilize their classification.
The main goal of this study is to provide more suitable quantitative techniques, generalize the
widely used correlation-based portfolio construction framework, discover the equity index network
and make diversification reliable.
The baseline concept follows the Sharpe (1964) Capital Asset Pricing Model (CAPM), in which
similarity measures are calculated from correlations between logarithmic returns (Yalamova 2009). The
anomalies of CAPM indicate a two-dimensional mean-beta framework that gives only a simplified
picture of the real market structure. In order to explain the residuals, financial variables appeared in
the famous regression (Fama and French 1996).
In this paper, we carry out a graph theory-based approach to unveil embedded network level
information (Shi and Malik 2000). We propose non-linear similarity kernels that are able to deal with
higher-order terms. We introduce novel jump-based similarity to investigate the effect of shocks.
In addition, we test whether relative entropy of the distribution functions, that captures non-Gaussian
behavior, conveys network level information. We also investigate the widely used Gaussian smoothing
and correlation (Von Luxburg 2007). We compare different spectral clustering techniques and introduce
the usage of the normalized Newman–Girvan cut (Bolla 2011).
Analyzing historical data supports the a priori assumption that clusters are homogenously
connected. Thus, normalized Laplacian based techniques (Takumasa et al. 2015) are not applicable.

JRFM 2018, 11, 88; doi:10.3390/jrfm11040088 161 [Link]/journal/jrfm

JRFM 2018, 11, 88

However, the proposed Newman-Girvan cut brings suitable, stationary clustering results. We
calculate correlation, jump, relative entropy and Gaussian-based similarities. The figures show
that Newman–Girvan cut outperforms normalized Laplacian technique. Analyzing the spectral
property of the jump-based similarity matrix unveils that exogenous shocks have minor effect on the
network. Thus, our novel results imply that shocks do not convey sufficient information about the
equity index graph. Regression analysis demonstrates the stationarity and explanatory power of the
clusters. Moreover, we shed some light on the node level equity index structure. We unveil that the
index network has scale free properties. Nevertheless, we show that geographical and qualitative
categorizations are in line with clusters.
The article structured as follows: in Section 2 we introduce our spectral clustering-based concept.
In Section 3 we analyze the equity index graph, compare different similarity matrices and clustering
techniques. Section 4 summarizes the article.

2. Materials and Methods

2.1. Data
The current study presents a detailed analysis of 59 stock indices. We apply USD denominated
stock splits and dividend-adjusted daily closing prices between 26 September 1990 and 21 September
2015; data is provided by Thomson Reuters.
Our selection criteria for covered stock indices is based on their classification in the International
Monetary Fund (IMF) Economic Outlook 2015, and the MSCI WORLD Index composition in 2015.
In our analysis we allocate approximately the same weight to each region, despite an unequal number
of countries and market capitalization. We rebalance the sample by choosing approximately ten indices
from each IMF group. We are also interested in the role of well diversified indices e.g., MSCI WORLD
and EURO STOXX600, which have also been analyzed.
In order to underline the highly different characteristics of individual stock indices, we present
some monthly descriptive statistics in Table 1.

Table 1. Descriptive statistics of monthly returns.

Index Mean Variance Skewness

.CSI300 0.018 0.056 −0.336
.XU100 0 0.026 −0.809
.DJI 0.012 0.009 −0.819
.UAX −0.034 0.037 −0.721
.WORLD 0.004 0.002 −1.889
Notes: Table 1 shows the descriptive statistics of the monthly returns, where CSI300, XU100, DJI, UAX, and WORLD
represent the Shanghai Composite 300, Brose Istanbul 100, Dow Jones, Ukraine UX index, and the MSCI World
index respectively.

2.2. Methodology
In the 20th Century, the appearance of large, complex data sets brought new challenges to
developing methods which could be used to understand complicated structures. The key concept
is to classify data points according to various similarity functions. The problem is computationally
extremely challenging. However, spectral clustering techniques provide optimal, lower dimensional
representation of multidimensional data sets. The idea is twofold: on the one hand, similarly to
principal component analysis we could calculate lower dimensional representation of the data points
from the eigenvalues and eigenvectors of the similarity matrix. On the other hand, we could represent
the data structure as a weighted graph and cut the graph along the different clusters. This approach
leads to penalized cut optimization problems. Linear algebra and cluster analysis provide powerful
methods to find the optimal representations and minimized cuts.

162
JRFM 2018, 11, 88

2.2.1. Similarity Matrix

If we would like to cluster different items, first the measurement of similarity has to be decided.
In this study similarity of two stock indices (i, j) will be denoted by Wi,j . The goal is to penalize
differences and reward similarities. Logarithmic returns are easy to handle and maintain all price
process information.
ri (t) = ln (Si (t)/Si (t − 1)), (1)

where Si (t) represents the price of index i. The current study analyses multiple similarity approaches.
First, the Markowitz-based squared correlation is considered a similarity metric.

Wi,j = Corr2 (ri , r j ), (2)

We argue this approach because logarithmic returns are not normally distributed, hence non-linear
effects may also be important. However, as correlation is linear, squared correlation similarities only
take into account linear dependences.
The problem of higher-order moments can be easily solved by using symmetric and positive-definite
kernel functions. The idea comes from the functional analysis. Data can be transformed into a reproducing
kernel Hilbert space (RKHS), where applying the usual statistics provides the same outcomes as can be
attained by using non-linear statistics in the original Hilbert space (Berlinet and Christine 2011); and,
in practice, the Gaussian-kernel is widely used (Gregory et al. 2008).

Wi,j = exp (−  ri − r j 2 ), (3)

We notice that, if the sets of the relevant information and sensitivities are similar, then the relative
entropy of the distribution of return processes is small. Otherwise, we can say stock indices are
sensitive to different sets of information in a different manner (Ormos and Zibriczky 2014). This means
that the similarity function has to be monotonically decreasing in symmetric Kullback–Leibler distance,
and so we can construct a similarity measure such that:
! "
Wi,j = 2/(2 + KL( p(ri )  p(r j )) + KL( p(r j )  p(ri )) ), (4)

where p(ri ) denotes the probability distribution function of logarithmic returns of index i and
def
KL( p(ri )  p(r j )) = ∑ p(ri = x ) ln ( p(ri = x )/p(r j = x )) the relative entropy of indices i and j.
Another perspective argues that large deviations are riskier, hence similarities should be defined
with tail distributions. We calculate the differences of the return series and count the number of at least
two standard deviation peaks. This logic implies that indices are similar if their price processes jump
together. Similarity function has to be decreasing in the number of large deviations, hence we propose
the following metric:
T  
Wi,j = 1/(1 + ∑ δ( zi (t) − z j (t) > 2)), (5)
t =1

where zi represents the normalized return of index i.

In the current study we compare each approach.

2.2.2. Normalized Modularity

The equity index structure is strongly connected. We cannot say that events in Africa do not
have any effect on European markets, hence we have to find methods which can be used to cluster
dense graphs.
Let G (VN ×1 , WN × N ) be a weighted graph, where V denotes the set of vertices and W represents
the weights of the edges. A k-partition of graph G (V, W ) can be defined as the partition of vertices
such that ∪ka=1 Va = V and Vi ∩ Vj = δi,j Vi , ∀i, j ∈ {1, . . . , k}.

163
JRFM 2018, 11, 88

The Wi,j value represents the strength of the connection between nodes (i, j). If we assume that
nodes are independently connected, then the guess of weight Wi,j will be the product of the average
connection strength of i and j. The average connection strength di and d j are given by W,

N
1
di =
N ∑ Wi,u ,
u =1

Thus, Wi,j − di d j captures the information of the network structure (Bolla 2011). If we want to
maximize the sum of information in each cluster, we get:

k , -
max
Pk ∈Pk
∑ ∑ Wi,j − di d j , (6)
a=1 i,j∈Va

where Pk stands for specific k-partition in Pk , which represents the set of all possible k-partitions.
Let M := W − dd T denotes the modularity matrix of G (V, W ). If we would like to get clusters
with similar volumes, then we have to add a penalty to Equation (6), hence we get the normalized
Newman–Girvan cut.
k
1 , -
max ∑ ∑ Wi,j − di d j , (7)
Pk ∈Pk a=1 Vol(Va ) i,j∈V
a

where Vol(Va ) = ∑u∈Va du .

Let us define the so called normalized modularity matrix:

MD := D −1/2 MD −1/2 , (8)

If we would like to cluster a weighted graph G (V, W ), then eigenvectors of its modularity ( M )
and normalized modularity matrices ( MD ) can be used. Modularity and normalized modularity
matrices are symmetric and 0 is always in the spectrum of MD :

N N −1
MD = ∑ λi ui = ∑ λi ui , (9)
i =1 i =1

where 1 > λ1 ≥ λ2 ≥ . . . ≥ λ N ≥ −1 denote the eigenvalues of MD .

If we would like to maximize Equation (7), then we can use the k-means clustering algorithm on
the optimal k-dimensional representation of vertices,
 T
D − 2 u1 , . . . , D − 2 u k
1 1
,

where u1 , . . . , uk denote the corresponding eigenvalues of |λ1 ( MD )|≥ . . . ≥|λk ( MD )| . Moreover, if

the normalized modularity matrix has large positive eigenvalues, then the graph has well-separated
clusters, otherwise clusters are strongly connected.
Another natural approach is to minimize the normalized cut (Von Luxburg 2007).

k −1,k
1 1
min
Pk ∈Pk
∑ ( + )W ,
Vol(Va ) Vol(Vb ) i,j
(10)
a=1,b= a+1

The optimization problem is similar to Equation (7). However, instead of the normalized-
modularity matrix the normalized Laplace matrix provides the solution (Shi and Malik 2000).

L D := D− 2 ( D − W ) D− 2 ,
1 1
(11)

164
JRFM 2018, 11, 88

This technique works when clusters are well separated, otherwise normalized modularity gives
better results.

2.2.3. Algorithm
In empirical analysis, the following steps are the backbone of the calculation (Maurizio et al. 2007).

1. Constructing the similarity matrix (W ).

2. Calculating the normalized modularity matrix ( MD ).
3. Based on the spectral gap, determining the number of clusters and optimal
k-dimensional representation.
4. Appling k-means clustering.

2.2.4. Assessment of Clustering Methods

The relevance of different clustering techniques can be tested in multiple ways. The most
common metrics follow a regression-based logic. In this framework we suppose that variance has two
components: the within, and the between cluster components. Therefore, the explanatory power of
given clusters can be described as:

N 2 N 2
∑kj=1 ∑ j=i 1 ( Xi,j − X ) − ∑ik=1 ∑ j=i 1 ( Xi,j − Xi )
Ni , Nj 2
, (12)
∑i,j=1 ( Xi,j − X )

where k represents the number of clusters, Ni shows the size of clusters and X, X i stands for the total
and cluster wise average (Zhao 2012). The formula penalizes dispersions within clusters, hence dense
clusters would give a number close to 1. Moreover, calculating the ratios with a different number of
clusters highlights the optimal number of clusters.

3. Results
This study presents a broad analysis of the equity index network structure. Logarithmic returns
of 59 stock indices are clustered in different ways. Our investigations reveal stock indices are
homogenously connected, and large price changes have limited effect on the network structure.

3.1. Similarity Metrics

Defining similarity is a key aspect in clustering. In general, it is not usually possible to find an
optimal kernel, but different approaches can be tested and compared to specific data sets.
This study analyzes correlation, jump, entropy, and Gaussian-based similarity kernels. When
calculating the similarity matrices, we expect strongly connected indices have coefficients close to
one, whereas loosely connected close to zero. Level plots (Figure 1) give a feeling about the network
structure which seems to be homogeneous; thus, clusters could not be well separated.
Figure 1 displays the correlation, Gaussian-kernel, relative entropy and jump-based similarity
structure of the equity index graph, in which the whiter the color the stronger the connection between
the indices. Indices are sorted alphabetically and (i, j) represents the similarity between index i and j.
Different similarity measures imply similar patterns, which is in line with our a priori intuition.
However, the spectra of normalized Laplace and normalized modularity matrices help us to find
the most adequate kernel function: the wider the spectral gap, the better the clustering property.
This means, we have to find similarity metrics, which in turn implies large gaps in the spectrum of
normalized Laplacian and modularity matrix (Chung 1997).

165
JRFM 2018, 11, 88

Figure 1. Level plots of daily similarity matrices.

Empirical evidences (Figures 2 and 3) show relative entropy, and Gaussian-kernel can also be used
to cluster the stock index network while correlation and jump-based similarities are not promising.

Figure 2. Eigenvalues of normalized modularity matrix in decreasing order.

Figure 3. Eigenvalues of normalized Laplacian matrix in decreasing order.

A correlation-based similarity approach implies roughly uniform eigenvalue density on [0, 1].
This means, a lot of gaps appear in the spectrum, hence we could not comment on the optimal number
of clusters. Moreover, lower dimensional representations will not contain all the information as some of
the large eigenvalues are not considered. These hurdles highlight the problems of squared correlation
similarity matrices.
Counting at least two standard deviation jumps results in a small number of eigenvalues with
large multiplicity. Therefore, lower dimension representation cannot be used to cluster the data points.
Accordingly, jumps are random and do not reflect the network structure; thus we could say all the
clusters are exposed to the same systematic risk. Thus, the results provide evidence of spillover effect.

166
JRFM 2018, 11, 88

Moreover, we show that shocks and market collapses have a minor effect on the equity index graph
i.e., network structure of equity indices.
Gaussian and relative entropy-based similarity matrices infer promising figures, especially in the
case of normalized modularity. Here, we get large well separated eigenvalues necessary to transform
the data into a lower dimensional space.
Notice that these results are in line with Figure 1 because the normalized Laplacian minimizes the
normalized cut (Equation (10)), which in turn, is small if, and only if, the clusters are loosely connected.
Whereas, the modularity approach maximizes the information of clustering, hence, it can also be used
in a homogeneous network structure as well.
Investigating the spectra, especially the positions of spectral gaps, gives some guidance on the
optimal number of clusters. Considering the previous results, the spectra of Gaussian and relative
entropy-based normalized modularity matrices are suitable. Figure 4 shows indices could be put into
2, 3, or 5 clusters.

Figure 4. Largest eigenvalues of Gaussian- and relative entropy-based normalized modularity matrices.

In order to identify the spectrum gap, we apply the elbow method to identify the optimal number
of clusters. This approach is rather computationally intensive, because of the percentage of variance
explained as a function of clusters has to be estimated (Equation (12)); thus, the whole process has to
be repeated many times. However, in our case, as we have 59 stock indices, the elbow method can also
be used. Figures 5–7 provide evidence for using 2, 3, 4, or 5 clusters.

Figure 5. Explained percentage variance of Gaussian-kernel based clusters of representations.

Figure 6. Histogram of 10,000 Gaussian similarities which are generated from i.i.d. 250 dim. standard
normal samples.

167
JRFM 2018, 11, 88

Figure 7. Explained percentage variance of Gaussian-kernel based clusters after zero out similarities
less than 0.2.

Analyzing the Gaussian similarity kernel shows that if we randomly generate data, then we
would get similarities smaller than 0.2, with probability more than 0.99.
This observation (Figure 6) implies that we have to filter out similarities less than 0.2 from the
adjacency matrix.
Figures 2–4 show the Gaussian-kernel infers the clearest spectrum property. The relative
entropy-based kernel also gives usable results, whereas, jump and correlation-based approaches
are ineffective.

3.2. Comparing Normalized Modularity and Laplacian

We propose the use of an accuracy ratio-based (Engelmann et al. 2003) measure to compare the
efficiency of different clustering techniques. Calculating the area between the variance explanation
function of the random and the different spectral clustering methods generates an appropriate statistic.
Considering this metric (Zhao 2012), it can be seen that the Gaussian-kernel over-performs relative
to the entropy-based approach; this is because in each case its variance explanation function is steeper.
Henceforth, the Gaussian-kernel based normalized modularity matrix is used.

3.3. Equity Index Network Structure

Spectral gap (Figure 4) and variance analyses (Figures 5 and 7) imply equity indices can be studied
by using 2, 3, and 5 clusters. The explanatory power of two clusters is 38%. This means roughly
one-third of the total variance comes from the sample heterogeneity. If we increase the number of
clusters and investigate the three cluster cases, we get a similar explanatory power. However, a spectral
gap appears between the third and fourth eigenvalues (Figure 4), so, theoretically, we propose the
three clusters. The next gap is between the fifth and sixth eigenvalues. The explanation power of five
clusters is 52%. This means, half of the total variance of data can be explained by five clusters.
This result (Figure 8) also suggests that additional clusters have little explanatory power, which is
in line with spectrum properties.

Figure 8. Explained percentage variance of Gaussian-kernel based clusters.

168
JRFM 2018, 11, 88

In practice, mean-variance plots can be used to represent risks and rewards. Intuitively, indices
with similar risk and return can be believed to be similar. This approach applies a k-means algorithm
to cluster the two-dimensional (mean, standard deviation) representation of logarithmic returns.
We have seen this naïve method does not give optimal cuts. However, if we calculate Gaussian
similarities and normalized modularity matrix based representation, then we get clusters with a higher
variance explanatory power. We have seen stock indices can be put into 2, 3, or 5 clusters. If we plot
the mean-variance representation of indices we get Figures 9–11, for 2 and 5 clusters, respectively.

Figure 9. Two Gaussian-kernel based normalized modularity clusters (part of the total graph).

In Figure 9 we can see clusters that are optimizing the modularity cut are concave in a
mean-variance framework. If we have a closer look at the indices in Appendix A (Table A1) we
could say that a qualitative approach also works, because east-west geographical clustering would
imply almost similar results.
Putting the indices into three different clusters (Figure 10) gives a complicated structure, but we
could still state that the first cluster is dominated by European countries, the second by American,
and the third is a mixture of indices from the rest of the world. Thus, applying geographical
diversification is in line with cluster property. The network generated by simple index returns
incorporates geographical information.
Calculating five different clusters helps us to gain a deeper understanding of the network. The
first surprising result is that despite the penalty of different cluster sizes, the Dhaka Stock Exchange
(.DS30) is separated into cluster three. In addition, cluster four contains only two African and two
American indices. Another interesting result is the first cluster, which includes the Arabian indices
except Morocco. Cluster two primarily comprises developed, while cluster five is dominated by
emerging market names. Hence, we could notice that spectral clustering-based classification is similar
to qualitative categorizations. However, these results also suggest that a portfolios constructed using
only geographical scope can integrate indices which behaves significantly differently compared to real
regional regimes.

169
JRFM 2018, 11, 88

Figure 10. Three Gaussian-kernel based normalized modularity clusters, edges with weights stronger
than 0.5.

Figure 11. Five Gaussian-kernel based normalized modularity clusters.

170
JRFM 2018, 11, 88

In order to compare our quantitative approach with geographical and MSCI classifications, we
run the following regressions:
r = β 0 + β 1 σ + β 2 cluster + , (13)

The regressions (Table 2) show that spectral clustering provides statistically reliable figures, while
geographical- and MSCI-based clusters are not statistically significant.

