ESSAYS ON REGIME SWITCHING MODELS WITH

ENDOGENOUS FEEDBACK

Shi Qiu

Submitted to the faculty of the University Graduate School

in partial fulfillment of the requirements

for the degree

Doctor of Philosophy

in the Department of Economics,

Indiana University

July 2019




ProQuest Number: 13904477




All rights reserved

INFORMATION TO ALL USERS
The quality of this reproduction is dependent upon the quality of the copy submitted.

In the unlikely event that the author did not send a complete manuscript
and there are missing pages, these will be noted. Also, if material had to be removed,
a note will indicate the deletion.






ProQuest 13904477

Published by ProQuest LLC (2019 ). Copyright of the Dissertation is held by the Author.


All rights reserved.
This work is protected against unauthorized copying under Title 17, United States Code
Microform Edition © ProQuest LLC.


ProQuest LLC.
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, MI 48106 - 1346
Accepted by the Graduate Faculty, Indiana University, in partial fulfillment of the requirements for

the degree of Doctor of Philosophy.

Doctoral Committee

Joon Y. Park, Ph.D., Co-Chair

Yoosoon Chang, Ph.D., Co-Chair

Grey Gordon, Ph.D.

Ke-Li Xu. Ph.D.

May 2, 2019

ii
Copyright © 2019

Shi Qiu

iii
For Ling and the girls.

iv
Acknowledgments

I am incredibly grateful to all of my research committee for their guidance and support. I am

particularly thankful to my committee co-chairs, Professor Joon Y. Park and Professor Yoosoon

Chang, for all the effort and training they put into helping me become a successful econometrician,

communicator, and colleague. I am forever indebted to Professor Grey Gordon for the invaluable

research experience he shared as well as his helpful comments and suggestions. I thank Professor

Ke-Li Xu for the helpful job market information and his encouragement. I also want to thank

Professor Michael V. Alexeev, for granting me the opportunity to pursue graduate school at Indiana

University.

I am honored to have worked with Professor Joon Park and Professor Yoosoon Chang toward

the three chapters. They are terrific coauthors, colleagues, and mentors. I am grateful for the

technical support from Charlie Beima, Jefferson Davis, and the research computing support of

Indiana University. I also thank Professor Todd Walker and Professor Yoosoon Chang for their

excellent service as directors of graduate study.

Also, I would like to acknowledge the support of several other people who have been instrumental

in my graduate study. I thank Nastassia Krukava, Boreum Kwak, Peng Shen, and Fei Tan for their

friendship, advice, and invaluable support during my hardest time in graduate school. I also want

to recognize Paulette Davidson and Brandie Roberts for their dedication to the department of

v
economics and all the efforts to help graduate students.

Finally, I want to thank my parents for their patience and unconditional support, and for always

believing in me and helping me reach this goal. I am also deeply grateful for all the support that

my wife’s parents have provided. More than anyone else, I am forever thankful to my wife, Ling,

for the company of hundreds of sleepless nights, and everything she has provided me during this

challenging process. Without her support and tremendous sacrifice, I would never have finished

this work. I can not possibly thank her enough.

vi
 Shi Qiu

ESSAYS ON REGIME SWITCHING MODELS WITH ENDOGENOUS FEEDBACK

This dissertation consists of three essays concerning the dynamics of switching economic states

such as the expansions and recessions. These essays extend existing methodologies along several

dimensions and bring these novel tools to the empirical studies of the U.S. real business cycle,

monetary-fiscal policy interaction, and the effect of financial market uncertainty on the macroecon-

omy.

The first chapter introduces the switching states driven by a nonlinear latent factor whose

innovation correlates with the shock of observed time-series in the previous period. We examine

the U.S. GDP growth data with this new method and present evidence of nonlinearity in the factor

driving the expansion-recession cycle after 1984.

The second chapter introduces unsynchronized switching in a system of equations by allowing

several factors, each driving a separate state. The essay also presents a novel identification result

for the model. We estimate the U.S. monetary and fiscal policy rules with this new approach

and uncover evidence that the fiscal policy shock feeds the dovish-hawkish cycle of monetary

policy. The estimation further reveals economically significant monetary-fiscal policy coordination

at frequencies longer than 18 months.

The third chapter investigates the effects of time-varying financial market conditions on macroe-

conomic variables by extending a standard dynamic stochastic general equilibrium model (DSGE)

model to incorporate switching degrees of financial friction derived from switching uncertainty

process in a costly-state-verification problem. We emphasize the expectation effect of switching

financial condition. Transition probabilities influence agents’ choice through expectation effect:

vii
upon an adverse shock, a bleak outlook of the financial market causes slow recovery of investment.