Table 2. Regression statistics.

Method Coeff. of Cluster p-Value

Geographical −0.000036 0.394
MSCI −0.000041 0.293
Spectral −0.000112 0.027
Notes: This table shows the daily linear regression coefficients and p statistics of geographical, MSCI, and spectral
clusters. Returns are regressed on standard deviations and clusters.

The outcomes highlight the difficulty of diversification, because the correlation structure of
the network is quite homogeneous. Moreover, geographical and other qualitative diversification
techniques do not give us statistically significant results. However, indices can be clustered by spectral
methods. This means indices in the same cluster are affected by the same risk factor, hence, only cluster
wise diversification can be used to eliminate non-systematic global risk.

3.4. Equity Index Graph

Clustering helps us to globally analyze the network. However, the local structure can be better
understood by node-specific attributes. Our aim is to find the most influential markets. Hubs can be
identified as vertices with the largest vertex weights. Vertex weight of node i can be defined as the
sum of the edge weights.
N
Vi := ∑ Wi,j δ(Wij > 0.2), (14)
j =1

Calculating the histograms, we get Figure 12.

Figure 12. Histogram of vertex weights, five Gaussian cluster, two nodes are connected if their Gaussian
similarity is stronger than 0.2.

The outcomes show that essentially cluster wise histograms differ. In each cluster, there are
vertices whose connection numbers substantially differ from the cluster wise mean (Figure 12). Note
that the vertex connection density of an Erdő-Rényi graph is binomial, hence hubs and separated nodes
cannot be generated (Erdős and Alfréd 1960). This implies that preferential attachment processes
should be used to model the network structure (Barabási and Réka 1999).

171
JRFM 2018, 11, 88

However, the randomness of vertex weights is twofold: one factor is the number of connections,
while the other factor is edge weights.
In order to distinguish the effects, we calculate the vertex weights as the sum of connections;

N
Vicount := ∑ δ(Wij > 0.2), (15)
j =1

Calculating the histogram of counting-weights we get Figure 13.

Figure 13. Histogram of vertex count-weights, five Gaussian cluster, two nodes are connected if their
similarity is stronger than 0.2.

We could say clusters 1 and 2 contain hubs, whereas, the vertex-count distribution in cluster 5 is
more balanced. There is no hub, but there are vertices with more than 40, and less than 10 connections.
The results show that the shape of the cluster wise vertex connection differs, hence, the vertex weight
distribution is also a mixed distribution.
Comparing Figures 12 and 13 shows that counting implies higher skewness, while having less
effect on the shape. When analyzing edge weights, it turns out that they are not uniformly distributed.
In addition, different clusters have different edge weight densities.
Moreover, it also can be seen (Figure 14), that if the average connection strength is higher, the
vertex has more connections; this is true cluster-wise as well.

Figure 14. Cluster-wise connection number and average connection strength.

All of this implies that spectral clustering techniques can be used to distinguish subgraphs.
Moreover, the number of connections of an index and its average edge weight, follow the preferential
attachment process.

172
JRFM 2018, 11, 88

3.5. Risk and Reward

To understand the connection between risk and reward, we can use the mean-standard deviation
framework. When calculating the regressions we arrive at Table 3. The outcomes imply that the
total sample regression does not provide reliable figures, nevertheless, cluster-wise regressions are
significant. This points to the conclusion that the relationship between risk and return, cluster wise has
different behavior.

Table 3. Descriptive statistics of daily linear regressions.

Clusters p-Value of Intercept p-Value of s.d. R2

Total Sample 0.62 0.12 0.05
First cluster 0.62 0.02 0.68
Second cluster 0.29 0.00 0.59
Fifth cluster 0.93 0.71 0.01
Notes: This table shows the p statistics and R2 values of daily linear regressions. Returns are regressed on standard
deviations. Calculation is done for total, only the first, second and fifth clusters.

Figure 11 and Table 3 show higher standard deviations, implying higher returns, because
regression lines slope upwards. In addition, it also turns out that connections between returns
and standard deviations are strong in Arabian and developed market cases. Nevertheless, emerging
markets show different statistics: index returns in the fifth cluster are not linear in standard deviation,
hence emerging market returns cannot be estimated in the Markowitz framework.

3.6. Time Stability

Making investment decisions vastly depends on the time stability of our strategy. Therefore, we
have to check the stationarity of our clustering method. By splitting the time series by years we get
25 periods. Calculating the stability of explained percentage variance of clustering could be a good
proxy of time stability. Stationarity can be analyzed by the augmented Dicky–Fuller (ADF) test.
Note that, the analysis covers 25 years’ data, hence we get 25 non-overlapping periods. The
t-values (Table 4) show that the variance explanation power process could be stationer, but because
of the small sample size the ADF p-value of 0.32. To gain a better understanding of the results, we
can compare them with the test statistics of randomly generated 25 long standard normal samples
(Figure 15).

Table 4. Augmented Dicky–Fuller (ADF) statistic of explained percentage variance process.

ADF t-Value ADF p-Value

−2.67 0.32
Notes: This table shows the ADF t and p statistics of yearly percentage variance process.

Figure 15. Histogram of ADF statistics of 10,000 independent 25 dim. standard normal sample.

However, we also have to study the time stability of cluster wise mean-standard deviation
regressions. Splitting the data into one-year periods, clustering them and calculating regressions shed
some light on the robustness of clusters (Figure 16).

173
JRFM 2018, 11, 88

Figure 16. ADF test of cluster wise time shifted regressions.

The results show that cluster wise mean-variance regressions are stationary in cluster 1 and 2.
Nevertheless, cluster 3 and 4 are outliers and clusters 5 mostly covers emerging market names. Thus,
the Gaussian-based normalized modularity clustering technique can be used to filter out outliers and
find robust clusters.

4. Discussion
Spectral clustering techniques can be used to discover the equity index structure. On the one hand,
clusters help us to overcome the hardship of heterogeneity and make diversification more efficient.
In our paper we shed some light on the relations between spectral, geographical and qualitative
clustering. It also turned out that Gaussian-kernel based clusters are more suitable than geographical
and qualitative categorizations. In addition, spectral cluster-wise linear regressions give time stationary
and significant results.
On the other hand, we stress that correlation does not convey enough information about the
network; hence linear dependency-based diversification is not optimal (Sharpe 1964; Maldonado
and Anthony 1981). We compared various similarity kernels and spectral clustering methods to
demonstrate the inadequacy of a normalized Laplacian approach (Takumasa et al. 2015) and underpin
the applicability of the proposed Newman–Girvan cut. Moreover, we highlighted that daily closing
prices incorporate the network level information. The results unveiled that tail events have little
effect on the dense network structure, in other words, market shocks have no effect on the cluster
components; thus, index co-movements are not affected by large price changes.
All of these imply spectral clustering can eliminate non-linear effects, thus regular mean-standard
deviation representation gives cluster-wise reliable figures. Instead of qualitative categorization,
we suggest that portfolio managers should use Gaussian-based normalized modularity clusters to
diversify global non-systematic risk.
An interesting field of further research would be analyzing the evolution of the network to identify
patterns that could help us to understand the life cycle of hubs and the vulnerability of the current
equity network.

Author Contributions: Conceptualization, L.N. and M.O.; Methodology, L.N. and M.O.; Validation, L.N. and
M.O.; Formal Analysis, L.N. and M.O.; Investigation, L.N. and M.O.; Writing-Original Draft Preparation, L.N. and
M.O.; Writing-Review & Editing, L.N. and M.O.; Visualization, L.N.; Supervision, M.O.
Funding: Research fund was provided by Pallas Athéné Domus Educationis.
Acknowledgments: The authors would like to gratefully acknowledge the valuable comments and suggestions
of three anonymous referees that contributed to a substantially improved paper. Mihály Ormos acknowledges
the support of the János Bolyai Research Scholarship of the Hungarian Academy of Sciences and the support
of the Pallas Athéné Domus Educationis Foundation. The views expressed are those of the authors and do not
necessarily reflect the official opinion of the Pallas Athéné Domus Educationis Foundation.

174
JRFM 2018, 11, 88

Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the
study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to
publish the results.

Appendix A

Table A1. Clusters of stock indices.

Country Two Clusters Three Clusters Five Clusters

United Arab Emirates 2 3 1
Saudi Arabia 2 3 1
Qatar 2 1 1
Kuwait 2 1 1
Egypt 2 3 1
Bahrain 2 1 1
Vietnam 2 1 1
Nigeria 2 1 1
Dow Jones 1 2 2
Denmark 1 1 2
Switzerland 1 1 2
Canada 1 2 2
Mexico 1 2 2
Chile 1 2 2
Argentina 1 2 2
Hungary 1 1 2
Morocco 2 1 2
S&P 500 1 2 2
MSCI World 1 2 2
Czech Republic 1 1 2
Togo 2 1 2
Spain 1 1 2
Norway 1 1 2
Luxembourg 1 1 2
France 1 1 2
South Africa 1 3 2
Euro Stocks 1 1 2
Sweden 1 1 2
UK 1 1 2
Netherlands 1 1 2
Finland 1 1 2
Poland 1 3 2
Germany 1 1 2
Belgium 1 1 2
Italy 1 1 2
Brazil 1 2 2
Colombia 1 3 2
Bangladesh 2 1 3
Costa Rica 2 1 4
Zambia 2 1 4
Malawi 2 1 4
Venezuela 2 1 4
South Korea 2 3 5
Hong Kong 2 3 5
Thailand 2 3 5
China 2 3 5
Kenya 2 3 5
India 2 3 5
Namibia 2 3 5
Turkey 2 3 5
Indonesia 2 3 5
Malaysia 2 3 5
Russia 2 3 5
Australia 2 3 5
Taiwan 2 3 5
Japan 2 3 5
Ukraine 2 3 5
Bulgaria 2 1 5
Romania 2 3 5
Notes: This table contains the list of indices and clustering results for 2, 3, and 5 clusters.

175
JRFM 2018, 11, 88

References
Barabási, Albert L., and Albert Réka. 1999. Emergence of Scaling in Random Networks. Science 26: 509–12.
[CrossRef]
Berlinet, Alain, and Thomas-Agnan Christine. 2011. Reproducing Kernel Hilbert Spaces in Probability and Statistics.
Berlin: Springer Science & Business Media, pp. 1–108. ISBN 978-1441990969.
Bolla, Marianna. 2011. Penalized version of Newman-Girvan modularity and their relation to normalized cuts
and k-means clustering. Physical Review E 84: 016108. [CrossRef] [PubMed]
Chung, Fan R. G. 1997. Spectral Graph Theory. Providence: American Mathematical Society, No. 92. pp. 14–81.
ISBN 978-0821803158.
Engelmann, Bernd, Evelyn Hayden, and Dirk Tasche. 2003. Measuring the Discriminative Power of Rating Systems.
Banking and Financial Supervision. Frankfurt: Deutsche Bundesbank.
Erdős, Péter, and Rényi Alfréd. 1960. On the Evolution of Random Graphs. Acta Mathematica Hungarica 5: 17–61.
Erdős, Péter, Mihály Ormos, and Dávid Zibriczky. 2011. Non-parametric and semi-parametric asset pricing.
Economic Modelling 28: 1150–62. [CrossRef]
Fama, Eugene, and Kenneth R. French. 1996. Multifactor explanations of asset pricing anomalies. The Journal of
Finance 51: 55–84. [CrossRef]
Maurizio, Filippone, Francesco Camastra, Francesco Masulli, and Stefano Rovetta. 2007. A survey of kernel and
spectral methods for clustering. Pattern Recognition 41: 176–90. [CrossRef]
Heiberger, Raphael H. 2014. Stock network stability in times of crisis. Physica A: Statistical Mechanics and Its
Applications 393: 376–81. [CrossRef]
Gregory, Leibon, Scott Pauls, Daniel Rockmore, and Robert Savell. 2008. Topological Structures in the Equities
Market Network. PNAS 105: 20589–94. [CrossRef]
Von Luxburg, Ulrike. 2007. Tutorial on Spectral Clustering. Statistics and Computing 17: 395–416. [CrossRef]
Maldonado, Rita, and Saunders Anthony. 1981. International portfolio diversification and the inter-temporal
stability of international stock market relationships, 1957–1978. Financial Management 10: 54–63. [CrossRef]
MSCI. 2018. Market Classification. Available online: [Link] (accessed on 3
November 2018).
Ormos, Mihály, and Dávid Zibriczky. 2014. Entropy-Based Financial Asset Pricing. PLoS ONE 9: E115742.
[CrossRef] [PubMed]
Shi, Jianbo, and Jitendra Malik. 2000. Normalized cuts and image segmentation. IEEE Pattern Analysis and Machine
Intelligence 22: 888–905. [CrossRef]
Sharpe, William F. 1964. Capital asset prices: A theory of market equilibrium under conditions of risk. Journal of
Finance 19: 425–42.
Song, Dong-Ming, Michele Tumminello, Wei-Xing Zhou, and Rosario N. Mantegna. 2011. Evolution of worldwide
stock markets, correlation structure, and correlation-based graphs. Physical Review E 84: 026108. [CrossRef]
[PubMed]
Takumasa, Sakakibara, Tohgoroh Matsuib, Atsuko Mutoha, and Nobuhiro Inuzuka. 2015. Clustering mutual
funds based on investment similarity. Procedia Computer Science 60: 881–90. [CrossRef]
Yalamova, Rossitsa. 2009. Correlations in Financial Time Series during Extreme Events-Spectral Clustering and
Partition Decoupling Method. Paper presented at World Congress on Engineering, London, UK, July 1–3,
Volume 2, pp. 1376–78.
Zhao, Yanchang. 2012. R and Data Mining: Examples and Case Study. Cambridge: Academic Press, pp. 49–59.

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

176
 Journal of
Risk and Financial
Management

Brief Report
Capital Structure and Firm Performance in Australian
Service Sector Firms: A Panel Data Analysis
Rafiuddin Ahmed 1, * and Rafiqul Bhuyan 2
1 Program of Accounting and Finance, James Cook University, Douglas, QLD 4814, Australia
2 Department of Accounting and Finance, Alabama A&M University, Normal, AL 35762, USA;
rafi[Link]@[Link]
* Correspondence: rafi[Link]@[Link]

Received: 19 August 2020; Accepted: 16 September 2020; Published: 18 September 2020

Abstract: Using cross-sectional panel data over eleven years (2009–2019), or 1001 firm-year
observations, this study examines the relationship between capital structure and firm performance
of service sector firms from Australian stock market. Unlike other studies, in this study directional
causalities of all performance measures were used to identify the cause of firm performance. The study
finds that long-term debt dominates debt choices of Australian service sector companies. Although
the finding is to some extent similar to trends in debt financed operations observed in companies in
developed and developing countries, the finding is unexpected because the sectoral and institutional
borrowing rules and regulations in Australia are different from those in other parts of the world.

Keywords: firm performance; causality tests; leverage; long-term debt; capital structure

1. Introduction
Capital structure is one of the most perplexing puzzles in the financial literature that deals with
solutions to optimal mix of debt and equity. The seminal work of Modigliani and Miller (1958) initiated
this body of work, other researchers later developed theories along the MM, and empirical researchers
validated the assumptions underlying the theoretical body of the literature by examining different
dimensions such as firm characteristics, time or industry sector category. A mirror image of capital
structure choice is essentially a decision to fund capital from the cheapest sources to maximize income
after taxes (Yazdanfar 2012). The seminal work of Jensen and Meckling (1976) posits managerial
behavior in the best interest of the shareholders which is to borrow at a level that will maximize
shareholder value and firm profitability. Since the work of Jensen and Meckling (1976) several
researchers have examined the relationship between leverage and profitability. The findings of these
studies are contradictory and mixed, some suggesting a positive relationship (Ghosh et al. 2000;
Hadlock and James 2002; Roden and Lewellen 1995; Taub 1975) and some suggesting a negative
relationship (Fama and French 1998; Gleason et al. 2000) between leverage and profitability (El-Sayed
Ebaid 2009). There are many studies on capital structure in the context of service sectors in Europe,
USA, the Middle East and other parts of the world (Chakrabarti and Chakrabarti 2019; Choi et al. 2018;
Park and Jang 2018; Sardo et al. 2020; Sermpinis et al. 2019; Szemán 2017). Compared to other sectors,
service sector capital structure research is at a nascent stage. Further research needs to be done to
enrich the understanding of the drivers of financial performance of this sector.
The key aim of this paper is to empirically examine the relationship between debt financing and
firm performance of service sector companies listed in the Australian Stock Exchange. The service
sector is chosen to reflect the changing configuration of the Australian economy from a resource-based
economy to a service-based economy. Over the last one and a half decades, the Australian service
sector contributed between 60–70% and is a major employer (Australian Bureau of Statistics 2019,
2020). This trend is expected to continue in the foreseeable future and it is important to get some

JRFM 2020, 13, 214; doi:10.3390/jrfm13090214 177 [Link]/journal/jrfm

JRFM 2020, 13, 214

insights into the effect of capital structure on this sector firms. Four performance measures are used
to capture firm performance: (a) return on asset, (b) return on equity, (c) return on capital employed
and (d) operating margin. The paper finds: (a) portability (measured by return on equity, ROE) and
leverage (measured by a ratio of short-term debt to total assets) is positively associated, (b) profitability
(measured by return on assets) and leverage (measured by short-term debt) is positively associated
and (c) no significant association between either ROE and ROA and long-term or total debt. The main
contribution of this paper is that it has extended the current body of literature on capital structure
by adding the Australian service industry context from very recent data. Australia’s move from a
resource-based economy to a service-based economy means the sector is growing, so the findings of
this paper are expected to shed light on this emerging frontier of capital structure practices of service
sector firms. The remainder of the paper is organized as follows. In Section 2, the literature is discussed.
In the third section, the empirical literature is reviewed, followed by three sections on data, results and
discussions. The final section concludes the paper with some possible directions for future research.

2. Literature Review
Numerous theories have been developed following the initial development of capital structure
theory by Modigliani and Miller (1958). These theories were later classified by their assumptions
about how they affect firm value in the financial market. The first of these theories is the Trade-off
theory of capital structure. This theory precedes some initial refinements in 1963 (Modigliani and
Miller 1963) of Modigliani and Miller’s (1958) initial work, in which taxes are added to theorize the
effect of taxes on a firm’s tax payable amount, increase in after tax income and its market value.
This development was later labelled as trade-off theory, a theory which states that a firm’s optimal
leverage is achieved by minimizing taxes, costs of financial distress and agency costs. Baxter (1967)
argued that increased debt levels increases the chances of bankruptcy and increases interest payable to
the debtholders. A firm’s optimal leverage is where tax advantage from debt exactly equals the cost of
debt. Kraus and Litzenberger (1973) argue that a firm’s market value declines if its debt obligations
are greater than its earnings. DeAngelo and Masulis (1980) propose the static trade off theory and
include other tax minimizing offsets such as depreciation and investment tax credits. They argue that
firms weigh tax advantages of debt against business risk (a cost). Their theoretical model proposes that
a firm’s optimum debt level is where the present value of tax savings from debt equals the present
value of costs of distress.
Myers (1984), in his theoretical explanation of the asymmetric information hypothesis, proposes
different information held by firms’ internal and external stakeholders. Managers hold real information
about firms’ income distribution plans (Ross 1977). Thus, firm’s leverage level signals its confidence
levels, suggesting lower leverage as a poor signal about income and its distribution potential and vice
versa. Pettit and Singer (1985) discuss the problems of asymmetric information and possible agency
costs affecting firms’ demand and supply of credit. They argue that small firms possess a higher level
of asymmetric information due to financial constraints for sufficient disclosure of financial information
to outsiders. This theory has laid the foundation for Pecking Order Theory (POT). Donaldson (1984)
proposes the concepts and ideas of Pecking Order Theory (POT) which was later refined by Myers (1984)
and Myers and Majluf (1984). The fundamental premise of this theory is that firms’ preferences for
funding is stacked by a pecking order of risk preferences and corresponding costs. Thus, firms use
the cheapest source of internal funds such as retained earnings, debt, convertible debt and preference
shares) and external equity (Myers 1984). The cost of sourcing extra funding is dependent on the extent
of information asymmetries of risk perceptions emanating from differential information needs held by
inside management and potential investors. In addition to a firm’s desire to source the cheapest fund
to finance its needs, other factors, such as the stage of development of a firm (a startup, a mature firm
etc.) influence the supply of funds (Macan Bhaird and Lucey 2010).
Agency theory (Jensen and Meckling 1976) addresses the fundamental problem of managing a
firm’s capital structure from the cheapest source of funds. While common equity is an expensive source

178
JRFM 2020, 13, 214

of funds, its use results in suboptimal firm value when equity holders insist on risk reduction from
lower leverage usage. If managers’ and shareholders’ interests are not aligned, it is highly unlikely
that optimal firm value is ever going to eventuate from managerial actions. The debtholders’ risk
perceptions encourage them to ask for debt covenants or other costly debt shielding instruments.
The tensions between the two subgroups of owners impose increased risk of monitoring by management,
resulting in costly monitoring and hence, agency costs. A number of remedial measures can be
implemented such as reduction in consumption of resources when debt and bankruptcy risks increase
(Grossman and Hart 1982), increasing the stake of managers in a firm or increasing the leverage
(Jensen 1986), commonly packed as ‘free cash flow hypothesis’. Free cash flow hypothesis proposes
adoption of measures to reduce free cash flow at managers’ disposal by increasing leverage (Stulz 1990)
so that less cash flow is available for desired investment choices.
The theories above are prevalent in different country specific studies. An empirical study by
El-Sayed Ebaid (2009) on Egyptian firms suggest a negative relationship between profitability and
shorter-term or total debt when return on asset is used to measure profitability. The results also suggest
no significant relationship between short-term or long-term debt and profitability when return on
equity or gross margin is used as a measure of profitability. Salim and Yadav’s (2012) study on 237 listed
Malaysian companies from 1995–2001 found a negative relationship between short-term and long-term
debts and all measures of profitability, return on assets, return on equity and earnings per share.
Ahmed Sheikh and Wang (2011) examined 240 listed Pakistani non-financial companies during the
2004–2009 period. Three statistical tests, fixed effects, random effects and ordinary least squares found
negative relationships between debt and return on assets. Weill (2008) used the maximum likelihood
estimation method to analyze the effect of financial leverage on the performance of 11,836 firms from
seven European countries over a three-year time period, 1998–2000. The results indicate that the
long-term debt ratio is positively related at statistically significant level in Spain and Italy but negatively
related at statistically significant level in Germany, France, Belgium and Norway, and insignificantly
in Portugal. Goddard et al. (2005) used the generalized methods of moments system to test the
determinants of profitability of manufacturing and service firms in Belgium, France, Italy and the
U.K. from 1993–2001. They found a negative relationship between the sample firms’ gearing ratio and
profitability, and higher profitability in more liquid firms. Abor (2007) used a generalized least squares
regression to study a sample of 160 Ghanaian and 200 South African Small and Medium Enterprises
(SMEs) from 1998–2003 and found a negative relationship between longer-term and total debt ratios and
profitability. Yazdanfar and Öhman’s (2015) study used 15,897 Swedish SMEs from five different sectors
from 2009–2012 to examine the effect of three different forms of debt ratios, trade credit, short-term
debt and long-term debt on profitability. The results suggest a negative relationship between all types
of debts and profitability, suggesting an increased use of equity capital to finance Swedish SMEs.
There are not many Australian studies on the relationship between capital structure and profitability.
Li and Stathis (2017) examined the determinants of the capital structure of Australian manufacturing
listed traded firms. The study used eight factors: profitability, log of assets, median industry leverage,
industry growth, market-to-book ratio, tangibility, capital expenditure and investment tax credits.
They found weak support for the pecking order hypothesis and increasing support for the trade-off
theory in Australia. Qiu and La (2010) examined the relationship between firm characteristics and
capital structure of 367 Australian firms over a 15 year period. Their study identified the role of debt
on profitability, tangibility, growth prospects and risk of these firms. They concluded that profitability
has the potential to reduce debt levels of Australian firms, implying debt reduction through increased
profits was possible in Australian firms. Barth et al. (2001) examined the relationship between
capital structure and profitability of 107 countries including Australia. They tested for regulatory
power, supervision, and other factors affecting the relationship between profitability and leverage
across the countries studied. Rashid and Islam (2009) examined 60 companies in the Australian
Financial services sector during the years 2002–2003. The results suggest that profitability is negatively

179
JRFM 2020, 13, 214

affected by leverage, and positively affected by board size, liquid markets and information efficiency
(all control variables).
Firm performance as a measure of the impact of different proxies for capital structure has
added new insights in recent times. Some country-specific studies have examined the direct effect
of using different types of debts on firm performance. Most of these studies reported a significant
negative relationship between debts and firm performance. Chakrabarti and Chakrabarti (2019)
examined firm-specific and macro-economic variables on 18 Indian non-insurance firms for seven
years. They found a positive relationship between low insurance, low input costs, low inflation
rates, higher return on investment, liquidity and profitability. Dalci (2018) examined the impact of
capital structure on 1503 listed manufacturing firms in the Chinese stock exchange between the years
2008–2016. They found an inverted U-shaped relationship between capital structure and profitability
and provided the causes of a negative and positive relationship between financial leverage (as a
measure of capital structure) and profitability. This is a major study that highlighted the importance of
the developments of credit market policies and rules for the advancement of different-sized Chinese
manufacturing firms.
Dave et al. (2019) examined the impact of capital structure and profitability of firms in the Indian
Steel industry and observed a significant negative relationship between long-term and short-term debts
as a ratio of total assets and profitability. Helmy et al. (2020) examined the impact of capital structure,
internal governance mechanism, and firm-performance of 183 Bursa-listed Malaysian companies
for the years 2007–2010. They found a positive impact of capital structure on firm performance.
Gharaibeh and Bani Khaled (2020) examined the factors that played key roles in the profitability of
46 Jordanian service sector companies between the years 2014–2018. They found that debt as a portion
of total assets and tangible assets have significantly negative relationships with profitability whereas
tangible size and business risk had a positive relationship with profitability.
Hussein et al. (2019) examined listed Jordanian firms between 2005–2017. Using three measures
of firm performance, return on assets, Tobin’s Q and return on assets, and total and short-term debt
as a proxy for capital structure, they observed a positively significant relationship between firm size,
asset growth, significant negative relationship between short-term debt and long-term debt and return
on assets. However, they did not find any significant negative relationship between short-term and
long-term debts and return on equity measure of firm performance. Lastly, Yazdanfar examined 15,897
firms working in five SME sectors of the Swedish economy between 2009–2012. They found debt ratios
(trade-credit, short-term and long-term debts) negatively affected firm profitability.
Capital structure studies that examined the relationship between different proxies for capital
structure and firm performance used a variety of measures to define profitability. Some studies used
a single measure (see, for example, Arifin 2017; Negasa 2016) while others used multiple measures
such as return on Equity (ROE), return on assets (ROA), and return on capital invested (ROCE) (see,
for example, Gharaibeh and Bani Khaled 2020; Musah and Kong 2019). In these studies, different types
of debts are used as proxies for capital structure and different control variables are used to measure
the collective impacts on firm performance. The relationship between firm performance and capital
structure is assumed to be unidirectional in most of the studies reviewed above. However, some recent
studies validated the causal relationship between capital structure and firm performance (Arifin 2017;
M’ng et al. 2017). Finally, the studies above showed a negative, positive, and mixed relationship
between capital structure measures and firm performance.
The studies are from diverse sectors and cover a wide range of firm year cross sectional observations.
There is a limited number of studies that examine the linkage between different measures of firm
performance (or profitability) and capital structure. Studies covering the services sector are hardly
noteworthy in the Australian and global contexts. Moreover, the directional causal relationship
between different types of borrowings and firm performance is hardly examined in detail in the studies
reviewed above. This study contributes to the growing body of literature in the study of capital
structure in the under-researched domains of service sector firms in Australia and internationally.

180
JRFM 2020, 13, 214

3. Data
We consider a comprehensive database from the Australian service sector (as classified by
Australian Bureau of Statistics) for the period 2009 to 2019. The data was collected from Datanalysis
database-a database that publishes financial data of companies in different Australian sectors. Although
our initial sample was much larger than what we have included in the study, due to matching
inconsistency in variable definition and the availability of all variables of all companies, we have
truncated the data to 91 companies that have same data set for the entire time period. These companies
are all listed in the Australian Stock Exchange.
Table 1 shows that a total of nine sectors are considered to conduct research for the period 2009 to
2019. Based on the availability of the entire data set with chosen variables, some sectors had the most
samples and others had only a few companies. The percentage column shows the degree of weight
from each sector of our sample. Based on the literature surveyed, we consider several variables shown
in Table 2 to investigate our research question.
To avoid spurious regression estimates in our empirical analysis, variables under consideration
should ideally be stationary. To confirm this, we used the panel unit root test of Levin et al. (2002).
Table 3 shows that the unit-roots hypothesis is rejected by all variables at the 1% level of significance.
Following (Canarella and Miller 2018; Köksal and Orman 2015; Khan et al. 2018; M’ng et al. 2017),
we also checked for stationarity using a unit root test and observed that all variables were stationary with
respect to the dependent variables (Return on Equity (ROE), Operating Margin (op_margin), Return
on Asset (ROA) and Return on Invested Capital (ROIC)), confirmed by the tests for heteroscedasticity
and autocorrelation diagnostics.
The panel regressions were run for four dependent variables (return on equity, return on assets,
return on invested capital, and operating margin), two treatment variables (leverage and long-term
debt to total assets ratio), and five control variables (size, liquidity, revenue growth for three years,
tangibility and depreciation tax shield). A series of regressions were run for these variables and
diagnostic tests were conducted to confirm the appropriateness of fixed or random effects panel
regressions models.
For each of the dependent variables, outputs for two models are presented, after eliminating the
inappropriate models using Hausman tests. The Breausch Pagan test was employed to confirm the
outputs of the Hausman test for this purpose. Earlier studies in capital structure used the Hausman
test to identify the appropriate panel data model from two available models: fixed effect model
and random effect model (Dalci 2018; Mayuri and Kengatharan 2019; Sivalingam and Kengatharan
2018; Suntraruk and Liu 2017). Breausch Pagan Lagrange Multiplier tests were used for confirming
the appropriateness of the random effects model (see for example, Dalci 2018; Ghasemi et al. 2018;
Khan et al. 2018). The tables below present the outputs of these models.

Table 1. Table shows the various Australian service sector companies considered for this research.

Sector-ID Frequency Percent Cumulative Percentage

Utilities 242 24.18 24.18
Construction 33 3.30 27.47
Retail trade 132 13.19 40.66
Transport 143 14.29 54.95
Communication services 308 30.77 85.71
Consumer discretionary 22 2.20 87.91
Commercial services 121 12.09 100.00
Total 1001 100.00
(Source: Authors’ compilation).

181
 Table 2. List of dependent and explanatory variables.

Variables Calculated as Sources

Earnings before Interest and tax (EBIT)/Total (Dalci 2018; Gharaibeh and Bani Khaled 2020; Goddard et al.
Return on assets (ROA)
assets 2005; Nunes et al. 2009)
JRFM 2020, 13, 214

Return on equity (ROE) EBIT/Total Equity (Arifin 2017; Dalci 2018; Gharaibeh and Bani Khaled 2020)
Operating margin (op_margin) EBIT/Operating revenue (Gharaibeh and Bani Khaled 2020)
Return on Invested Capital (ROIC) EBIT/Invested capital (Musah and Kong 2019)
Leverage Total liabilities/total assets (Gharaibeh and Bani Khaled 2020; Nunes et al. 2009)
Long-term debt to total asset (LTd_TA) Long-term debt/total assets (Yazdanfar and Öhman 2015)
Liquidity Current assets/Current liabilities (Nunes et al. 2009)
(Fitim et al. 2019; Gharaibeh and Bani Khaled 2020; Nunes et
Tangibility Fixed assets/Total assets
al. 2009; Shalini and Biswas 2019)
(Fitim et al. 2019; Shalini and Biswas 2019; Yazdanfar and
Tax shield Depreciation/Total assets
Öhman 2015)
Operating revenue (size) Log of Operating revenue (Fitim et al. 2019; Shalini and Biswas 2019)
Revenue growth (3-year) % of revenue growth (3 yearly average, given) (Chadha and Sharma 2015; Chakrabarti and Chakrabarti 2019)

Table 3. Unit root tests results.

182
ROIC ROE ROA OP_MARGIN LTD_TA LEVERAGE TAXSHIELD TANBIBILITY LIQUIDITY LNSALES
−6.4783 −6.04 −9.75 −7.0134 −9.1691 −3.2673 −5.5693 −12.2987 −5.9927 −11.4026
(0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000)
JRFM 2020, 13, 214

4. Results
In the following section, we have presented the results of our analysis. Four different measures
of financial performance and six explanatory variables were analyzed to identify the important
explanatory variables affecting firm performance of Australian service sector firms between the years
2009 and 2019. In each table below, two models are presented: Model 1 and Model 2. In Model
1, leverage is used as a treatment variable and in Model 2, long-term debt is used as the treatment
variable. Size, depreciation tax-shield, revenue growth for 3 years, operating revenue (measure of size),
liquidity and tangibility are used as control variables in measuring firm performance.
As shown in Table 4, the fixed effect model (FEM) in Model 1 identified leverage, tangibility,
liquidity and operating revenue as important predictors of operating margin. The random effect model
(REM) in Model 1 identified leverage, tangibility, operating revenue and revenue growth for three years
as significant predictors of operating margin. The constant is also important at 1% level of significance.
The Granger causality test shows a unidirectional relationship between leverage and operating margin.
This relationship is positive as evidenced by a significant positive coefficient of leverage.

Table 4. Panel regression outputs for operating margin (dependent variable).

Dependent Variable Model 1 Model 2

FEM REM FEM REM
0.5972 * 0.2411 **
leverage
(4.48) (2.23)
−0.0447 −0.0295
LTD_TA
(−1.21) (−0.88)
−0.0041 −0.1265 −0.0208 −0.0179
Depreciation tax shield
(−0.10) (−0.36) (−0.49) (−0.51)
−0.04682 ** −0.0584 *** −0.0640 −0.0648 **
Tangibility
(−1.21) (−1.80) (−1.63) (−1.97)
0.1589 ** 0.0191 −0.0983 *** −0.0893 ***
Liquidity
(2.15) (0.32) (−1.69) (−1.78)
0.0075 * 0046 * 0.0086 * 0.0059 *
Operating revenue
(2.39) (2.84) (2.68) (3.68)
−0.0081 −0.0074 ** −0.0091 * −0.0080 *
Revenue growth (3-year)
(−2.66) (−2.45) (−2.95) (−2.65)
0.0454 0.5986 * 0.9164 * 0.9457 *
Constant
(0.22) (3.92) (9.34) (12.37)
6.49 3.32
F-test
(0.0000) (0.0031)
24.85 3.80
Hausman test
(0.0000) (0.7033)
445.16 480.54
Breusch-Pagan test
(0.0000) (0.0000)
*, ** and *** are used for 1%, 5% and 10% level of significance.

In Model 2, the fixed effect model identified liquidity, operating revenue and revenue growth
as significant predictors of operating margin. In the random effects model, tangibility, liquidity and
operating revenue are significant predictors of operating margin. In both models, the constants are
significant at 1% level. The Granger causality test revealed a unidirectional relationship between
long-term debt to total asset and operating margin. This relationship is negative but not significant at
any level.
In Table 5, the random effect model of Model 1 identified leverage, operating revenue, revenue
growth as significant predictors of return on assets. In the fixed effect model of Model 1, leverage,

183
JRFM 2020, 13, 214

operating revenue and growth are significant predictors of return on assets. The Granger causality test
indicates a bi-directional causality between leverage and return on asset. Leverage significantly pulls
return on assets down, as evidenced by the negative coefficient in the equations in Model 1.

Table 5. Panel regression outputs for Return on Asset (ROA).

Variables Model 1 Model 2

REM FEM REM FEM
−1.1143 * −1.6487 *
leverage
(−3.75) (−4.58)
0.2475 ** 0.1819 **
LTD_TA
(2.49) (1.98)
0.01242 −0.1814 −0.1288 0.0429
Depreciation tax shield
(0.13) (−1.60) (−1.13) (0.44)
−0.01436 0.05435 0.0810 0.0045
Tangibility
(−0.16) (0.52) (0.77) (0.05)
0.2029 0.1243 0.9096 * 0.7347 *
Liquidity
(1.24) (0.62) (5.82) (5.35)
0.02992 * 0.01709 ** 0.0120 0.0232 *
Operating revenue
(6.61) (2.01) (1.39) (5.19)
0.0168 ** 0.01639 ** 0.0193 ** 0.01966 **
Revenue growth (3-year)
(2.06) (1.98) (2.32) (2.38)
0.3210 1.2644 ** −1.2744 * −1.3402 *
Constant
(0.76) (2.03) (−4.82) (−6.41)
10.06 7.48
F-test
(0.0000) (0.0000)
26.13 21.04
Hausman test
(0.0002) (0.0016)
544.07 532.38
Breusch-Pagan test
(0.0000) (0.0000)
*, ** and *** are used for 1%, 5% and 10% level of significance.