The novelty of this paper is that we introduce feedback from past fundamental shocks to switching

dynamics through time-varying transition probability given explicitly as a function of these shocks.

Empirically, we uncover evidence of time-varying transition in the U.S. data and quantify the

contribution of each fundamental shock.

Joon Y. Park, Ph.D., Co-Chair

Yoosoon Chang, Ph.D., Co-Chair

Grey Gordon, Ph.D.

Ke-Li Xu. Ph.D.

viii
Contents

Chapter 1 Regime Switching Model with Nonlinear and Self-Excited Latent Dynamic

Factor 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 The Model and Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2.1 Regime Switching with STAR Latent Factor . . . . . . . . . . . . . . . . . 4

1.2.2 Computation of Stationary Distribution for ESTAR Process . . . . . . . . 8

1.3 Likelihood, Filtering, Smoothing and Forecasting . . . . . . . . . . . . . . . . . . 9

1.3.1 Likelihood Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.3.2 Basic Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.3.3 Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.3.4 Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.4 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

Chapter 2 Regime Switching Models with Multiple Dynamic Factors 19

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2 The Model and Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.2.1 A Model with Multiple Unsynchronized Regime Switchings . . . . . . . . 23

ix
 2.2.2 Transition Probabilities and Their Computations . . . . . . . . . . . . . . 25

2.3 A Modified Markov-Switching Filter . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.3.1 Likelihood Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.3.2 Extraction of Latent Factor . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.4 Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.4.1 Identification of Mixture Distribution . . . . . . . . . . . . . . . . . . . . 32

2.4.2 Identification of State Transition . . . . . . . . . . . . . . . . . . . . . . . 33

2.4.3 Identification of Endogenous Feedback . . . . . . . . . . . . . . . . . . . 38

2.5 Empirical Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.5.1 Interactions of U.S. Monetary and Fiscal Policies . . . . . . . . . . . . . . 40

2.5.2 Data and Identified Regimes . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.5.3 Estimation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2.6 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

2.6.1 Simultaneity Bias Reduction . . . . . . . . . . . . . . . . . . . . . . . . . 52

2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

Appendices 55

2.A A Parallelizable Algorithm for Likelihood Evaluation . . . . . . . . . . . . . . . . 55

2.B Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

2.B.1 Computation of Latent Factor . . . . . . . . . . . . . . . . . . . . . . . . 56

2.B.2 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

2.C Omitted Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

2.C.1 Extraction of Latent Factor . . . . . . . . . . . . . . . . . . . . . . . . . . 58

2.D Additional Tables and Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

x
Chapter 3 Expectation Effects of Switching Financial Frictions 61

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3.2 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

3.3 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

3.3.1 The Real Sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

3.3.2 The Financial Sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

3.3.3 Policy Rules and Aggregate Resource Constraint . . . . . . . . . . . . . . 77

3.3.4 Shocks Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

3.3.5 Regimes, Regime Factor and Feedback . . . . . . . . . . . . . . . . . . . 80

3.3.6 Agents’ Information Set . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.3.7 Solution Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.4 Effects of Switching Financial Conditions . . . . . . . . . . . . . . . . . . . . . . 85

3.4.1 Effects of Financial Friction in Fixed Regime . . . . . . . . . . . . . . . . 85

3.4.2 Switching Financial Conditions and Expectation Effect . . . . . . . . . . . 91

3.5 Time-Varying Outlook of Financial Market . . . . . . . . . . . . . . . . . . . . . 93

3.5.1 Data Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

3.5.2 Sub-Sample Evidence of Time-Varying Outlook of Financial Market . . . . 95

3.6 Sources of Time-Varying Uncertainty Outlook . . . . . . . . . . . . . . . . . . . . 98

3.6.1 Filtering Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

3.6.2 Priors and Posterior Modes . . . . . . . . . . . . . . . . . . . . . . . . . . 101

3.6.3 Close the Gap between Uncertainty and Spread . . . . . . . . . . . . . . . 105

3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

xi
Appendices 109

3.A Perturbation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

3.B Estimation, Optimization and Filtering Method . . . . . . . . . . . . . . . . . . . 113

3.B.1 Bayesian Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

3.B.2 Numerical Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

3.B.3 Chang, Maih and Tan Filter . . . . . . . . . . . . . . . . . . . . . . . . . 115

3.C Additions Figures and Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

Bibliography 128

Curriculum Vitae

xii
List of Tables

1.1 ML Estimates for US Real GDP Growth Rate . . . . . . . . . . . . . . . . . . . . 17

2.1 Maximum Likelihood Estimation (1949:Q2-2014:Q2) . . . . . . . . . . . . . . . . 46

2.2 Leading FP vs. Leading MP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

2.3 Selected Parameter Estimates for Error Component Model (1949:Q2-2014:Q2) . . 51

2.4 Relative Bias against DGP at Joint MLE . . . . . . . . . . . . . . . . . . . . . . . 53

2.D.1VAR(1) Fit for Extracted Latent Factor . . . . . . . . . . . . . . . . . . . . . . . . 60

3.1 List of Fundamental Shocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

3.2 List of Feedback Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

3.1 Calibrated Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

3.2 Priors and Posterior Mode: Fixed Regime . . . . . . . . . . . . . . . . . . . . . . 87

3.3 Simulation Setup of Expectation Effect . . . . . . . . . . . . . . . . . . . . . . . . 92

3.1 Priors and Posterior Mode: Switching Regimes, Constant Transition . . . . . . . . 97

3.1 Priors and Posterior Mode: Endogenous Switching . . . . . . . . . . . . . . . . . 103

3.2 Priors and Posterior Mode: Endogenous Switching, Cont’d . . . . . . . . . . . . . 104

3.C.1 CMR vs. Fixed Regime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

3.C.2 Calibrated vs. Estimated Adjustment Costs . . . . . . . . . . . . . . . . . . . . . 126

xiii
3.C.3 Calibrated vs. Estimated Adjustment Costs, Cont’d . . . . . . . . . . . . . . . . . 127

xiv
List of Figures

1.1 Contours of State Transition Probabilities . . . . . . . . . . . . . . . . . . . . . . 6

1.2 One-to-one Correspondence between Transition Matrix and (τ, λ) . . . . . . . . . 7

2.1 Historical Data and Estimated Policy Regimes . . . . . . . . . . . . . . . . . . . . 43

2.2 Extracted Latent Factor (MLE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

2.3 Impulse Responses of Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2.4 Comovements and Correlation of Extracted Latent Factors . . . . . . . . . . . . . 50

2.5 Magnitude-Squared Coherence of Extracted Latent Factors . . . . . . . . . . . . . 51

3.1 Credit Spread Against Idiosyncratic Uncertainty . . . . . . . . . . . . . . . . . . . 64

3.1 Normal vs. Large Uncertainty Shock in Fixed Regime . . . . . . . . . . . . . . . . 88

3.2 Low vs. High Steady State Uncertainty in Fixed Regime . . . . . . . . . . . . . . 89

3.3 Low vs. High Monitoring Cost in Fixed Regime . . . . . . . . . . . . . . . . . . . 90

3.4 Pessimistic vs. Optimistic Outlook in Switching Regimes . . . . . . . . . . . . . . 93

3.1 Credit Spread Against Idiosyncratic Uncertainty . . . . . . . . . . . . . . . . . . . 99

3.1 Credit Spread Against Uncertainty, Full Sample Endogenous Switching . . . . . . 106

3.2 Regime Factor and Identified High Uncertainty Regimes . . . . . . . . . . . . . . 107

3.C.1 Filtered Idiosyncratic Uncertainty: CMR against Pmode . . . . . . . . . . . . . . . 120

xv
3.C.2 Normal vs. Large Uncertainty Shock in Fixed Regime, Deviation . . . . . . . . . . 121

3.C.3 Low vs. High Steady State Uncertainty in Fixed Regime, Deviation . . . . . . . . . 122

3.C.4 Low vs. High Monitoring Cost in Fixed Regime, Deviation . . . . . . . . . . . . . 123

3.C.5 Credit Spread Against Idiosyncratic Uncertainty, Full Sample Exogenous Switching 124

3.C.6 High Uncertainty Regime Probability . . . . . . . . . . . . . . . . . . . . . . . . 128

3.C.7 Uncertainty, Exogenous vs. Endogenous . . . . . . . . . . . . . . . . . . . . . . . 129

xvi
Chapter 1

Regime Switching Model with Nonlinear and Self-Excited Latent Dynamic Factor

1.1 Introduction

Regime switching models have been used extensively by numerous researchers in various fields.

Since the seminal work by Hamilton (1989a), many authors, including Kim (1994, 2004, 2009),

Diebold et al. (1994), Chib (1996), Chib and Dueker (2004), Kim et al. (2008), Bazzi et al.

(2014), Kang (2014), Kalliovirta et al. (2015), and Chang et al. (2017), among many others, have

extended his regime switching model to make it more widely applicable. There also exists a large

literature, which studies various statistical properties of regime switching models such as Hansen

(1992), Hamilton (1996), Garcia (1998), Timmermann (2000), and Cho and White (2007). For an

introduction and overview of regime switching models, the reader is referred to the monograph by

Kim and Nelson (1999).