The constant is also significant at the 5% level. We also observe that in both fixed effect model
and random effect model of Model 2, long-term debt to total assets, liquidity, operating revenue and
revenue growth for three years are significant predictors of return on assets. The constant is also
significant at the 1% level.
In the Table 6, the fixed effect model in Model 1 (leverage as treatment variable), leverage,
tax shield, tangibility and operating revenue are significant explanatory variables of return on capital
invested. In the random effects model, leverage, tax shield, tangibility and operating revenue are
significant predictors of return on invested capital. In both models, the constants are significant at the
10% level of significance. The Granger causality test indicates a bi-directional relationship between
leverage and return on invested capital. This relationship is negative, as evidenced by the negative
coefficient of leverage.
In Model 2 (long-term debt as a treatment variable), long-term debt, tax shield, tangibility, liquidity
and operating revenue are significant predictors of return on capital employed. In the random effects
model in Model 2, long-term debt, tax shield, tangibility, liquidity and operating revenue are significant
predictors of return on invested capital. The constant is also significant at the 10% level. When we run
the Granger’s causality test, we only observe a unidirectional relationship between long-term debt
to total assets and return on invested capital. That is, return on invested capital is largely influenced
by debt positively. Granger causality test indicates a uni-directional relationship between long-term

184
JRFM 2020, 13, 214

debt and return on invested capital. This relationship is positive, as evidenced by the coefficient of
long-term debt above.

Table 6. Return on invested capital (RoIC).

Dependent Variable Model 1 Model 2

FEM REM FEM REM
−7.6385 ** −7.509 **
Leverage
(−2.17) (−2.42)
2.1416 ** 1.6844 ***
LTD_TA
(2.22) (1.83)
−3.5081 * −2.4426 ** −3.2105 * −2.2261 **
Depreciation tax shield
(−3.16) (−2.45) (−2.9) (−2.23)
−2.4543 ** −3.152 * −2.4969 ** −3.0967 *
Tangibility
(−2.40) (−3.41) (−2.43) (−3.33)
−0.4048 0.09087 3.8359 ** 4.0090 *
Liquidity
(−0.21) (0.05) (2.53) (2.86)
0.2143 ** 0.2100 * 0.17597 ** 0.1623 *
Operating revenue
(2.58) (4.09) (2.09) (3.15)
0.05104 0.0623 0.06777 0.0803
Revenue growth (3-year)
(0.63) (0.78) (0.84) (1.00)
8.9454 *** 8.1662 *** −3.893121 −3.6585 ***
Constant
(1.66) (1.83) (−1.52) (−1.70)
4.52 4.56
F-test
(0.0000) (0.000)
13.67 16.76
Hausman test
(0.00) (0.0102)
1149.89 1166.45
Breusch-Pagan test
(0.00) (0.0000)
*, ** and *** are used for 1%, 5% and 10% level of significance.

In Table 7 below, in Model 1, the random effect regression identified leverage and tangibility as two
important explanatory variables of return on equity. The fixed effect model identified tangibility, size and
liquidity as significant explanatory variables at the 5% level of significance. The Granger causality
test indicates a unidirectional relationship between leverage and return on equity. This relationship is
positive in the REM regression of Model 1 above.
In model 2 (long-term debt as a second measure of debt level), long-term debt to total assets,
tangibility, liquidity and operating revenues were identified as significant explanatory variables of
return on equity at the 1% level of significance. In the random effects model, long-term debt, tangibility
and operating revenues were identified at the 5%, 1% and 10% levels of significance, respectively.
The Granger causality test suggests a bi-directional relationship between long-term debt and return on
equity. The positive coefficient of long-term debt to total assets in the equations in Model 2 in Table 7
indicates long-term debt to finance the purchase of assets for operations is beneficial to Australian
service sector firms.

185
JRFM 2020, 13, 214

Table 7. Panel regression output for Return on Equity (ROE).

Dependent Variable Model 1 Model 2

REM FEM FEM REM
3.8214 * 3.7031
Leverage
(2.76) (1.41)
3.3723 * 1.2797 **
LTD_TA
(4.73) (2.45)
0.1380 −0.9583 −0.8636 0.2167
Depreciation tax shield
(0.29) (−1.15) (−1.05) (0.45)
−1.3156 * −1.5104 ** −2.2271 * −1.7410 *
Tangibility
(−3.01) (1.97) (−2.93) (3.89)
1.1551 3.1308 ** 3.7468 * 0.9349
Liquidity
(1.54) (2.15) (3.34) (1.28)
0.02765 −0.1364 ** −0.1853 * 0.0335 ***
Operating revenue
(1.56) (−2.20) (−2.97) (1.91)
−0.0013 0.03421 0.03896 −0.0113
Revenue growth (3-year)
(−0.02) (0.56) (0.65) (−0.20)
−4.1381 ** −1.596 −0.14843 −1.2106
Constant
(−2.17) (−0.40) (−0.08) (−1.08)
3.44 6.90
F-test
(0.0023) (0.0000)
17.72 38.67
Hausman test
(0.0000) (0.0000)
0.62 2.88
Breusch-Pagan test
(0.2150) (0.0449)
*, ** and *** are measured for 1%, 5% and 10% level of significance.

5. Discussion
The relationship between capital structure and firm performance can be summarized in two
different ways: leverage and firm performance, and long-term debt and firm performance. These themes
are discussed in the following section.
In the Table 8 above, leverage is significantly associated with operating margin but has significant
negative association with two measures of firm performance: return on assets and return on invested
capital. In Table 9, other control variables have influenced profitability in positive and negative ways.
Tangibility has affected operating margin, return on vested capital and return on equity negatively,
suggesting that Australian firms are overinvesting on fixed assets. Revenue growth has also affected
operating margin and return on assets significantly. Depreciation tax shield is observed to have affected
operating margin negatively and liquidity has affected return on equity positively. The constant is
significant in both operating margin and return on invested capital, suggesting a guaranteed minimum
return from the presence of service sector firms in the economy. However, return on assets and return
on equity, not assumed as constants, are not significant at any level of confidence.
In the table below, all relevant regression models are summarized to demonstrate the effect
of long-term debt being used to finance total assets in order to improve firms’ performance in the
Australian services sectors.

186
JRFM 2020, 13, 214

Table 8. Leverage and its effect of firm performance.

Variables Performance Measures Used

Operating Return on Return on Return on
Margin Assets Invested Capital Equity
REM REM REM FEM
0.2411 ** −1.1143 * −7.509 ** 3.7031
Leverage
(−2.23) (−3.75) (−2.42) (−1.41)
1265 0.01242 −2.4426 ** −0.9583
Depreciation tax shield
(−0.36) (−0.13) (−2.45) (−1.15)
−0.0584 *** −0.01436 −3.152 * −1.5104 **
Tangibility
(−1.80) (−0.16) (−3.41) (−1.97)
0.0191 0.2029 0.09087 3.1308 **
Liquidity
(−0.32) (−1.24) (−0.05) (−2.15)
0.0046 * 0.02992 * 0.2100 * −0.1364 **
Operating revenue
(−2.84) (−6.61) (−4.09) (−2.20)
Revenue growth −0.0074 ** 0.0168 ** 0.0623 0.03421
(3-year) (−2.45) (−2.06) (−0.78) −0.56
0.5986 * 0.321 8.1662 *** −1.596
contant
(3.92) (−0.76) (−1.83) (−0.40)
3.44
F-test
(−0.0023)
24.85 26.13 35.01
Hausman test
(0.0000) (−0.0002) (0.000)
445.16 544.07 13.67
Breusch-Pagan test
(0.0000) (0.0000) (0.000)
*, ** and *** are measured for 1%, 5% and 10% level of significance.

In the Table 9 above, long-term debt to finance assets is significantly associated (positively) with
all measures of firm performance, except operating margin. Depreciation tax shield significantly
influenced return on invested capital negatively while tangibility has negatively affected, at different
levels of statistical significance, all measures of firm performance except return on assets. Liquidity has
affected firm performance in all instances, natively in operating profit, and positively (at different levels
of statistical significance) in improving return on assets and return on invested capital. Operating
revenue positively influenced all measures of firm performance except return on assets. Finally,
revenue growth for three years was a significantly influential negative factor in affecting operating
margin but a positive factor in improving return on assets.
In light of the discussions above, we can say, ‘capital structure matters.’ It enhances the performance
of the service sectors in Australia not only through improved operating margin and higher return on
invested capital; it also increases shareholder value by improving return on equity and return on assets.
As observed in our data analysis, we have used four different measures of firm performance. Three of
these measures relate to balance sheets and the other one relates to profit and loss in the short-term.
Leverage in all measures of performance was significant at the 1% level except when return on invested
capital was used to measure firm performance. Even in the presence of other variables (used in the
literature as explanatory variables) that showed significant influence in firm performance, leverage
remains significant in shaping the performance of service sector firms in Australia. The positive and
significant relationship between leverage and operating margin implies that the service sectors in
Australia can greatly benefit by increasing the debt level in its capital structure. The negative and
significant relationship with tangibility also makes economic sense and implies that tying funds in
fixed assets can be detrimental to operating profits as the company will have lees funds available for
generating revenues.

187
JRFM 2020, 13, 214

Table 9. Long-term debt and its effect on firm performance.

Variables Performance Measures Used

Operating Return on Return on Return on
Margin Assets Invested Capital Equity
REM REM REM REM
−0.0295 0.2475 ** 1.6844 *** 1.2797 **
LTD_TA
(−0.88) (−2.49) (−1.83) (−2.45)
−0.0179 −0.1288 −2.2261 ** 0.2167
Depreciation tax shield
(−0.51) (−1.13) (−2.23) (−0.45)
−0.0648 ** 0.081 −3.0967 * −1.7410 *
Tangibility
(−1.97) (−0.77) (−3.33) (−3.89)
−0.0893 *** 0.9096 * 4.0090 * 0.9349
Liquidity
(−1.78) (−5.82) (−2.86) (−1.28)
0.0059 * 0.012 0.1623 * 0.0335 ***
Operating revenue
(−3.68) (−1.39) (−3.15) (−1.91)
Revenue growth −0.0080 * 0.0193 ** 0.0803 −0.0113
(3-year) (−2.65) (−2.32) (−1.000) (−0.20)
0.9457 * −1.2744 * −3.6585 *** −1.2106
Constant
(−12.37) (−4.82) (−1.70) (−1.08)
3.800 21.04 16.76 38.67
Hausman test
(−0.7033) (−0.0016) (−0.0102) (0.000)
480.54 532.38 1166.45 2.88
Breusch-Pagan test
(0.0000) (0.0000) (0.0000) (−0.0449)
*, ** and *** are measured for 1%, 5% and 10% level of significance.

Long-term debt to total assets does not play any role in operating margin as the service sectors
can generate frequent cash flows and turnover rate is very high. As such, need for long-term debt is
irrelevant. The economic impact is different for return on assets where our analysis shows that both
leverage and long-term debt to total assets are positively significant in impacting return on assets under
both REM and FEM. It further strengthens our earlier arguments that service sectors greatly benefit
from increased debt level in its capital structure. Revenue growth for three years also shows promising
impact on return on asset which also remain positively significant. When we tested our model for ROIC,
we find that leverage is negatively significant whereas long-term debt to total asset remains positively
significant. The negative relationship with leverage implies that firms face challenges in return on
invested capital if too much of the funds are tied with short-term borrowing as they are payable quickly.
In addition, just like in operating margin, tangibility remains negatively significant implying that firms
are adversely affected by increased tangible assets holding. Finally, when analyzing the relationship
with return on equity, just like ROA, we observe that both leverage and long-term debt to total assets
remain positively significant. Unlike ROA, tangibility remains negatively significant with ROE. In
conclusion, we can assert that in the case of service sector firms in Australia, a high level of leverage
and a high level of long-term debt in capital structure is beneficial to increasing shareholders’ wealth.

6. Conclusions
This paper has examined firm level characteristics and firm performance (or profitability) of
service sector firms listed in the Australian Stock Exchange (ASX). Using a panel regression approach
on data collected over an eleven-year period (2009–2019), the effect of capital structure and leverage was
examined. Four measures of firm performance were used: return on assets, return on equity, operating
margin ratio and return on capital employed. The analysis of data reveals a significant association
between return on equity and leverage levels. Leverage affects firm performance at a statistically

188
JRFM 2020, 13, 214

significant level in these service sector firms. For every dollar of increase in leverage, operating margin
improves by 0.24 times, return on assets reduces by 1.11 times and return on invested capital reduces
by 7.59 times (all statistically significant at the 1% and 5% level), suggesting that Australian services
sector firms are not benefitting much from the use of debts to finance their operations. This finding
is in sharp contrast to asymmetric information theory that suggests that lower debt levels hide firm
performance (Myers 1984). In fact, they are overburdened with debts. When long-term debt is used to
finance total assets, the picture changes dramatically. Return on assets, return on invested capital and
return on equity changes by 0.24 times (significant at 5% level), return on capital employed increases by
1.68 times (significant at 10% level) and return on equity improves by 1.27 times (significant at the 5%
level), suggesting the positive value adding contributions of the use of long-term debt. The directional
causality tests, as captured in the Granger causality test, indicated a positive unidirectional association
between leverage and operating margin, bi-directional causality with return on assets (negative) and
return on invested capital, with return on assets (negative) and a unidirectional (positive) causality
between leverage and return on equity. The test also identified a bidirectional causality between
long-term debt to total assets and operating margin, and a bidirectional relationship with return on
assets, return on invested capital and return on equity. The presence of unidirectional and bidirectional
causality between different types of debts to finance operations mean significant interdependencies
and negative effects of debt on service sector firms in Australia.
The study has the inherent limitations of any research project. The sample size may be questioned
for two reasons: the number of firms from the Australian service sector and the years included in the
data. Due to unavailability of data, only three years of data are used. Inclusion of more years can
be a possibility for the extension of the current research project. Interested researchers may consider
a robust dataset, encompassing industries from all sectors of the Australian economy. The current
study has examined a limited number of constructs to reflect on the profitability of Australian service
sector listed firms. The influence of extra-organizational factors may contribute to the profitability of
Australian service sector companies. Researchers willing to pursue the line of inquiry in this paper may
include economic factors such as inflation, interest rate and GDP in future research. Finally, service
sector heterogeneity may be partially responsible for poor reflection of profitability. So the inclusion of
industry effects may be worthwhile before a conclusion can be reached about the industry sector effect
on Australian service sector performance.

Author Contributions: R.A. developed the research idea and conducted review of literature. R.A. collected the
data under the guidance and recommendation of R.B. and run the statistical analysis which is also suggested by
R.B. R.B. provided the insights of the statistical results and analysis. Both authors have read and agreed to the
published version of the manuscript.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflict of interest.

References
Abor, Joshua. 2007. Corporate governance and financing decisions of Ghanaian listed firms. Corporate Governance:
The International Journal of Business in Society 7: 83–92. [CrossRef]
Ahmed Sheikh, Nadeem, and Zongjun Wang. 2011. Determinants of capital structure: An empirical study of
firms in manufacturing industry of Pakistan. Managerial Finance 37: 117–33. [CrossRef]
Arifin, Agus Zainul. 2017. Interactions between capital structure and profitability: Evidence from Indonesia stock
exchange. International Journal of Economic Perspectives 11: 117–21.
Australian Bureau of Statistics. 2019. Australian System of National Accounts, 2018–19 5204.0. Canberra: Australian
Bureau of Statistics.
Australian Bureau of Statistics. 2020. Labour Force, Australia, Detailed, Quarterly, May 2020. (6291.0.55.003).
Canberra: Australian Bureau of Statistics.

189
JRFM 2020, 13, 214

Barth, James Richard, Gerard Caprio Jr., and Ross Levine. 2001. Banking Systems around the Globe: Do Regulation
and Ownership Affect Performance and Stability? In Prudential Supervision: What Works and What Doesn’t.
Chicago: University of Chicago Press, pp. 31–96.
Baxter, Nevins D. 1967. Leverage, risk of ruin and the cost of capital. The Journal of Finance 22: 395–403.
Canarella, Giorgio, and Stephen Michael Miller. 2018. The determinants of growth in the U.S. information and
communication technology (ICT) industry: A firm-level analysis. Economic Modelling 70: 259–71. [CrossRef]
Chadha, Saurabh, and Anil Kumar Sharma. 2015. Capital Structure and Firm Performance: Empirical Evidence
from India. Vision 19: 295–302. [CrossRef]
Chakrabarti, Anindita, and Ahindra Chakrabarti. 2019. The capital structure puzzle—Evidence from Indian
energy sector. International Journal of Energy Sector Management 13: 2–23. [CrossRef]
Choi, Serin, Seoki Lee, Kyuwan Choi, and Kyung-A Sun. 2018. Investment–cash flow sensitivities of restaurant
firms: A moderating role of franchising. Tourism Economics 24: 560–75. [CrossRef]
Dalci, Ilhan. 2018. Impact of financial leverage on profitability of listed manufacturing firms in China.
Pacific Accounting Review 30: 410–32. [CrossRef]
Dave, Ashvin, Ashwin Parwani, Ashish Joshi, and Tejas Dave. 2019. A study of capital structure and profitability
of Indian steel sector companies. International Journal of Advanced Science and Technology 28: 866–73.
DeAngelo, Harry, and Ronald W. Masulis. 1980. Optimal capital structure under corporate and personal taxation.
Journal of Financial Economics 8: 3–29. [CrossRef]
Donaldson, Gordon. 1984. Managing Corporate Wealth. New York: Praeger.
El-Sayed Ebaid, Ibrahim. 2009. The impact of capital-structure choice on firm performance: Empirical evidence
from Egypt. The Journal of Risk Finance 10: 477–87. [CrossRef]
Fama, Eugene Frank, and Kenneth Richard French. 1998. Value versus growth: The international evidence.
Journal of Finance 53: 1975–99. [CrossRef]
Fitim, Deari, Matsuk Zoriana, and Lakshina Valeriya. 2019. Leverage and Macroeconomic Determinants: Evidence
from Ukraine. Studies in Business and Economics 14: 5–19. [CrossRef]
Gharaibeh, Omar Krishna, and Marie Hal Bani Khaled. 2020. Determinants of profitability in Jordanian services
companies. Investment Management and Financial Innovations 17: 277–90. [CrossRef]
Ghasemi, Maziar, Nazrul Hisyam Ab Razak, and Komeli Dehghani. 2018. Determinants of debt structure in ACE
Market Bursa Malaysia: A panel data analysis. Journal of Social Sciences Research 2018: 390–95. [CrossRef]
Ghosh, Chinmoy, Raja Nag, and C. F. Sirmans. 2000. The pricing of seasoned equity offerings: Evidence from
REITs. Real Estate Economics 28: 363–84. [CrossRef]
Gleason, Kimberley Clark, Linette Knowles Mathur, and Ike Mathur. 2000. The interrelationship between culture,
capital structure, and performance: Evidence from European retailers. Journal of Business Research 50: 185–91.
[CrossRef]
Goddard, John, Manouche Tavakoli, and John Oliver Wilson. 2005. Determinants of profitability in European
manufacturing and services: Evidence from a dynamic panel model. Applied Financial Economics 15: 1269–82.
[CrossRef]
Grossman, Sanford John, and Oliver David Hart. 1982. Corporate financial structure and managerial incentives.
In The Economics of Information and Uncertainty. Chicago: University of Chicago Press, pp. 107–40.
Hadlock, Charles Jones, and Christopher M. James. 2002. Do banks provide financial slack? The Journal of Finance
57: 1383–419. [CrossRef]
Helmy, Muhammad Harith Zulqarnain Bin Noor, Goh Chin Fei, Tan Owee Kowang, Ong Choon Hee, Tan Seng Teck,
Lim Kim Yew, and Wong Chee Hoo. 2020. Capital structure, internal governance mechanisms and firm
performance. International Journal of Psychosocial Rehabilitation 24: 7313–21. [CrossRef]
Hussein, Mohammed Jalal, Alrabba Hussein, Muhannad Akram Ahmad, and Mashhoor Hamadneh. 2019. Capital
structure and firm performance: Evidence from Jordanian listed companies. International Journal of Scientific
and Technology Research 8: 364–75.
Jensen, Michael Clark. 1986. Agency cost of free cash flow, corporate finance, and takeovers. Corporate Finance,
and Takeovers. American Economic Review 76: 323–29.
Jensen, Michael Clark, and William H. Meckling. 1976. Theory of the firm: Managerial behavior, agency costs and
ownership structure. Journal of Financial Economics 3: 305–60. [CrossRef]
Khan, Tasneem, Mohd Shamim, and Jatin Goyal. 2018. Panel data analysis of profitability determinants: Evidence
from Indian telecom companies. Theoretical Economics Letters 8: 3581–93. [CrossRef]