In this paper, we consider a regime switching model, where regimes are determined by a latent

dynamic factor depending upon whether it takes a value above or below some threshold level.

The latent factor is modeled as evolving in a nonlinear fashion and exciting itself by a shock

to the observed time series. More explicitly, we specify the latent factor as a smooth transition

autoregressive model (STAR) to allow for nonlinear dynamics. Moreover, to give a channel through

which the factor is self-excited, we let the innovation of the latent factor be correlated with the

1
innovation of the observed time series in the previous period. Our regime switching model may

be regarded as a generalization of the model introduced by Chang et al. (2017), which relies on a

linear autoregressive (AR) latent factor in place of our STAR latent factor.

Chang et al. (2017) show that their model fits the US GDP growth rates and the volatilites of SP

500 equity returns quite well. In particular, it was made clear that there exists a strong endogenous

feedback effect on regime switching of the US GDP growth rates and SP 500 equity returns. The

overall performance of their model allowing endogenous feedback is shown to be significantly

better than that of the conventional regime switching model relying on a completely exogenous

Markov chain. There are two common features of the regime switching models estimated by their

approach. First, the latent factor is highly persistent with the estimated coefficient typically close

to unity. Second, the endogenous feedback effect is very strong. It is indeed too strong, i.e., the

estimated correlation coefficient between the innovation of the latent factor and the innovation of

the observed time series is too close to unity, which implies that the latent factor is indeed driven

entirely by the innovation of the observed time series.

Their empirical findings support our use of the AR factor with nonlinear dynamics. It is well

known that an AR process with a nonlinear transition becomes hardly distinguishable from a unit

root process, if its nonlinearity is ignored and fitted by a linear AR model. See, e.g., Moreover, it is

highly probable that the observed perfect endogenous feedback may well be an artifact caused by

ignoring a nonlinear transition. In the paper, we specify the latent factor as an exponentially smooth

transition autoregressive (ESTAR) process, one of the most commonly used smooth transition

autogressive (STAR) processes. This is just for concreteness, and our methodology in the paper

can also be easily implemented for regime switching models driven by different classes of STAR

factors only with minor and obvious modifications. The STAR models, which may be interpreted

2
as the AR models with stochastic and state dependent autoregressive coefficients, generally provide

more flexible and realistic dynamics. The reader is referred to Teräsvirta for more discussions on

the STAR and other nonlinear AR models.

As shown in the paper, the performance of our regime switching model relying on an ESTAR

factor looks quite promising. Examination of U.S. GDP growth as in Chang et al. (2017) unveils

several pieces of new results. First, on all sample periods, the ESTAR coefficients are either close to

zero or statistically insignificant from zero, suggesting highly diffusive behavior of the latent factor

if close to zero. Second, on the sub-sample spanning from 1952-1984, the ESTAR model identifies

the model parameters with much smaller standard errors compared to the linear AR model. Third,

the on sub-sample spanning from 1984-2012, the ESTAR model identifies the regime feedback

parameter to be well within the unit interval, while the linear AR model estimates the parameter to

be on the boundary.

Finally, a word on notation. Throughout the paper, we denote by φ(·; µ, σ 2 ) the normal density

function with mean µ and standard deviation σ, and by Φ(·; µ, σ 2 ) the corresponding distribution

function. By convention, we will simply write φ(·; 0, 1) = φ(·) and Φ(·; 0, 1) = Φ(·). For a time

series (zt ), we let Za:b = (z a, z a+1, · · · , z b )0 for any 0 ≤ a < b.

1.2 The Model and Preliminaries

In this section, we introduce our regime switching model and some preliminaries required to

understand our subsequent theory and methodology presented in later sections.

3
1.2.1 Regime Switching with STAR Latent Factor

The typical regime switching model for a time series (yt ) is given as

yt = m(St−k:t, Yt−k:t−1, x t ; π) + σ(St−k:t, x t ; π)ut, (1.2.1)

where m and σ are known functions up to unknown parameter π, (st ) is a binary state variables

taking values 0 and 1, (x t ) is an exogenous variable assumed to be predictable, and (ut ) is an i.i.d.

error term. In the paper, we further specify (1.2.1) by setting

st = 1 wt ≥ τ ,

(1.2.2)

whether τ is an unknown threshold level, and (wt ) is a latent factor generated as

wt = f (wt−1 ; λ) + vt (1.2.3)

with
 
f (w; λ) = w exp −λ 2 w 2

and i.i.d. innovation (vt ). Here we assume that

*. ut +/ = N *.0, *. 1 ρ +/+/ (1.2.4)

d
, vt+1 - , , ρ 1 --

for ρ ∈ (−1, 1).