190
JRFM 2020, 13, 214

Köksal, Bülent, and Cüneyt Orman. 2015. Determinants of capital structure: Evidence from a major developing
economy. Small Business Economics 44: 255–82. [CrossRef]
Kraus, Alan, and Robert H. Litzenberger. 1973. A state-preference model of optimal financial leverage. The Journal
of Finance 28: 911–22. [CrossRef]
Levin, Andrew, Chien-Fu Lin, and Chia-Shang James Chu. 2002. Unit root tests in panel data: Asymptotic and
finite-sample properties. Journal of Econometrics 108: 1–24. [CrossRef]
Li, Hui, and Petros Stathis. 2017. Determinants of Capital Structure in Australia: An Analysis of Important Factors.
Managerial Finance 43: 881–97. [CrossRef]
M’ng, Jacinta Chan Phooi, Mahfuzur Rahman, and Selvam Sannacy. 2017. The determinants of capital structure:
Evidence from public listed companies in Malaysia, Singapore and Thailand. Cogent Economics and Finance
5: 1418609. [CrossRef]
Macan Bhaird, Ciarán, and Brian Lucey. 2010. Determinants of capital structure in Irish SMEs. Small Business
Economics 35: 357–75. [CrossRef]
Mayuri, Tulsi, and Lingesiya Kengatharan. 2019. Determinants of Capital Structure: Evidence from Listed
Manufacturing Companies in Sri Lanka. SCMS Journal of Indian Management 16: 43–56.
Modigliani, Franco, and Merton H. Miller. 1958. The cost of capital, corporation finance and the theory of
investment. The American Economic Review 48: 261–97.
Modigliani, Franco, and Merton H. Miller. 1963. Corporate income taxes and the cost of capital: A correction.
The American Economic Review 53: 433–43.
Musah, Mohammed, and Yusheng Kong. 2019. The relationship between capital structure and the financial
performance of non-financial firms listed on the Ghana Stock Exchange (GSE). International Journal of Research
in Social Sciences 9: 92–123.
Myers, Stewart Charles. 1984. The capital structure puzzle. The Journal of Finance 39: 574–92. [CrossRef]
Myers, Stewart C., and Nicholas Samual Majluf. 1984. Corporate financing and investment decisions when firms
have information that investors do not have. Journal of Financial Economics 13: 187–221. [CrossRef]
Negasa, Tran. 2016. The Effect of Capital Structure on Firms’ Profitability (Evidenced from Ethiopian). Preprints,
2016070013. [CrossRef]
Nunes, Paulo J. Maçãs, Zélia M. Serrasqueiro, and Tiago N. Sequeira. 2009. Profitability in Portuguese service
industries: A panel data approach. The Service Industries Journal 29: 693–707. [CrossRef]
Park, Kwangmin, and SooCheong Jang. 2018. Is franchising an additional financing source for franchisors?
A Blinder–Oaxaca decomposition analysis. Tourism Economics 24: 541–59. [CrossRef]
Pettit, Richardson, and Ronald F. Singer. 1985. Small business finance: A research agenda. Financial Management
14: 47–60. [CrossRef]
Qiu, Mei, and Bo La. 2010. Firm characteristics as determinants of capital structures in Australia. International
journal of the Economics of Business 17: 277–87. [CrossRef]
Rashid, Kashif, and Sardar Islam. 2009. Capital structure and firm performance in the developed financial market.
Corporate Ownership and Control 7: 189–201. [CrossRef]
Roden, Dianne M., and Wilbur Grand Lewellen. 1995. Corporate Capital Structure Decisions: Evidence from
Leveraged Buyouts. Financial Management 24: 76–87. [CrossRef]
Ross, Stephen Allen. 1977. The determination of financial structure: The incentive-signalling approach. The Bell
Journal of Economics 8: 23–40. [CrossRef]
Salim, Mahfuzah, and Raj Yadav. 2012. Capital Structure and Firm Performance: Evidence from Malaysian Listed
Companies. Procedia-Social and Behavioral Sciences 65: 156–66. [CrossRef]
Sardo, Filipe, Zélia Serrasqueiro, and Elisabete GS Félix. 2020. Does Venture Capital affect capital structure
rebalancing? The case of small knowledge-intensive service firms. Structural Change and Economic Dynamics
53: 170–79. [CrossRef]
Sermpinis, Georgios, Serafeim Tsoukas, and Ping Zhang. 2019. What influences a bank’s decision to go public?
International Journal of Finance and Economics 24: 1464–85. [CrossRef]
Shalini, Ramaswamy, and Mahua Biswas. 2019. Capital structure determinants of SandP BSE 500: A panel data
research. International Journal of Recent Technology and Engineering 8: 377–80. [CrossRef]
Sivalingam, Logavathani, and Lingesiya Kengatharan. 2018. Capital Structure and Financial Performance: A Study
on Commercial Banks in Sri Lanka. Asian Economic and Financial Review 8: 586–98. [CrossRef]

191
JRFM 2020, 13, 214

Stulz, René. 1990. Managerial discretion and optimal financing policies. Journal of Financial Economics 26: 3–27.
[CrossRef]
Suntraruk, Phassawan, and Xiaoxing Liu. 2017. The impacts of institutional characteristics on capital structure:
Evidence from listed commercial banks in China. Afro-Asian Journal of Finance and Accounting 7: 337–50.
[CrossRef]
Szemán, Judith. 2017. Relevance of Capital Structure Theories in the Service Sector. Theory, Methodology, Practice
13: 53–64. [CrossRef]
Taub, Allan Jay. 1975. Determinants of the firm’s capital structure. The Review of Economics and Statistics 57: 410–16.
[CrossRef]
Weill, Laurent. 2008. Leverage and corporate performance: Does institutional environment matter? Small Business
Economics 30: 251–65. [CrossRef]
Yazdanfar, Darush. 2012. The Impact of Financing Pattern on Firm Growth: Evidence from Swedish Micro Firms.
International Business Research 5: 16. [CrossRef]
Yazdanfar, Darush, and Peter Öhman. 2015. Debt financing and firm performance: An empirical study based on
Swedish data. The Journal of Risk Finance 16: 102–18. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

192
 Journal of
Risk and Financial
Management

Project Report
Examination and Modification of Multi-Factor Model
in Explaining Stock Excess Return with Hybrid
Approach in Empirical Study of Chinese Stock Market
Jian Huang * and Huazhang Liu
Division of Business Management, Beijing Normal University-HongKong Baptist University United
International College, Zhuhai 519087, China; spot_light@[Link] or k530002087@[Link]
* Correspondence: k530002046@[Link] or jianhuang.951111@[Link]; Tel.: +86-186-6454-9728

Received: 28 April 2019; Accepted: 22 May 2019; Published: 25 May 2019

Abstract: To search significant variables which can illustrate the abnormal return of stock price, this
research is generally based on the Fama-French five-factor model to develop a multi-factor model.
We evaluated the existing factors in the empirical study of Chinese stock market and examined
for new factors to extend the model by OLS and ridge regression model. With data from 2007 to
2018, the regression analysis was conducted on 1097 stocks separately in the market with computer
simulation based on Python. Moreover, we conducted research on factor cyclical pattern via chi-square
test and developed a corresponding trading strategy with trend analysis. For the results, we found
that except market risk premium, each industry corresponds differently to the rest of six risk factors.
The factor cyclical pattern can be used to predict the direction of seven risk factors and a simple moving
average approach based on the relationships between risk factors and each industry was conducted
in back-test which suggested that SMB (size premium), CMA (investment growth premium), CRMHL
(momentum premium), and AMLH (asset turnover premium) can gain positive return.

Keywords: multi-factor model; risk factors; OLS and ridge regression model; python; chi-square test

1. Introduction
Financial markets are rife with uncertainties which make it difficult to forecast future market trends.
Especially in the stock market, while providing investors with remarkable return, at the same time,
it entails tremendous risk. As people pursue higher returns, they must bear the corresponding risks.
A pricing model with multiple factors is a promising approach to predict future stock prices. If the
relationship between the risk and return can be expressed by multiple factors in a mathematical model,
the standards of stock selection and corresponding investment strategies can be established. To be
specific, each investor has their own risk preference, for instance, risk averse investors tend to bear
lower risk and receive lower return. This specific type of investors has their preference in stock selection.
In order to find suitable selection criteria, we need to quantify the relationship of risk premium factor
and excess return. Investors can refer to the selection criteria to establish a corresponding trading
strategy which can achieve their target excess return.
In the process of model examination, regression was conducted on single stocks which differed
from the previous research which used portfolios. Especially, ridge regression was conducted instead of
OLS regression. As for the process of model modification, two new factors were added to achieve higher
explanatory power. Furthermore, we discussed the endogeneity and exogeneity for the risk premium
factor. On the basis of economic objectives, we established a trading strategy for Chinese stock market.
Although previous research has worked well in the American stock market, these findings may be
less practically applied to the Chinese market due the investor component. Since individual investors
contribute nearly 80% of the trading volume, investment behavior and preference can largely impact

JRFM 2019, 12, 91; doi:10.3390/jrfm12020091 193 [Link]/journal/jrfm

JRFM 2019, 12, 91

the market average return. However, the asymmetric information and investment concepts may lead
to irrational behaviors. Therefore, we use risk premium factors to explain the excess return and the
coefficients to measure the sensitivity of investor reactions to the risk premium.
To begin, we conducted single stock regression to examine the effectiveness of a five-factor model
in Chinese stock market. Then we compare the coefficients of different factors under specific company
types to discover the leading factor. We use the t-statistic to evaluate the significance of each factor.
Regarding previous articles and research, we will add new factors in the model for better performance.
Moreover, we conducted inter-factor cyclical research to study the pattern of factors. We can predict
the rise and fall of the coefficient on a quarterly basis.

2. Literature Review
The multi-factor risk model has undergone a series of developments which can be divided into
four major steps. At the beginning, in order to figure out the leading factors for security return, several
researchers have contributed to the development of an asset pricing model. Initially, Markowitz (1952)
proposed a mean-variance model to illustrate the statistical relationship between security risk and
return in terms of standard deviation and expected rerun. It has established the foundation of modern
finance theory. However, he had not specified the factors that explain expected security returns.
Based on the modern portfolio theory, the capital asset pricing model (CAPM) was developed
by Sharpe (1964) and Lintner (1965) to illustrate the linear relationship between expected return and
market risk premium.
Ri = Rf + βmarket (Rm − Rf) (1)

With the empirical tests in the stock market, they managed to find out the pattern of stock returns
in line with the general stock market index. To be specific, the model estimates the relationships
between the return on market index (the explanatory variable) and the return on the stock (the
dependent variable). The regression coefficient of the single index model is referred to as beta which
is a measure of the sensitivity of a stock to general movement in the market index. This empirical
research symbolized the transition from qualitative analysis to quantitative analysis which also laid
the foundation for the subsequent asset pricing models.
Since then, the CAPM model has been widely applied in research and empirical testing. However,
with increasing abnormal returns which cannot be explained by existing factors, the market beta
was no longer sufficient to describe expected return (Fama 1996). Therefore, Fama and French
proposed a multifactor model consisting of three factors for market risk (Rm − Rf), market value,
and book-to-market ratio (Fama and French 1993). In this model, Ri, Rf, and Rm stands for security
expected return, risk free rate, and market return. SMB and HML are the risk premium factors. They
illustrated that small stocks can generate higher returns than large stocks while value stocks can
generate higher returns than growth stocks.
 
Ri − R f = a + βmarket Rm − R f + βsize SMB + βBM HML (2)

It is also a supplement of arbitrage pricing theory (Ross 1976) which emphasizes that the
expected return is not only affected by market risk but also a series of other factors. The generalized
model (Bodie et al. 2014) illustrated the linear relationship of expected return and different factors
(Bodie et al. 2017). In Equation (3), F j represents the factors and b represents the coefficients.
k
ri = ai + bij F j + i, i = 1, 2 . . . , N (3)
j=1

Moreover, Banz (1981) illustrated that return of securities are affected by several index including
B/M ratio and E/P ratio which represent a series of risk premium. These articles largely contribute to
the development of the multi-factor model.

194
JRFM 2019, 12, 91

In latter decades, the three-factor model was faced with a series of challenges. As the three-factor
model was applied to stock trading and empirical testing, it appeared that some of the phenomenon
cannot be explained by the model which can lead to unpredictable abnormal return. Therefore, more
specific factors need to be added into the model to improve the accuracy. Novy-marx (2013) proved
that profitability, measured by gross profits-to-assets, has roughly the same power as book-to-market
value in explaining the average return. The following equation is based on the dividend discount
model, Yt+τ represents the earning for period t + τ, dBt+τ is the change in book equity, r represents
the expected return. It implied that higher market value leads to lower expected return, while higher
earnings imply higher expected return.


∞
Mt = E(Yt+τ − dBt+τ )/(1 + r)τ (4)
τ=1

Aharoni et al. (2013) recorded an insignificant but statistically reliable relationship between
investment pattern and average return. Other evidence also illustrated that the profitability and
investment factors can explain some of the variation in average return.
Carhart (1997) conducted a study on the common factors in stock return and investment expense
by adding momentum factor in three-factor model. It measures the tendency of price changes with
a portfolio of long previous-12-month return winners and short previous-12-month loser stocks,
which had an 8% accumulated yield. The momentum factor can explain 6.4% of the excess return.
Therefore, Fama and French (2015) added the two new factors, investment (CMA) and profitability
(RMW), to build the five-factor model for measuring the effects of company size, valuation, profitability,
and investment pattern in average stock returns. In this equation, they divide both sides by book
value at time t to create book-to-market ratio and present the relationship between return r and
valuation factor.
 
Ri − R f = a + βmarket Rm − R f + βsize SMB + βBM HML + βpro f itability RMW + βinvestment CMA + ε (5)

Moreover, they use the RMW factor to represent the profitability and the CMA factor to represent
investment. According to empirical study, the five-factor model has a higher effectiveness than the
three-factor model, which can explain 71–94% of the variation of average return. In the test, researchers
divided all the stocks with three sets of factors. The result also implied that book-to-market factor is
redundant for describing average return using the American stock data from 1963 to 2013. Also, the
model fails to capture the low average return for small firms with low profitability and high investment.
Sehgal and Vasishth (2015) tested the model in various emerging markets and discovered that the
change of price and trading volume are partly risk based and partly behavioral. The research indicated
that behavioral factors are necessary in the study of the multi-factor model. Except for the existing risk
premium factors in Fama and French model, Peng et al. (2014) studied the effect of different investment
sentiments on the market return from which they found out customer satisfaction is a significant factor
for abnormal return. Moreover, a human capital factor was considered in terms of compensation level
(Moinak and Balakrishnan 2018). Zahedi and Rounaghi (2015) applied an artificial neural network to
assess the components of a multi-factor model and predict future stock prices.
Considering the features of data and the time-varying factors, a range of studies applied various
methods to address problems via multi-factor models. Akter and Nobi (2018) examined the distribution
and frequency distribution for both daily stock returns and volatility. Furthermore, Chen and
Kawaguchi (2018) distinguish two significant regimes (a persistent bear market and a bull market)
to examine market time-varying risk factors to achieve Markov regime-switching. These studies
examined the model specifically in certain periods which enables the researchers to compare the model
performances in different situations which can improve the applicability of the model.
Some researchers conducted modification on the basic model with specific structure. Ronzani et al. (2017)
suggested that β (systematic risk) evolves over time and the model with time-varying β provide less

195
JRFM 2019, 12, 91

conservative VaR measures than the static β. While Cisse et al. (2019) examined the dynamics of the model
with Kalman filter and Markov switching (MS) model and proved that the former method fits better in
the model. Bhattacharjee and Roy (2019) proposed a social network dependence structure to address such
misspecifications. For the investment aspect, Frazzini et al. (2013) suggesting that Buffett’s returns are more
due to stock selection than to his effect on management.

3. Methodology

3.1. Hypotheses

Hypothesis 1. Market premium has a positive relationship with excess return.

Hypothesis 2. Size premium has a positive relationship with excess return.

Hypothesis 3. Book-to-market premium has a positive relationship with excess return.

Hypothesis 4. Profitability premium has a positive relationship with excess return.

Hypothesis 5. Investment growth premium has a positive relationship with excess return.

Hypothesis 6. Momentum premium has positive a relationship with excess return.

Hypothesis 7. Asset turnover premium has a positive relationship with excess return.

Hypotheses are conducted on the relationship between dependent variable and independent
variables. To be specific, the positive relationship stated that risk premium factor can generate higher
excess return. For instance, a company with higher profitability can achieve higher returns than a low
profitability company. Since previous researchers have not examined the effect of factors on single
stocks, we decided to retest the effect of seven factors on excess return for each security and summary
by industries.

3.2. Research Design

The data were collected from the Choice, Tongdaxin, and Resset database. Choice and Tongdaxin
can provide historical values of stock financial ratios and Resset provides the basic information of
companies and interest rates. The samples are 1097 stocks in the Chinese stock market, including
Shanghai securities exchange market and Shenzhen securities exchange market from, 2007 to 2018
because large quantity of firms transformed their non-tradable shares into tradable shares in the process
of Chinese Reformation of National Owned Stock from 2005 to 2006 and the time range after 2007
provides the research with more companies’ data. The type of data is quarterly data which includes 47
quarterly data points for each stock. In the process of data cleaning, those stocks with missing data
will be deleted in the samples.
Even if Fama and French (2015) had delivered the five-factor model, yet they only consider the
factor performance in different groups of portfolio but not the factors’ effect on each security in the
stock market. Thus, by applying variable-intercept models mentioned in the book of Hsiao (2003),
we ran the panel data for each stock to test the effect of these five premiums in the Model 1. The study
bridged the relationships between the excess return of each security and different premium factors
via OLS (ordinary least squares regression), which means each stock has their individual coefficient
and there are more than 1000 regression functions (Johnson and Wichern 2008). During the research
process, this paper classified the securities into different groups to calculate risk premium factors
based on market value, book-to-market ratio, ROE, and the growth of investment. Then, this research
considered the average coefficients’ level, average significant level, and their distribution.