Thf regime switching model given by (1.2.1) includes a very large class of models that have

4
been proposed and analyzed by many authors. The model considered initially by Hamilton (1989a),

and subsequently by many others, is

γ(L)(yt − µt ) = σut, (1.2.5)

Pk i, γ
where γ(z) = i=0 γi z 0 = 1 is a k th order polynomial in z, σ is the volatility of the innovation

(ut ), and µt = µ(1 − st ) + st µ, with some unknown parameters µ and µ, denotes the state dependent

mean of (yt ). The model for the state dependent volatility is similarly given as

yt = σt ut (1.2.6)

with the state dependent volatility σt = σ(1 − st ) + st σ specified by some unknown parameters σ

and σ. Such a model has been considered by xxx, among many others. The model

k
X
yt = αt−i yt−i + βt x t + σut (1.2.7)
i=1

has also been used, where αt and βt are in particular defined to be state dependent and given as

functions of st similarly as above. See, for instance, Chang and Kwak (2017).

The regime switching given by (1.2.2) and (1.2.3) is just an equivalent reformation of the

conventional Markov switching. More precisely, there is a one-to-one correspondence between

them if ρ = 0 in (1.2.4), in which case the state variable (st ) becomes strictly exogenous also in

our model as in the conventional Markov switching model. This will be shown below. Our novel

formulation of regime switching is primarily introduced to allow for the feedback effect that makes

self-excitation possible in regime switching, via the nonzero correlation ρ between the innovation

5
(ut ) of the observed time series and the innovation (vt+1 ) in the next period of the latent factor (wt ).

One important advantage of our formulation of regime switching is that it provides an operational

notion of the strength of regime, which is measured by how far (wt ) is away from the threshold level

τ at any given period of time t. In this sense, our formulation of regime switching may be preferred

even for the case of ρ = 0, when it becomes equivalent to the conventional Markov switching.

For the innovations (ut ) and (vt ), we assume joint normality and set their variances to be unity for

identification.

To demonstrate that there is a one-to-one correspondence between our regime switching model

with the conventional Markov switching model, we present the contour plots of low-low and

high-high transition probabilities across pairs of (τ, λ) in Figure 1.1. As further demonstrated in

6 6
99
0.

4 4
0.99
0.1
0.2
2 0.9 0.9 2 0.1 0.3
0.4
0.8 0.2 0.5
0.7 0.3 0.6
0.4
0.6 0.5 0.7
0 0
0.5 0.6
0.4
0.7 0.8
0.3
0.2 0.8

-2 0.1 -2 0.9
0.9

0.99
0.8

-4 -4 0.99
0.7
0.
0.
00.4
65
0.2.1
.3
0

0.9

-6 -6
3 2.5 2 1.5 1 0.5 0 3 2.5 2 1.5 1 0.5 0

(a) Low-Low (b) High-High

Figure 1.1: Contours of State Transition Probabilities

Figure 1.2, it is clear that each pair of (τ, λ) pins down uniquely a transition probability matrix,

provided ρ = 0.

The latent factor (wt ) defined in (1.2.3) follows the so-called exponential smooth transition

6
 6

0.901
2

0.796 0.796

(0.126,-1.549)

-2 0.901 0.901

-4

0.796
-6
3 2.5 2 1.5 1 0.5 0

Figure 1.2: One-to-one Correspondence between Transition Matrix and (τ, λ)

autoregressive (ESTAR) process, which is one subclass of processes that are more broadly defined

as smooth transition autoregressive (STAR) processes. The general STAR process is defined as

wt = λ t wt−1 + vt,

where λ t is given as a smooth function of wt−1 such that |λ t | < 1. It may naturally be viewed

as a stationary AR process with a time-varying and stochastic AR coefficient λ t . As well known,

the STAR models give rich dynamics while retaining their stationarity. The reader is referred

to Terasverta for a general introduction and more detailed discussions on the STAR models. In

particular, they are very effective in modeling stationary, yet highly persistent, Markovian dynamics,

as demonstrated by, e.g., Park and Shintani (2016).

 
For our latent factor, we set λ t = exp −λ 2 wt−1
2 . Under this specification, λ t → 0 as wt−1 →

7
±∞, marking strong mean-reverting behavior as latent factor (wt ) moves away from 0, and λ t → 1

as wt−1 → 0, marking diffusive behavior of (wt ) around 0. If ρ = 0 and the latent factor is strictly

exogeneous, the specification of latent factor in our model becomes unimportant. In fact, in this

case, a large class of models of latent factor yields the regime switching models that are equivalent

to the conventional Markov switching models. The specification of latent factor only changes how

the innovation ut of the observed time series in the current period affects the latent factor wt+1 in

the next period through a nonzero correlation between ut with vt+1 .

1.2.2 Computation of Stationary Distribution for ESTAR Process

To estimate our regime switching model by the maximum likelihood (ML) method, we need to

compute the stationary distribution of ESTAR process with all values of the parameter λ > 0.

The stationary distribution Ψλ of (wt ) in ESTAR specification (1.2.3) plays a critical role in

the estimation and filtering method in our setup. The analytic form of Ψλ is unknown. It is obvious

that ψ∞ is the density of standard normal distribution. Moreover, it is evident that Ψλ is symmetric

around zero for all positive λ. In what follows, we devise an approximation of Ψλ .

We approximate the density function ψ λ by linear combination of zero-mean normal densities

with different variances in light of its symmetry. This approach is slightly different from Alspach

and Sorenson (1972) since some of our coefficients will be inevitably negative. Note by definition

of stationary distribution

Z ∞
ψ λ (x) = φ(x − f (y; λ); 0, 1)ψ λ (y)dy.
−∞

The method can be described as follows:

8
 Pn Pn
1. Write ψ λ ≈ k=1 a k φ k such that k=1 a k = 1, and (φ k )’s are normal densities φ(x; 0, σ 2k )

for a prescibed sequence of {σ k }.

2. Estimate the vector of coefficients

n "
X Z ∞ #2
â = arg min{(ak ) n } a k φ k (x) − a k φ(x − f (y; λ))φ k (y)dy
k=1
k=1 −∞

on a sufficiently fine grid of test points X.

The choices of sequence {σ k }, and test points X are critical to the precision of approximated

stationary density. Our simulation result suggests that both the proper sequence of {σ k } and the

test points X are contingent on the value of λ.

1.3 Likelihood, Filtering, Smoothing and Forecasting

In this section, we describe the method of obtaining likelihood, filtering and smoothing. The

obtained likelihood enables maximum likelihood (ML) and Bayesian estimation.

For the ease of notation, we denote θ in this section to be the vector of all parameters and Θ to

be the parameter space. In addition, we suppress the exogenous variable x t and denote through out

this section that

f λ (x) = f (x; λ) (1.3.1)

mt = m(St−k,t, Yt−k:t−1 ; π) (1.3.2)

σt = σ(St−k:t ; π) (1.3.3)

9
1.3.1 Likelihood Function

It is shown in Chang et al. (2017) that ρ has nontrivial impact on dynamics of latent factor (wt ) and

state dependent variable (yt ). The joint transition of state process (st ) and state dependent variable

(yt ) is fully specified in light of the following proposition.

Proposition 1.3.1 If | ρ| < 1, the bivariate process (yt, st ) on R × {0, 1} is a (k + 1) st order Markov

process whose transition density

p(yt, st |St−k−1:t−1, Yt−k−1:t−1 ; θ) = p(yt |St−k:t, Yt−k:t−1 ; θ)p(st |St−k−1:t−1, Yt−k−1:t−1 ; θ)

with respect to the product of the counting and Lebesgue measure. Moreover,

p(yt |St−k:t, Yt−k:t−1 ; θ) = φ(yt ; mt, σt ) (1.3.4)

p(st |St−k−1:t−1, Yt−k−1:t−1 ; θ) = (1 − st )ω ρ,λ (St−k−1:t−1, Yt−k−1:t−1 ) (1.3.5)

+st [1 − ω ρ,λ (St−k−1:t−1, Yt−k−1:t−1 )]

with the transition probability ω ρ,λ of endogenous state process (st ) to low state given by

Z
1st−1 (x)Φ ρ τ − f λ (x) − ρ yt−1σ−m
  
t−1
t−1
dΨλ (x)
ω ρ,λ (St−k−1:t−1, Yt−k−1:t−1 ) = (1.3.6)
[Ψλ (τ)]1−st−1 [1 − Ψλ (τ)]st−1

in which Φ ρ (x) = Φ(x/ 1 − ρ2 ).

p

Proof of Proposition 1.3.1 One may adapt to the proof of Theorem 3.1 in Chang et al. (2017)

via (i) replacing the standard invariant distribution by our non-standard invariant distribution Ψλ ,

10
and (ii) replacing the linear transition equation by the nonlinear transition equation (1.2.3). QED.