196
JRFM 2019, 12, 91

In Model 2, the momentum factor and turnover factor are added into the model. We developed an
innovative method to study each single stock in the market, in order to find out the general pattern of
the stock market. In contrast to the former ways that examine the model with the diversified portfolio,
we conducted ridge regression on single stocks and divided the stocks in industries. The 28-industry
classification standards are from ShenWan industry index.
Figure 1 illustrates the research structure and their corresponding functions for the results.
The research includes two major parts, which are multi-factor examination with single stock regression
and time-series analysis for risk factors.

Figure 1. Multi-factor model examination with single stock regression and time-series analysis.

The first part involved stability test, OLS regression, ridge regression, and robustness test to find
the significant correlations between stock excess return and seven risk factors. In addition, a chi-square
test was conducted to examine whether the effect of factors various in different industries. Moreover,
by measuring the percentage of positive and negative correlations in each industry, the significant
relationships between factors and industry was discovered.
The second part of research covered chi-square test to find out the endogeneity and exogeneity (or
called pattern of fluctuation) for seven risk factors and a back test of trading strategy with trend analysis.
For details, for chi-square testing, we firstly recorded the rise and fall pattern in the neighboring
quarters. There were four kinds of pattern (rise after rise, rise after fall, fall after rise, fall after fall).
We calculated each patterns’ amount and conducted chi-square testing on the pattern of each factor.
If a factor passes the test, it can be inferred that the current pattern is in accordance with the previous
period. Therefore, with the endogeneity and exogeneity of the factors, we can predict the future rise
and fall based on the current pattern. Moreover, a trading strategy was established based on the trend
analysis of factors. As for the investment portfolio, stocks were selected according to the effect of
seven factors on industries. A simple back test was also conducted to examine the performance of the
trading strategy.
The result of general significant correlations and the factors’ effect in various industry can be
used in some economic phenomenon explanation. Furthermore, combing the significant relationships
between risk factors and corresponding industries with trend analysis, trading strategies can be built.

Factor Selection
We modified the multi-factor model to adapt for the Chinese stock market. Since some of the
existing factors in the model have poor performance in the regression analysis and empirical test, it is
significant to add or delete some of the factors to improve the model.
In accordance with previous studies, we selected two new factors (momentum factor and turnover
factor). Firstly, the momentum effect is remarkable in the medium term of stock market. Momentum

197
JRFM 2019, 12, 91

effect was proposed by Jegadeesh and Titman (1993) to illustrate the phenomenon that stock performed
well in the past is more likely to achieve higher return. Regarding the composition of investor and
investment behavior in stock market, we suggested that the momentum effect can explain the part
of the excess return. Secondly, we added a turnover factor into the model. Asset turnover rate is a
component of DuPont analysis, which is used to analyze the return of shareholder’s equity (Wild 2016).
Asset turnover rate is a vital measure of a company’s operation capacity. A higher asset turnover
rate indicates that a company has better operation capacity and higher efficiency. When conducting
value investing, this financial ratio is primarily regarded before investment, as it can directly reflect the
condition of a company, for example profitability, operation capacity, and so on.

3.3. Research Process

3.3.1. Multi-Factor Model Examination with Single Stock Regression Analysis

• Stability Test

DV: Ri-Rf
IV: Rm-Rf, SMB, RMW, HML, CMA, CRMHL, AMLH
CV: Time, Timeˆ2, Season

If the three control variables which refers to time, the square of time and season can pass the t-test
with high p-level, these three variables should be put into the regression models, otherwise they can
be deleted.

• OLS Regression

We conducted OLS regression on 47 quarters for each single stock, in order to examine the
multi-factor model. According to OLS, to calculate the beta (coefficient) of each independent variable,
the matrix operation combines X and y. In this formula, X is a 47 × 7 matrix, in which 47 is the 47
time-series of data and 7 represents seven proposed risk factors. The y is a 47 × 1 matrix and 1 indicates
the excess return of a single stock. Since there are 1097 stock, X matrix is fixed yet y matrix is the data
of different stocks.
β = (XT X)−1 XT y (6)
⎡ ⎤ ⎛ ⎡ ⎤
⎢⎢ β1 ⎥⎥ ⎜⎡ ⎤T ⎡ ⎤⎞−1 ⎡ ⎤T ⎢⎢ y1 ⎥⎥
⎢⎢
⎢⎢ β2
⎥⎥ ⎜⎜⎢⎢
⎥⎥ ⎜⎜⎢⎢
SMB1 RMW1 . . . AMLH1 ⎥⎥ ⎢⎢ SMB1 RMW1 . . . AMLH1 ⎥⎥⎟⎟⎟ ⎢⎢ SMB1 RMW1 . . . AMLH1 ⎥⎥ ⎢⎢⎢ ⎥⎥
⎥⎥
⎢⎢ ⎥⎥ ⎜⎜⎢⎢ ⎥⎥ ⎢⎢ ⎥⎥⎟⎟ ⎢⎢ ⎥⎥ ⎢⎢ y2 ⎥⎥
⎢⎢ . ⎥⎥ ⎜⎜⎢⎢ ... ... ... ⎥⎥ ⎢⎢ ... ... ... ⎥⎥⎟⎟ ⎢⎢ ... ... ... ⎥⎥ ⎢⎢ ⎥⎥
⎥⎥ ⎢⎢ ⎥⎥⎟⎟ ⎢⎢ ⎥⎥ ⎢ ...
⎢⎢ ..
⎢⎢
⎢⎢ .
⎥⎥ = ⎜⎜⎢⎢
⎥
⎥⎥⎥ ⎜⎜⎜⎜⎢⎢⎢⎢ .. .. .. ⎥⎥⎥ ⎢⎢⎢
⎥⎥ ⎢⎢ .. .. .. ⎥⎥⎥⎟⎟⎟ ⎢⎢⎢
⎥⎥⎟⎟ ⎢⎢ .. .. .. ⎥⎥⎥ ⎢⎢⎢⎢
⎥⎥⎥ ⎢⎢⎢
⎥⎥
⎥⎥
⎥⎥
(7)
⎢⎢ .
⎢⎢ .
⎥
⎥⎥⎥ ⎜⎜⎜⎝⎢⎢⎣
. . . ⎥⎥⎥ ⎢⎢⎢ . . . ⎥⎥⎥⎟⎟⎟ ⎢⎢⎢ . . . ⎥⎥ ⎢⎢ .. ⎥⎥
⎥⎥
⎢⎣ ⎥⎦ ⎦ ⎣ ⎦⎟⎠ ⎣ ⎦ ⎢⎢ . ⎥⎦
SMB47 RMW47 . . . AMLH47 SMB47 RMW47 . . . AMLH47 SMB47 RMW47 . . . AMLH47 ⎣
β7 y7

We used computer simulation in Python, which can help us automatically run regression 1097
times. The python program would print out the result of regression and calculate the number of mean
value of coefficients and t-value, which were collected in a 7 × 1097 matrix. We also drew the frequency
histogram for each factor to study the distribution of the coefficients. Based on these assumptions, the
majority of the coefficients should be positive. If most of them come out negative, we may infer that
the five-factor model is not suitable for the Chinese stock market.

• Ridge Regression

Since we added two more factors in the model, we applied ridge regression instead of OLS
regression. On one hand, ridge regression can prevent an over-fitting result from extra factors. On the
other, it can prevent multi-collinearity and increase the significance of the factors. Ridge regression is
developed based on OLS regression by adding regularization term λI. The regularization term can
discover the factor with collinearity and force the coefficient to approach zero to ensure that the effect
of multi-collinearity is minimized. To calculate the beta (coefficient) of each independent variable,

198
JRFM 2019, 12, 91

the matrix operation combines X, y, λ, and I. X and y are the same as the matrix in the OLS model. λ is
an optimization hyperparameter and I is a 7 × 7 identity matrix.

β = (XT X + λI)−1 XT y (8)

⎡ ⎤ ⎛ ⎡ ⎤
⎢⎢ β1 ⎥⎥ ⎜⎡ ⎤T ⎡ ⎤ ⎡ ⎤⎞−1 ⎡ ⎤T ⎢⎢ y1 ⎥⎥
⎢⎢ ⎥⎥ ⎜⎜⎢⎢ SMB1 RMW1 . . . AMLH1 ⎥⎥ ⎢⎢ SMB1 RMW1 . . . AMLH1 ⎥⎥ ⎢ λ 0 0 0 ⎥⎥⎟⎟ ⎢⎢⎢ SMB1 RMW1 . . . AMLH1 ⎥⎥ ⎢⎢⎢ ⎥⎥
⎢⎢ β2 ⎥⎥ ⎜⎜⎢⎢ ⎥ ⎢ ⎥⎥ ⎢⎢ ⎥ ⎥⎥
⎥⎥⎥ ⎢⎢⎢⎢
y2
⎢⎢
⎢⎢ .
⎥⎥ ⎜⎜⎢⎢
⎥⎥ ⎜⎜⎢⎢
.. .. .. ⎥⎥⎥ ⎢⎢⎢ .. .. .. ⎥⎥ ⎢⎢ 0 λ ⎥⎥⎥⎟⎟⎟ ⎢⎢⎢ .. .. .. ⎥⎥
⎥⎥
⎥ ⎢ ⎥⎥ ⎢⎢ 0 0 ⎥⎥⎟⎟ ⎢⎢ ⎥
⎢⎢ ..
⎢⎢
⎥⎥ = ⎜⎜⎢⎢
⎥
. . . ⎥⎥⎥ ⎢⎢⎢ . . . ⎥⎥ +⎢⎢ . ⎥⎥⎟⎟ ⎢⎢ . . . ⎥⎥⎥ ⎢⎢⎢⎢ ... ⎥⎥
⎥⎥ (9)
⎥⎥⎥ ⎜⎜⎜⎜⎢⎢⎢⎢ ⎥⎥ ⎢⎢ ⎥⎥ ⎢⎢ . ⎥⎥⎟⎟ ⎢⎢ ⎥⎥ ⎢⎢
.. .. .. .. .. .. .. . . .. .. .. ..
⎢⎢ . ⎥ ⎢ ⎥ ⎢ ⎥⎥⎟⎟ ⎢⎢ ⎥ ⎥⎥
⎥⎥⎥ ⎢⎢⎢⎢
. .
⎢⎢ .
⎢⎢ .
⎥⎥ ⎜⎜⎢⎢
⎥⎥ ⎜⎝⎣
. . . ⎥⎥⎥ ⎢⎢⎢ . . . ⎥⎥⎥ ⎢⎣⎢ . .
⎦⎥⎠⎟ ⎢⎢⎣ . . . .. ⎥⎥
⎥⎥
⎢⎣ ⎥⎦ ⎦ ⎣ ⎦ 0 0 0 λ ⎦ ⎢⎢ . ⎥⎦
SMB47 RMW47 . . . AMLH47 SMB47 RMW47 . . . AMLH47 SMB47 RMW47 . . . AMLH47 ⎣
β7 y7

When we applied the ridge regression, we need to find a λ that can provide the largest sum of
R-square which are the number of positive factors and percentage of positive coefficient. We set the
summary of these three measures as our target function to find out the optimal λ with iteration.

λ = argmax(Q) (10)

Q = P average value + number of positive coefficients + R square mean + number of variables

whose p value over 0.9 (9).

• Robustness Test

In this part, a zero-mean test for OLS and ridge regression model is adopted to check the robustness
of models. Both OLS and ridge regression models are fixed-coefficient models, which means that
they have an assumption that the covariance of errors between different units (stocks) are equal to
zero. Thereafter, a residual covariance matrix between 1097 stocks’ regression can be drawn by Python
and calculate the mean and standard deviation of all numbers in covariance matrix. Then a t-test is
conducted to test whether E(ui, j ) is equal to zero.
Covariance matrix of residual
⎡ ⎤
⎢⎢ u1,1 u1,2 ... u1,1097 ⎥⎥
⎢⎢ ⎥⎥
⎢⎢ .. . ⎥⎥
⎢⎢ u2,1 . . . . .. ⎥⎥
⎢⎢ ⎥⎥
⎢⎢ ⎥⎥ (11)
⎢⎢ .. .. . . .. ⎥⎥
⎢⎢ . . . . ⎥⎥
⎢⎣ ⎥⎦
u1097,1 ... . . . u1097,1097

u1,2 means the covariance between stock 1 and 2.

 
E ui, j
T-value =   (12)
STDE ui, j

3.3.2. Time-Series Analysis for Risk Factors

• Chi-Square Test

We use chi-square in two steps respectively. Firstly, it was used to examine the different effect of
factors in different industries. In the test we divided the significance of factors based on industries and
calculated the total value χ2 .
Secondly, it was applied to find out the pattern in the inter-factor direction prediction analysis.
We divided the pattern into four types with a combination of increase and decrease. Then we used
chi-square to find out whether there is a pattern for factor change direction. With the increase and
decrease pattern, we can predict the next quarter movement based on the current situation. As we can
see in Figure 2, suppose “0” represents the increase of risk factor in one term (or one season) and “1”
represents the decrease of risk factors.

R = Max(F1, F2) + Max(F3, F4) (13)

199
JRFM 2019, 12, 91

W = Min(F1, F2) + Min(F3, F4) (14)

Figure 2. Pattern of factor’s change.

In Figure 3, if investors spontaneously make investment decision, the frequencies of right and
wrong decisions are equitable. Both are T/2. (T is number of total transaction time)
H0: There is no relationship between right or wrong investment decisions and direction prediction rules.
H1: There is a relationship between right or wrong investment decisions and direction prediction rules.

Figure 3. Observed frequencies: fo.

In Figure 4, the expected frequency of right decisions (Re ) and wrong decisions (We ) made by
the rule can be calculated by Equations (15) and (16). Also, the expected frequency of right or wrong
decision made by random selection ((T/2)e ) can be calculated by Equation (17).

Re = (R + W ) × (R + T/2)/(R + W + T ) (15)

We = (R + W ) × (W + T/2)/(R + W + T ) (16)

(T/2)e = T × (R + T/2)/(R + W + T ) (17)

Figure 4. Expected frequencies: fe.

According to the significant table (Figure 5) of chi-square test, if total chi-square level is >2.706,
we have 90% confidence to reject the null hypothesis (H0). If total chi-square level is >3.841, we have
95% confidence to reject the null hypothesis (H0). If total chi-square level is >6.635, we have 99%
confidence to reject the null hypothesis (H0).

Total chi-square level = (R − Re )2 /Re + (W − We )2 /We + 2[T/2 − (T/2)e 2 ]/(T/2)e (18)

200
JRFM 2019, 12, 91

The degree of freedom = (column − 1)(row − 1) = (2 − 1)(2 − 1) = 1 (19)

Figure 5. Chi-square level: χ2 = (fo − fe )2 f.

• Back-Test for Trading Strategy

After factors’ cyclical research and trend analysis for fluctuation of risk factors, the trading strategy
based on moving average approach and correlations between factors and industries can be formed.
By deciding the time to do the transactions and stocks which should be bought or sold, the float return
from 2007S4 to 2018S3 and annual expected return can then be calculated.

3.4. Assumptions for Multi-Factor Examination

The following assumptions are made in applying the multi-factor model:

1. Perfect market: there are no tax and transaction costs.

2. People are risk averse or rational.
3. People can lend or borrow money at a risk-free rate freely.
4. There is a trade-off between risk and return for all securities.
5. Everyone can obtain market information equally and freely.
6. Investors have the same expectations for this market.

The intuition and assumption behind the hypotheses are presented in the Appendix A with
measurements of factors and graphs.

3.5. Models and Variable Definitions

Model 1:
 
Ri − R f = a + βmarket Rm − R f + βsize SMB + βBM HML + βpro f itability RMW + βinvestment CMA + ε

Model 2:
 
Ri − R f = a + βmarket Rm − R f + βsize SMB + βBM HML + βpro f itability RMW+
βinvestment CMA + βmomentum CRMHL + βturnover AMHL + ε

The details of the factors (Rm-Rf, SMB, HML, RMW, CMA, CRMHL, AMHL) including explanation
and graphs are presented in Appendix A.
Table 1 displays the dependent variable, independent variables and the corresponding labels
and explanations.

Table 1. Dependent variable and independent variables.

Name Label Note

DV
The return of security mines risk-free rate of
Excess return Ri − R f
return

201
JRFM 2019, 12, 91

Table 1. Cont.

Name Label Note

IV
Abnormal return a The constant term of formula
The return of market index (in this model, market
Market premium Rm − R f index is Shanghai stock exchange market index)
mines risk-free rate of return
Size premium The return on a diversified set of small stocks
SMB
(Small minus Big) minus the return on a diversified set of big stocks.
Book-to-market premium (High The difference between the returns on diversified
HML
minus low) portfolios of high and low B/M stocks.
The difference between the returns on diversified
Profitability premium (Robust
RMW portfolios of stocks with robust and
minus weak)
weak profitability.
The difference between the returns on diversified
Investment growth premium portfolios of the stocks of low and high
CMA
(Conservative minus aggressive) investment firms, which we label as conservative
and aggressive.
The difference between higher momentum
Momentum premium (High (higher accumulated return) companies’ average
momentum minus low CRMHL return and lower momentum (lower accumulated
momentum) return) companies’ average return in a diversified
portfolio or in the market.
Asset turnover premium The difference between low asset turnover
(Low turnover rate minus high AMLH companies’ average return and higher turnover
turnover rate) companies’ average return or in the market.

4. Results

4.1. Multi-Factor Model Examination with Single Stock Regression Analysis

4.1.1. Stability Test

DV: Ri-Rf
IV: Rm-Rf, SMB, RMW, HML, CMA, CRMHL, AMLH
CV: Time, Timeˆ2, Season

According to the result of stability test (presented in Appendix E), the control variable including
Time, Timeˆ2, Season did not pass the t-test, the significance level is far less than 0.9. Thus, they are
removed from the model. It means that excess return is not affected by time.

4.1.2. Regression Models

• OLS Regression

Model 1: Five-factor model

DV: Ri-Rf
IV: Rm-Rf, SMB, RMW, HML, CMA

Model 2: Seven-factor model

DV: Ri-Rf
IV: Rm-Rf, SMB, RMW, HML, CMA, CRMHL, AMLH

202
JRFM 2019, 12, 91

Model 3: Modified optimal model

DV: Ri-Rf
IV: Rm-Rf, RMW, HML, CMA, CRMHL, AMLH

• Ridge Regression

Model 4: Ridge regression model

DV: Ri-Rf
IV: Rm-Rf, SMB, RMW, HML, CMA, CRMHL, AMLH

We conducted OLS regression on Model 1, 2, and 3, and ridge regression on Model 4. After conducting
the regression analysis, we recorded each factors’ coefficients and mean of R square. We created frequency
histograms to observe the distribution of coefficients. The graphs provide a general view of the coefficients.
In Table 2, above each histogram, there are two numbers, mean of coefficients and the proportion
of coefficients that passed the t-test (in the parenthesis).
To be specific, we divided the coefficients into uncorrelated, positively correlated, and uncorrected,
based on 95% confidence interval. Table 3 shows the percentage of coefficient. SMB, CMA, Rm-Rf, CRMHL,
and AMHL have more positive coefficients. Whereas RMW and HML have more negative coefficients.
In ridge regression, we obtain the λ by calculating the maximum of the objective function Q.
When λ equals to 0.008, the result of ridge regression is maximized. In Figure 6, we can find out the
largest Q value at 0.008.

Figure 6. Target function optimization.

Some factors have negative effects on the excess return, which deviates from our assumption.
Moreover, regarding the factors that have an even distribution. We need to classify them in industry
groups to discover a further pattern.

203
 Table 2. R square mean, coefficients’ mean and p-level for four models.

OLS Ridge Regression

Model 1 Model 2 Model 3 Model 4
JRFM 2019, 12, 91

0.5692 0.6078 0.5911 0.6048

Mean of R square

−0.29 0.29 0.82 0.21

(0.9535) (0.9991) (0.9991) (0.9991)

Rm-Rf

204
−0.29 0.34 0.29
(0.8669) (0.1641) (0.8368)

Mean of coefficient
(p-level) SMB

−0.74 0.24 0.40 0.22

(0.9189) (0.9891) (0.9763) (0.9298)

RMW
 Table 2. Cont.

OLS Ridge Regression

Model 1 Model 2 Model 3 Model 4
JRFM 2019, 12, 91

0.13 0.30 −0.27 0.22

(0.3874) (0.9690) (0.9444) (0.9681)

HML

−0.11
(0.9335) 0.58 −0.50 0.38
(0.9900) (0.9763) (0.9909)

CMA

205
Mean of coefficient
(p-level)

−0.14 0.33 0.02

(0.9617) (0.9243) (0.9535)

CRMHL

−0.04 −0.72 0.06

(0.9991) (0.9991) (0.9991)

AMLH
JRFM 2019, 12, 91

Table 3. Correlations between seven factors and excess return of stocks.

SMB RMW HML CMA AMLH CRMHL Rm-Rf

Positive 61.0% 24.2% 23.0% 56.2% 62.4% 77.8% 99.8%
Negative 22.7% 68.7% 73.8% 42.8% 37.5% 17.6% 0.1%
Uncorrelated 16.3% 7.0% 3.2% 0.9% 0.1% 0.1% 0.1%

4.1.3. Robustness Test: Zero Mean Residual Testing

According to Table 4, since t-values in four models are less than 2.021, we have 95% confidence
not to reject the null hypothesis, which means the residuals are randomly independent for all stocks,
both the OLS and ridge regression model pass the zero mean residual test.

Table 4. Zero mean residual testing.

OLS t-Value Ridge Regression t-Value

Model 1 Model 2 Model 3 Model 4
0.5811 0.5393 0.5416 0.5387

4.1.4. Industry Analysis

According to the result of chi-square test, each industry corresponds differently to the factors.
Therefore, effect of factors needs to be discussed based on different industries respectively. The details
of the test results are presented in the Appendix B.
We classify the stocks in 28 industries. Also, we divided the coefficients in positive correlation
and negative correlation. As for data cleaning, we only preserved the factors that contained over 70%
coefficients are of the same signs. The specific table is presented in the Appendix C.
In the Table 5, the industries are classified in manifold groups based on the significance level and
the relationship with each factor.
According to the table, there is no industry that can be explained by seven factors simultaneously.
The banking and steel industries are best fitted in the seven-factors model, since they have six significant
factors and the highest sum of percentage of significant coefficients.
Each factor has different relationships with different industries. For market premium (Rm-Rf), it
can apply to all industries. For SMB, it is positively related with seven industries (Group A), yet SMB is
negatively correlated with Steel industry. With respect to RMW, it has negative relationships with most
industries (Group D). However, RMW has significant positive relationships with bank industry. HML is
negative related to multiple industries (Group F). Also, HML has significant positive relationships with
the banking industry. CMA is positively correlated with five industries (Group G). Meanwhile, CMA is
negatively correlated to three industries (Group H) including extractive, banking, and nonferrous metal.
CRMHL is positive correlated with most of industries (Group I) except media, electrical equipment,
non-bank finance, animal husbandry and fishery, commercial, and comprehensive industries. AMLH
has a negative correlation with Group K. At the same time, AMLH is positively related with three
other industries (Group J) including housing, non-bank finance, and banking.
Moreover, the industries with the same pattern are regarded as coordinated industry. Additionally,
the industries with different pattern are regarded as deviated industries. To be specific, the banking
industry is considered as a deviated industry since it has a positive RMW and HML factor, while other
industries have negative factors. The graph shows four groups of coordinated industry—including
telecommunication and electronics, chemistry and automobile, light manufacturing and defense,
and nonferrous metal and extractive—while banking is regarded as a deviated industry.
Considering the trend of the factors, we can formulate corresponding strategies. To be specific,
based on the condition of each factor and the classification from Group A to L which consist of different
industries presented in Table 5, we formulated manifold strategies in Table 6.

206
 Table 5. Relationships of industries and seven factors.

Positive Relationships Negative Relationships

(Group A: 7)
(Group B: 1)
SMB Electrical Equipment, Electronics, Computer, Construction Material,
JRFM 2019, 12, 91

Steel
Food, Telecommunication, and Leisure Service
(Group D: 16)
Extractive, Media, Electrical Equipment, Textiles & Garments,
(Group C: 1) Steel, Defense, Chemistry, Mechanical Equipment, Computer,
RMW
Banking Construction Material, Transportation, Automobile, Light
Manufacturing, Leisure Service, Nonferrous Metal, and
Comprehensive Industries.
(Group F: 19)
Extractive, Media, Electrical Equipment, Electronics, Housing,
Textiles & Garments, Steel, Defense, Chemistry, Mechanical
(Group E: 1)
HML equipment, Computer, Animal Husbandry and Fishery,
Banking
Automobile, Light Manufacturing, Commercial,
Telecommunication, Leisure Service, Medical, Nonferrous Metal,
and Comprehensive Industries.

207
(Group G: 5)
(Group H: 3)
CMA Textiles & Garments, Utilities, Defense, Light Manufacturing, and
Extractive, Banking, and Nonferrous Metal
Commercial
(Group I: 20)
Extractive, Electronics, Housing, Textiles & Garments, Steel,
Utilities, Defense, Chemistry, Mechanical Equipment, Computer,
CRMHL
Domestic Appliance, Construction material, Transportation,
Automobile, Light Manufacturing, Telecommunication, Leisure
Service, Medical, Banking, and Nonferrous Metal.
(Group K: 14)
Extractive, Electrical Equipment, Electronics, Steel, Defense,
(Group J: 3)
AMLH Chemistry, Domestic appliance, Automobile, Light
Housing, Non-Bank Finance, and Banking.
Manufacturing, Food, Telecommunication, Leisure Service,
Medical and Nonferrous Metal.
(Group L: 28)
Rm-Rf
All industries
 Table 6. Trading strategies based on condition of factors.

Condition When Factor Is Positive When Factor Is Negative

Strategies BUY SELL BUY SELL
JRFM 2019, 12, 91

Big MV Big MV
Small MV companies in Small MV companies in
companies in companies in
Group A Group A
SMB Group A Group A
Big MV companies in Big MV companies in
Small MV companies in Small MV companies in
Group B Group B
Group B Group B
Robust ROE companies in Weak ROE companies in Weak ROE companies in Robust ROE companies in
Group C Group C Group C Group C
RMW
Weak ROE companies in Robust ROE companies in Robust ROE companies in Weak ROE companies in
Group D Group D Group D Group D
High B/M companies in Low B/M companies in Low B/M companies in High B/M companies in
Group E Group E Group E Group E
HML
Low B/M companies in High B/M companies in High B/M companies in Low B/M companies in
Group F Group F Group F Group F
Conservative (low growth rate of
assets) companies in Aggressive companies in Aggressive companies in Conservative companies in

208
Group G Group G Group G Group G
CMA Aggressive (High growth rate of Conservative Conservative Aggressive
assets) companies in companies in companies in
companies in Group H Group H Group H
Group H
High CR High CR
CRMHL companies in companies in
Group I Group I
High Asset turnover High Asset turnover
Low Asset turnover companies in companies in companies in Low Asset turnover companies in
Group J Group J Group J Group J
AMLH
High Asset turnover companies in Low Asset turnover Low Asset turnover High Asset turnover companies in
Group K companies in companies in Group K
Group K Group K
All companies All companies
Rm-Rf
(Group L) (Group L)
JRFM 2019, 12, 91

Here we can take an example to illustrate the above table, when SMB is positive, it means in this
market, small market value (MV) companies outperform big companies, thus for those industries
which is significantly positive correlated with SMB, we should buy small MV stocks and sell large MV
stocks. However, for those industries which is negatively correlated with SMB, the strategy should
purchase large MV stocks and sell small MV stocks. When SMB is negative, vice versa.
According to the significant level and correlation effect between 7 factors and 28 industries,
the trading portfolio and strategy can be conducted. Since the strategic making depends on the positive
or negative condition of seven factors, if we can forecast the conditions of seven factors, the trading
strategies can be easily made. Thus, for the next two parts, we explored the fluctuation of seven factors
to answer two questions:

1. From analysis of factors’ changing pattern, can we find the reasons or elements to illustrate the
fluctuation of seven factors? (if we find out the driving factor which stimulate the moving of
other factors, we can explain the most essential ratio that investor may focus on.)
2. Can we find an approach to forecast the moving of each factor and apply it to do the back test for
trading strategies?

4.2. Time-Series Analysis for Risk Factors

4.2.1. Endogeneity and Exogeneity for Factors’ Cyclical Research

• Endogeneity

According to Pearson Correlation matrix, last term data of SMB, RMW, HML, CMA, CRMHL,
and AMLH has significant correlation with this term at the level of 0.01. Nevertheless, for the market
factor (Rm-Rf), its last term data are nearly independent with this term, which means except market
risk factor, other factors are possible to forecast themselves with last term data.
However, to test the endogeneity of these factors, chi-square test are necessary for examination
and application of these relationships. After the test, the results show that only SMB, RMW pass the
test with 99% confidence level and CMA also has a significance level of 0.05, which means these three
factors exist endogeneity and can be truly used to forecast their direction and investment cycle.

• Exogeneity

Based on Pearson correlation matrix, the fluctuation of HML, CRHML, and AMLH depend upon
the history data of other factors. For details, HML is related to the AMLH of last term. CRHML
has a small correlation with HML AMLH bridge its relationships with RMW of last term. Finally,
with respect to Rm-Rf, its volatility is partly connected to last term’s RMW and AMLH.
Figure 7 summarizes the endogeneity and exogeneity for predicting direction of seven risk factors.
In the chi-square test, RMW, SMB, and CMA factor can be only predicted by the last term direction
of themselves. With confidence level of 95%, CRHML’s direction can forecasted by its history data
with HML and the increase and decrease of HML can be predicted by last term direction and AMLH
at 90% confidence level. For market factor (Rm-Rf)—even if single stock history data cannot predict
next term—with RMW and AMLH, market factor can estimate the direction of its next term with
95% confidence. With a confidence level of 90%, AMLH’s fluctuation direction can forecast by RMW.
The details of prediction are presented in Appendix D.

209
JRFM 2019, 12, 91

Figure 7. Summary of endogeneity and exogeneity for predicting direction of seven risk factors.

4.2.2. Trading Strategy with Trend Analysis

In the last two parts, with the examination of chi-square test, the directions of fluctuation of seven
factors are predictable, yet the range of fluctuation is still unknown.
Therefore, in order to forecast the positive and negative levels of each risk factor, trend analysis
is conducted.
The above seven graphs in Figure 8 describe the trend of each factor. The orange line is the
four-season moving average. SMB, HML, and CMA are mostly below 0. By contrast, RMW and CRMHL
are above 0. While AMLH and Rm-Rf fluctuate around 0.4.

Figure 8. Four seasons moving average of seven factors.

4.2.3. Back-Testing
Table 7 illustrates the results of the back-testing from 2007S4 to 2018S3 for the trading system.
With the strategy, we can achieve positive annual return in SMB, CMA, CRMHL, and AMLH for
15.43%, 3.23%, 29.13%, and 8.4% respectively. To some extent, the result proved the practicability of
the seven-factor model, especially for the factor CRMHL in terms of trading strategy. Consequently,
with alternative trading strategies which can predict the factor fluctuation more precisely, we can
witness a greater margin of improvement.

210
JRFM 2019, 12, 91

Table 7. Back-testing float return from 2007S4 to 2018S3 and expected annual return for factors.

Factors Float Return Expected Annual Return

SMB 165.89% 15.43%
RMW −6.525% −0.61%
HML −22.29% −2.07%
CMA 34.67% 3.23%
CRMHL 313.09% 29.13%
AMLH 90.26% 8.40%

5. Discussion

5.1. Analysis of Multi-Factor Model

We compared three multi-factor models in OLS regression and proceed to ridge regression for
higher significance and explanatory power. In the seven-factor model, only RMW and HML factor
have negative effects on the stock excess return, which is deviated from the assumption. To be specific,
negative RMW means that a company with lower ROE can achieve higher excess return than higher
ROE company. Also, negative HML means lower book-to-market ratio result in higher excess return.
According to the previous empirical studies in Chinese stock market, the low level of ROE often
corresponds to low stock price. On the contrary, relatively high ROE means that the stock price has
reached a periodic top. We can infer that stocks with relatively low ROE can attract investment which
can boost the stock price. As for the CMA, a company with lower growth of investments has higher
excess return. We can infer that investors prefer purchase stock with lower growth of investment.

5.2. Industry Analysis

We merely discussed the industries with more than four significant factors. Since they are
consistent with the model to the higher extent. As we can see in Figure A10 in Appendix C, among
12 industries, there are four groups of coordinated industries and one group of deviated industry.
We found out that the coordinated industries are usually cyclical industries which are highly related to
the economic wave. Cyclical industry primarily consists of two categories, resources and industrial
raw material. They are closely related to macro economy cycle and supply–demand relationship such
as manufacturing, automobile, metal and chemistry industry. Moreover, the coordinated industry
is distributed at upstream and downstream; for example, electronics is the upstream industry for
telecommunication. Thereby, the chain effect is transmitted from the upstream industry through a
series of companies all the way to the downstream industry.
To be specific, regarding different factors, industries may react differently in terms of positive and
negative relationship. For SMB, it has a negative effect on steel industry, while six industries stand
in the opposite position including electrical equipment, electronics, computers, construction material,
telecommunication, etc. There are two reasons for the adverse effect in steel industry. Firstly, steel industry
is widely regarded as traditional industry. Secondly, it is dominated by a few giant companies which are
supported by the government. Whereas the six companies with positive effects are in opposite condition.
They are relatively small in terms of market value. Most companies in electrical equipment and computer
industry are in early stage which have potential to achieve higher excess return.
As for RMW, over half of the industry are in a negative relationship, while banking is positive.
Since banking is closely related to the macro-economic policy. Therefore, it responds to market
situations like ROE more swiftly. Moreover, we ran the regression on the same period which may
reveal the lag of the information dissemination in some of the industry. Since the market is imperfect,
the companies may not respond to the information in the exact same period.
For HML, over 70% of the companies have a negative effect, while banking has a positive
relationship. A higher book-to-market ratio reflects that a bank is increasing the amount of loan
issue. Since the major source of incomes of banking come from loan issue, thus a bank with higher

211
JRFM 2019, 12, 91

book-to-market ratio can achieve a higher excess return. While other industries with low book-to-market
ratio can achieve higher excess return.
As for CMA, five companies—including textiles & garments, utilities, and light manufacturing—consist
of manufacturing processes. They have constant cash flow to maintain daily operation and dividend
payment. They are also relatively conservative in investment. However, banking and non-bank finance is
negative related to CMA factor. It can be inferred that they are more aggressive in investment to achieve
higher excess return.
For CRMHL, none of the companies are in negative relationship, while over 70% of the industries
are positive to the momentum factor. This means that the momentum effect widely applied to industries,
which can result in the similar movement in the next period.
When it comes to AMLH, the housing and finance industries are in a positive relationship. In these
industries, assets in terms of land and money reserves are placed in the foremost position, which means
that lower asset turnover rates can result in higher excess return. However, for industries with a
negative relationship, since asset turnover rate is the reflection of operating capacity, a higher asset
turnover rate can lead to higher excess return.

5.3. Factor Cyclical Research

According to the result of chi-square test, the future direction of SMB and CMA can be merely
predicted based on the previous period. While RMW has the initial effect to the rest of the factors,
this chain effect conveys through AMLH and HML to Rm-Rf and CRMHL respectively. Although
the size of this effect cannot be predicted, we can find the pattern of investment in this chain effect
which can reflect investor behavior in the stock market. Firstly, investors primarily focus on the ROE
of the companies. Together with the asset turnover rate, we can predict the future direction of the
market premium (Rm-Rf). Secondly, the future direction of CRMHL can also be predicted by HML
and AMLH. It can suggest that investors usually refer to the profitability and operation capacity
of companies before making investment. These two factors can affect the company evaluation and
eventually reflected on the change of momentum factor. According to the investment pattern, we can
establish a corresponding investment strategy.
Since the direction of market premium factor (Rm-Rf) can be predicted, to some extent, we can
find out the pattern of the index. Therefore, it can be also be applied in the investment of index futures
and options through call and put.

5.4. Trading Strategy and Back Test

According to the trend analysis, a moving average model is applied in the model (Hanke and
Wichern 2014). However, the span of moving average is not specifically decided, which can affect the
result of back-testing. We only conducted a four-season moving average as an example. In order to find
out the optimal moving average, iteration can be applied to figure out the optimal trading strategies.