The log-likelihood function can be written as

T
X
`(y1, · · · , yT |θ) = log p(y1 |θ) + log p(yt |Ft−1 ; θ) (1.3.7)
t=2

in which Ft = σ(Y1:t ). It can be evaluated recursively. One may write the conditional density

X
p(yt |Ft−1 ; θ) = p(yt |St−k:t, Ft−1 ; θ)p(St−k:t |Ft−1 ; θ)
St−k:t
X
= p(yt |St−k:t, Yt−k:t−1 ; θ)p(St−k:t |Ft−1 ; θ) (1.3.8)
St−k:t

and note that p(yt |St−k:t, Yt−k:t−1 ; θ) is given by (1.3.4). To evaluate (1.3.8), we evaluate recursively

through prediction and updating steps:

1) Prediction:

X
p(St−k:t |Ft−1 ; θ) = p(st |St−k−1:t−1, Ft−1 ; θ)p(St−k−1:t−1 |Ft−1 ; θ). (1.3.9)
st−k−1

2) Updating:

p(yt |St−k:t, Ft−1 ; θ)

p(St−k:t |Ft ; θ) = p(St−k:t |Ft−1 ; θ) (1.3.10)
p(yt |Ft−1 ; θ)

Finally, the log-likelihood function (1.3.7) can be evaluated by taking summation of all log-

conditional likelihoods.

11
1.3.2 Basic Filter

We can extract E(wt |Ft ; θ) recursively. For each t = 1, · · · , T,

p(yt |wt, St−k:t−1, Ft−1 ; θ)p(wt, St−k:t−1 |Ft−1 ; θ)

p(wt, St−k:t−1 |Ft ; θ) = , (1.3.11)
p(yt |Ft−1 ; θ)

in which p(yt |wt, St−k:t−1, Ft−1 ; θ) is given by (1.3.4), and

X
p(wt, St−k:t−1 |Ft−1 ; θ) = p(wt |St−k−1:t−1, Ft−1 ; θ)p(St−k−1:t−1 |Ft−1 ; θ). (1.3.12)
st−k−1

Adapting corollary 3.3 of Chang et al. (2017),

Z
1st−1 (x)φ ρ wt − f λ (x) − ρ yt−1σ−m
  
t−1
t−1
dΨλ (x)
p(wt |St−k−1:t−1, Ft−1 ; θ) = , (1.3.13)
[Ψλ (τ)]1−st−1 [1 − Ψλ (τ)]st−1

in which φ ρ denotes the probability density function of N(0, 1 − ρ2 ). Marginalizing (1.3.11) with

respect to St−k:t−1 gives p(wt |Ft ; θ). And the latent factor can be extracted as

Z
E(wt |Ft ; θ) = wt p(wt |Ft ; θ)dwt .

12
1.3.3 Smoothing

A smoothing filter for latent factor wt can be devised utilizing products of the basic filter. The

smoothing is conducted recursively backward in direction. First note

p(wT |FT ; θ) (1.3.14)

p(wt, St−k:t−1 |Ft ; θ), t = 1, · · · , T (1.3.15)

p(wt+1, St−k+1:t |Ft ; θ), t = 0, · · · , T − 1 (1.3.16)

can be stored while performing basic filtering. Also, by transition equation (1.2.3)

yt − mt
! !
p(wt+1 |wt, St−k:t−1, Yt−k:t ; θ) = φ wt+1 ; f λ (wt ) + ρ ,1− ρ .
2
(1.3.17)
σt

For each t = 1, · · · , T − 1,

Z
p(wt, St−k:t−1 |FT ; θ) = p(wt, St−k:t−1, wt+1 |FT ; θ)dwt+1
Z
= p(wt+1 |FT ; θ)p(wt, St−k:t−1 |wt+1, FT ; θ)dwt+1
Z
= p(wt+1 |FT ; θ)p(wt, St−k:t−1 |wt+1, Ft ; θ)dwt+1

p(wt, St−k:t−1 |Ft ; θ)p(wt+1 |wt, St−k:t−1, Ft ; θ)

Z
= p(wt+1 |FT ; θ) dwt+1
p(wt+1 |Ft ; θ)
p(wt+1 |FT ; θ)p(wt+1 |wt, St−k:t−1, Ft ; θ)
Z
= p(wt, St−k:t−1 |Ft ; θ) dwt+1,
p(wt+1 |Ft ; θ)
(1.3.18)

in which the first equality holds by marginalizing joint density of (wt, wt+1, St ) with respect to

13
wt+1 , the second equality holds by definition of conditional probability, the third equality holds by

equation (1.2.3) and independence of vt+1 and yt+1, · · · , yT implied by (1.2.4), and the rest is given

by Bayes rule and algebra.

Marginalizing the joint density of wt and St−k:t−1 gives

X
p(wt |FT ; θ) = p(wt, St−k:t−1 |FT ; θ).
St−k:t−1

Finally,
Z
E(wt |FT ; θ) = wt p(wt |FT ; θ)dwt .