5.5. Significance and Limitations of Research

With the comparison of previous work and our research, academic significance can be illustrated
in two aspects.
For one thing, prior studies mainly focus on the effect of factors on excess return from a portfolio
aspect. Stocks are classified into different groups based on their characteristics. For example,
the companies are divided into two groups based on size, then each group is divided into two groups
based on the book-to-market ratio. By comparing the average return of the four groups, they can find
out the relationship between factors and return. Thus, their methods can only be applied to investments
of a specific portfolio, which may lack of practicability and explanatory power in application to a
single stock or specific industries. For this research, nevertheless, we focus on discovering different
effects of risk premium factors to the excess return of single stock via OLS and ridge regression and
then summarize the significant correlations between factors and different industries, which means our

212
JRFM 2019, 12, 91

research is more practical. For example, based on our model, investors can judge the factors which
significantly affect a specific stock or industry.
Additionally, previous works have never discussed the risk premium factors from the aspect of
time-series analysis. While forecasting the investment cycle, prior research only investigated from a
macro-economic level or technical analysis. For one example, in Business Cycles (Lars 2006), the author
merely analyzed the cycle of housing, credit, and inventory, which belonged to macro-economic
level. For another example, technical analyses, like Elliott wave principle and candlestick charts,
aim to describe the market price wave pattern on time series. Thereafter, there is an academic gap of
discussing investment cycle of fundamental analysis. In this paper, however, the fluctuations of seven
factors are studied by chi-square test of endogeneity and exogeneity of factors. Moreover, it is a new
type of cycle analysis because it explains the companies’ value behind the investment decision.
The contribution consists of four points. Firstly, we applied a novel method to find out the
correlations between seven risk factors and each single stock. Secondly, we find out and explain
the correlations between seven risk factors and each industry and specify the situations (positive or
negative) of risk factors to buy or sell the industries’ stocks. Thirdly, the result of cyclical analysis can be
applied in forecasting the direction of risk factors, especially the market risk factor (Rm-Rf) which can
be used in transaction of options and futures of market index. Lastly, a back-test was conducted in a
simple trading system which suggested that SMB (size premium), CMA (investment growth premium),
CRMHL (momentum premium), and AMLH (asset turnover premium) can gain positive returns.
As for the limitations, this research conducted a four-seasons-moving average method to forecast
the level of seven factors. Even if it proved that SMB, CMA, CRMHL, and AMLH can be applied in
trading system. However, more forecasting methods like ARIMA, VAR, and ANN can be constructed
to form better trading strategies. More limitations are presented with our direction for further study.

6. Conclusions and Further Study

In the research of relationship between risk premium and excess return, the hybrid approach
takes a primary position. In the examination and modification of multi-factor model, OLS and ridge
regression are conducted on Models 1 to 4. The seven-factor model has an optimal combination of p
level and R square. In chi-square test, the effect of factor was responded differently in each industry.
Therefore, p level was calculated based on the industry in order to find out the well-fitted coordinated
and deviated industry. Bank and steel industry are well-fitted in the model, while industries within
the same stream—e.g., telecommunication and electronics—are found to be coordinated industries.
Moreover, cyclical industry is usually coordinated except for banking.
In the factor cyclical research, chi-square and correlation tests are applied to find out the endogeneity
and exogeneity. SMB, RMW, and CMA have endogeneity while the remaining four factors have both
endogeneity and exogeneity. With the pattern of direction in factors, an investment strategy was
established. To be specific, when moving average SMB is positive, small market value (MV) companies
outperform big companies, thus for those industries which are significantly positively correlated with
SMB, it would be better to buy small MV stocks and sell large MV stocks. With the strategy, we can
achieve a positive return in SMB, CMA, CRMHL, and AMLH respectively in the back test.
As for the further study, the proportion of individual investors in the Chinese stock market
contribute nearly 80% of the trading volume. Therefore, investor preference and the irrational behavior
should be considered. The investor sentiment factor may be added to improve the explanatory power.
Previous studies also illustrated that the reversal effect exists in the long-term stock market. If our
period is set to be longer than one year, we can add the reversal factor to explain the reversal effect.
Dynamic analysis can be applied to data processing, since the economic environment changes
over time. We can also monitor the stock market and provide suggestion in stock selection which can
fit for the target return.

213
JRFM 2019, 12, 91

Moreover, we can further study the inter-factor drive relationship, in order to establish more
investment strategies. To be specific, by optimizing the back-test ratio and Sharpe ratio, we can construct
a better investment portfolio.

Funding: This research received no external funding form government or any institution.
Conflicts of Interest: We declare that there is no conflict of interest in this research.

Appendix A. Intuition and Assumption Behind the Hypotheses

There is a trade-off between return and risk. In order to find the corresponding risk for the excess
return, according to CAPM model, there is a market risk. To illustrate the abnormal return, Fama used
three-factor model which added size premium and book-to-market premium in 1993. In 2013, in their
five-factor model, profitability and investment growth are also considered to be significant coefficients.
Next, the intuition behind these premiums will be explain one by one.
1. Market Premium

Figure A1. Positive relationship between expected return and market premium.

Market premium is represented by the difference between market return and risk-free rate.
Since the fluctuation of stock market’s expected return is higher than the risk-free rate, namely stock
market has higher risk, the expected return of stock market should higher than risk-free rate.
2. Size Premium

Figure A2. Positive relationship between expected return and size premium.

Size premium is the difference between small companies’ average return and big companies’
average return in a diversified portfolio or in the market. Normally, small companies have higher
returns than the bigger ones. Because at the same period of time, small companies’ profit is easy to
grow faster than big companies and the growth rate of their dividend also is higher.
According to the DDM model, the price of security depends on the discounted present value of
future dividend. In this formula, ‘P’ is the expected present price of one stock. ‘D’ is the dividend of

214
JRFM 2019, 12, 91

this year and ‘g’ is the constant growth rate of the dividend (it also is the growth rate of profit). For the
last, ‘i’ is the dividend interest rate, which may refer to risk-free rate.

D(1+ g) D(1+ g)2 D(1+ g)3 D(1+ g)n

P = lim [ 1+i + 2 + +...+ (1+i)n
]
n→∞  (1+ ni) (1+i)3
1+ g
D(1+ g) 1+i −1
= lim [ 1+i ∗ 1+ g ]
1+i −1
n→∞

Given by i > g,
n
1+g
→0
1+i
D(1+ g)
So, P = i−g
D1 (1 + g1 ) D2 (1 + g1 )
∵ P1 = , P2 =
i − g1 i − g1
If we assume D1 = D2
Hence, the return rate of the stock can be represented by

P2 − P1 ( g2 − g1 )(i + 1)
Ri = =
P1 (1 + g1 )(i − g2)

In terms of small companies and big companies, g1 (the growth rate o f f irst term) and i (risk − f ree
rate) are the same.
However, in the second term, with the control of other effects, growth rate (g2 ) of small companies
is higher than the big one, so the expected return (Ri ) of small companies is generally higher than the
large one.
In other words, high returns of small companies indicate they also carry higher risk. If investors can
suffer the risk brought by small companies, they can gain the risk premium which is called ‘size premium’.
3. Book-to-Market Premium

Figure A3. Positive relationship between expected return and book-to-market premium.

Book-to-market premium is the difference between high B/M ratio companies’ average return and
low B/M ratio companies’ average return in a diversified portfolio or in the market.
There is an effect named B/M effect which indicates that higher B/M ratio companies has higher
excess return. It can be illustrated by prospect theory easily. However, in this case, the x-axis is MV
and the y-axis is the value.

215
JRFM 2019, 12, 91

Figure A4. Relationship between true and cognitive value of stocks with the growth of MV/B.

For the higher B/M ratio companies, the MV/B is relative lower and people always overprice the
true value of a stock, it leads to the demand for those companies is higher. With the growth of demand
and stock price, the return of those stocks also is higher.
Vice versa, higher B/M ratio companies has lower expected return and excess return (expected
return minus risk of free rate).
Therefore, if the investor prefers higher B/M ratio companies, they take the risk of B/M effect on
one hand, they gain the B/M premium on the other hand.
4. Profitability Premium

Figure A5. Positive relationship between expected return and profitability premium.

Profitability premium is the difference between robust profitability (higher ROE) companies’
average return and weak profitability (lower ROE) companies’ average return in a diversified portfolio
or in the market.
With the control of other effects, companies with robust profitability—measured by the level of
ROE (return of equity)—outperform in their expected rate of return and take greater variance. This is
because companies with high profits also distribute high dividends.

D(1 + g)
P =
i−g

According DDM (dividend discount model) formula, higher dividend means higher price and
demand which will enhance the level of expected return. If an investor purchases companies with robust
profit, they may get higher excess return and fluctuation at the same time. Vice versa, weak profitability
companies bring people low excess return and risk.

216
JRFM 2019, 12, 91

5. Investment Growth Premium

Figure A6. Positive relationship between expected return and investment growth premium.

Investment growth premium is the difference between conservative (lower growth rate of investment
or lower growth rate of assets) companies’ average return and aggressive (higher growth rate of investment
or higher growth rate of assets) companies average return in a diversified portfolio or in the market.
The reason why aggressive companies may have low excess return and risk is that these kinds of
firms allocate more profit into reinvestment rather than dividends, thus it decreases the expected price
and return, which leads to low risk. Vice versa, conservative companies bring higher excess return and
risk because of larger amount of dividend rather investment.
6. Momentum Premium

Figure A7. Positive relationship between expected return and momentum premium.

Momentum premium is the difference between higher momentum (higher accumulated return)
companies’ average return and lower momentum (lower accumulated return) companies’ average
return in a diversified portfolio or in the market.
Momentum is the accumulated return in one quarter. The higher one means the stock is popular
with high return and risk. Vice versa, people invest in low momentum companies with low premium
and risk.

217
JRFM 2019, 12, 91

7. Asset Turnover Premium

Figure A8. Positive relationship between expected return and asset turnover premium.

Asset turnover premium is the difference between higher asset turnover companies’ average
return and lower asset turnover companies’ average return in a diversified portfolio in the market.
Asset turnover is the total revenue divided by total asset. The higher one means the stock is
popular with high return and risk. Vice versa, people invest in low momentum companies with low
premium and risk. The regression analysis was conducted with a opposite direction.

Appendix B. Chi-Square Test of Industry in Different Factors

Table A1 recorded the total chi-square value of 28 industries for seven factors. By conducting
the chi-square test, we can find out the different effect of factors to various industries. When the total
chi-square value is larger than the critical value (when df = 27), it means that each industry is responding
differently to the specific factors. According to the result, six factors passed the test, while Rm-Rf is
insignificant. Therefore, we need to discuss the effect of factors based on different industry.

Table A1. Chi-square test of industry in different factors.

Factors Chi-Square (χ2 )

SMB 97.58155
RMW 102.9963
HML 134.683
CMA 85.7257
Rm-Rf 38.17769
CRMHL 66.58794
AMLH 179.6194

Appendix C. Significance Level and Correlation Effect

Table A2. Significance level of factors in different industries.

Title SMB RMW HML CMA CRMHL AMLH Rm-Rf

Extractive 0 −0.76 −0.8 −0.72 0.72 −0.72 1
Media 0 −0.78571 −1 0 0 0 1
Electrical equipment 0.741935 −0.80645 −0.93548 0 0 0 1
Electronics 0.833333 0 −0.875 0 0.729167 −0.70833 1
Housing 0 0 −0.72527 0 0.714286 0.791209 1
Textiles & garments 0 −0.70833 −0.875 0.708333 0.916667 0 1
Non-bank finance 0 0 0 0 0 0.714286 0.964286
Steel −0.73684 −0.94737 0.789474 0 1 −0.89474 1

218
JRFM 2019, 12, 91

Table A2. Cont.

Title SMB RMW HML CMA CRMHL AMLH Rm-Rf

Utilities 0 0 0 0.710526 0.828947 0 1
Defense 0 −0.8 −0.72 0.8 0.8 −0.84 1
Chemistry 0 −0.74699 −0.78313 0 0.795181 −0.84337 1
Mechanical equipment 0 −0.76667 −0.76667 0 0.833333 0 1
Computer 0.777778 0 −0.92593 0 0.777778 0 1
Domestic appliance 0 0 0 0 0.857143 −0.7619 1
Construction material 0.714286 −0.7619 0 0 0.904762 0 1
Construction ornament 0 0 0 0 0.933333 0 1
Transportation 0 −0.81633 0 0 0.857143 0 1
Animal husbandry and
0 0 −0.71429 0 0 0 1
fishery
Automobile 0 −0.86364 −0.75 0 0.818182 −0.75 1
Light manufacturing 0 −0.73077 −0.76923 0.730769 0.730769 −0.80769 1
Commercial 0 0 −0.88525 0.754098 0 0 1
Food 0.794118 0 0 0 0.794118 −0.76471 1
Telecommunication 0.727273 0 −0.77273 0 0.772727 −0.77273 1
Leisure service 0.8125 −0.875 −0.8125 0 0.8125 −0.8125 1
Medical 0 0 −0.80412 0 0.835052 −0.74227 0.989691
Bank 0 0.714286 1 −0.71429 1 1 1
Nonferrous metal 0 −0.88372 −0.74419 −0.83721 0.790698 −0.86047 1
Comprehensive 0 −0.92308 −0.80769 0 0 0 1

In this table, 0 represents that the factor is insignificant to the industry. The absolute value
represents the percentage of the significant coefficients. A positive number represents that the factor
has positive effect on the industry, while negative number represents that the factor has a negative
effect on the industry.

Appendix D. Forecasting the Direction of Factors

Figure A9. Use SMB to forecast the direction of SMB.

219
JRFM 2019, 12, 91

Figure A10. Use RMW to forecast the direction of RMW.

Figure A11. Use CMA to forecast the direction of CMA.

Figure A12. Use AMLH and HML to forecast the direction of HML.

Figure A13. Use HML and CRMHL to forecast the direction of CRMHL.

220
JRFM 2019, 12, 91

Figure A14. Use RMW and AMLH to forecast the direction of AMLH.

Figure A15. Use RMW, AMLH, and Rm-Rf to forecast the direction of Rm-Rf.

Appendix E. Result of Stability Test

Figure A16. Significance level of variables.

The figures in the red frame are the significance level of time, timeˆ2 and season.

References
Aharoni, Gil, Bruce Grundy, and Qi Zeng. 2013. Stock returns and the miller modigliani valuation formula:
Revisiting the fama French analysis. Social Science Electronic Publishing 110: 347–57. [CrossRef]
Akter, Nahida, and Ashadun Nobi. 2018. Investigation of the financial stability of s&p 500 using realized volatility
and stock returns distribution. Journal of Risk and Financial Management 11: 22.

221
JRFM 2019, 12, 91

Banz, Rolf W. 1981. The relationship between return and market value of common stocks. Journal of Financial
Economics 9: 3–18. [CrossRef]
Bhattacharjee, Arnab, and Sudipto Roy. 2019. Abnormal Returns or Mismeasured Risk? Network Effects and Risk
Spillover in Stock Returns. Journal of Risk and Financial Management 12: 50. [CrossRef]
Bodie, Zvi, Alex Kane, Alan J. Marcus, and Ravi Jain. 2014. Investments. New York: Mc Graw Hill Education.
Bodie, Zvi, Alex Kane, and Alan J. Marcus. 2017. Essentials of Investments. New York: McGraw-Hill Education.
Carhart, Mark M. 1997. On persistence in mutual fund performance. The Journal of Finance 52: 26. [CrossRef]
Chen, Jieting, and Yuichiro Kawaguchi. 2018. Multi-factor asset-pricing models under markov regime switches:
Evidence from the chinese stock market. International Journal of Financial Studies 6: 54. [CrossRef]
Cisse, Mamadou, Mamadou Konte, Mohamed Toure, and Smael A. Assani. 2019. Contribution to the Valuation of
BRVM’s Assets: A Conditional CAPM Approach. Journal of Risk and Financial Management 12: 27. [CrossRef]
Fama, Eugene F. 1996. Multifactor portfolio efficiency and multifactor asset pricing. The Journal of Financial and
Quantitative Analysis 31: 441–65. [CrossRef]
Fama, Eugene F., and Kenneth R. French. 1993. Common risk factors in the returns on stocks and bonds. Journal of
Financial Economics 33: 3–56. [CrossRef]
Fama, Eugene F., and Kenneth R. French. 2015. A five-factor asset pricing model. Journal of Financial Economics
116: 1–22. [CrossRef]
Frazzini, Andrea, David Kabiller, and Lasse H. Pedersen. 2013. Buffett’s alpha. CEPR Discussion Papers 3: 583–90.
Hanke, John E., and Dean W. Wichern. 2014. Business Forecasting. Zug: Pearson Schweiz Ag.
Hsiao, Cheng. 2003. Analysis of Panel Data. Cambridge: Cambridge University Press.
Jegadeesh, Narasimhan, and Sheridan Titman. 1993. Returns to buying winners and selling losers: Implications
for stock market efficiency. The Journal of Finance 48: 65–91. [CrossRef]
Johnson, Richard Arnold, and Dean W. Wichern. 2008. Applied Multivariate Statistical Analysis. Beijing: Tsinghua
University Press.
Lars, Tvede. 2006. Business Cycles: History, Theory and Investment Reality. Hoboken: Wiley.
Lintner, John. 1965. The valuation of risk assets and the Selection of Risky Investments in Stock Portfolios and
Capital Budgets. The Review of Economics and Statistics 47: 13–37. [CrossRef]
Markowitz, Harry. 1952. Portfolio selection. Journal of Finance 7: 77–91.
Moinak, Maiti, and A. Balakrishnan. 2018. Is human capital the sixth factor? Journal of Economic Studies 45: 710–37.
Novy-marx, Robert. 2013. The other side of value: The gross profitability premium. Journal of Financial Economics
108: 1–28. [CrossRef]
Peng, Chi-Lu, Kuan-Ling Lai, Maio-Ling Chen, and An-Pin Wei. 2014. Investor sentiment, customer satisfaction
and stock returns. Social Science Electronic Publishing 49: 827–50. [CrossRef]
Ronzani, André, Osvaldo Candido, and Wilfredo Maldonado. 2017. Goodness-of-fit versus significance: A capm
selection with dynamic betas applied to the brazilian stock market. International Journal of Financial Studies 5:
33. [CrossRef]
Ross, Stephen A. 1976. The arbitrage theory of capital asset pricing. Journal of Economic Theory 13: 341–60. [CrossRef]
Sehgal, Sanjay, and Vibhuti Vasishth. 2015. Past price changes, trading volume and prediction of portfolio returns.
Journal of Advances in Management Research 12: 330–56. [CrossRef]
Sharpe, William F. 1964. Capital asset prices: A theory of market equilibrium under conditions of risk. The Journal
of Finance 19: 18.
Wild, John J. 2016. Fundamental Accounting Principles. New York: McGraw-Hill Education.
Zahedi, Javad, and Mohammad Mahdi Rounaghi. 2015. Application of artificial neural network models and
principal component analysis method in predicting stock prices on tehran stock exchange. Physica A:
Statistical Mechanics and its Applications 438: 178–87. [CrossRef]

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license ([Link]

222
 MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel. +41 61 683 77 34
Fax +41 61 302 89 18
[Link]

Journal of Risk and Financial Management Editorial Office

E-mail: jrfm@[Link]
[Link]/journal/jrfm
MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel: +41 61 683 77 34
Fax: +41 61 302 89 18
[Link] ISBN 978-3-03943-574-6