1.3.4 Forecasting

In forecasting, we aim to obtain conditional expectations

E(yT+q |FT ; θ), q = 1, 2, · · ·

E(yT+q |ST−k:T , FT ; θ).

It is unclear how to forecast with general measurement equation. With measurement equation

(1.2.5), it amounts to obtaining conditional probabilities

P(sT+q = 1|FT ; θ), q = 1, 2, · · ·

P(sT+q = 1|ST−k:T , FT ; θ).

14
To this end, we start with products of the basic filter

p(wT+1 |FT ; θ),

p(wT+1 |ST−k:T , FT ; θ),

and compute recursively for q = 2, 3, · · ·

(1) Given p(wT+q−1 |ST−k:T , FT ; θ),

Z
p(wT+q |ST−k:T , FT ; θ) = p(wT+q |wT+q−1, ST−k:T , FT ; θ)p(wT+q−1 |ST−k:T , FT ; θ)dwT+q−1,

in which

p(wT+q |wT+q−1, ST−k:T , FT ; θ) = p(wT+q |wT+q−1 ; θ) = φ(wT+q ; f λ (wT+q−1 ), 1),

by the fact that vT+q being independent of (ST−k:T , FT ).

(2) Given p(wT+q |ST−k:T , FT ; θ),

Z ∞
P(sT+q = 1|ST−k:T , FT ; θ) = p(wT+q |ST−k:T , FT ; θ)dwT+q .
τ

(3) Given p(wT+q−1 |FT ; θ), by similar argument in (1) and (2),

Z
p(wT+q |FT ; θ) = p(wT+q |wT+q−1 ; θ)p(wT+q−1 |FT ; θ)dwT+q−1,
Z ∞
P(sT+q = 1|FT ; θ) = p(wT+q |FT ; θ)dwT+q .
τ

15
1.4 Application

We demonstrate our method by examining the US real GDP growth rate. In this section, we perform

ML estimation using global numerical optimization routine DIRECT-L (Dividing Rectangles with

local search) and extract latent factor using our filter with ESTAR latent factor. Moreover, we

compare the estimates and extracted latent factors with respect to ESTAR and AR latent factor.

The data set consists of the seasonally adjusted quarterly real GDP growth rate from 1952:Q1

to 2012:Q4. The measurement equation takes form (1.2.5) with k = 4 as in Hamilton (1989b).

We compare our ML estimates to Chang et al. (2017)’s (CCP, hereafter) assuming endogenous

regime switching. Table 1.1 reports this comparison. In particular, CCP found with post-1984

sample that ρ = 0.999. We find ρ = 0.476. There is not a direct comparison between the ESTAR

coefficient and CCP’s AR coefficient. The ML estimates for full sample are also reported.

16
 Table 1.1: ML Estimates for US Real GDP Growth Rate

ESTAR Estimates CCP’s AR Estimates

Parameters 1952-2012 1952-1984 1984-2012 1952-1984 1984-2012
µ -0.003 -0.061 -0.037 -0.083 -0.758
(0.081) (0.051) (0.557) (0.161) (0.311)
1.125 1.289 0.700 1.212 0.705
µ (0.014) (0.012) (0.092) (0.095) (0.085)
0.550 0.744 0.226 0.784 0.452
σ (0.006) (0.013) (0.045) (0.057) (0.032)
0.333 0.333 0.185 0.147 0.169
γ1 (0.014) (0.015) (0.120) (0.104) (0.105)
0.086 0.007 0.358 0.044 0.340
γ2 (0.006) (0.009) (0.104) (0.096) (0.103)
-0.111 -0.170 -0.086 -0.260 -0.076
γ3 (0.006) (0.010) (0.110) (0.090) (0.128)
-0.034 -0.081 0.004 -0.067 0.049
γ4 (0.009) (0010) (0.117) (0.096) (0.112)
-0.667 -0.658 -1.999 -0.758 -2.782
τ (1.207) (0.000) (0.358) (0.883) (1.144)
0.071 0.110 1.460 – –
λ (0.013) (0.002) (9.478) – –
-0.788 -0.931 0.476 -0.923 0.999
ρ (15.601) (0.202) (0.194) (0.151) (0.012)
p-value 0.0233 0.0288 0.0018
log-likelihood -293.19 -179.25 -81.04
Note: The p-value reports the likelihood ratio test for our model of nonlinear latent factor
with respect to H0 : ρ = 0. CCP’s estimated AR coefficients are 0.927(0.041) for pre-84
sample and 0.809(0.145) for post-84 sample.

17
1.5 Conclusion

This paper introduces the regime switching driven by a latent factor, which has two important novel

characteristics: nonlinear dynamics and self-excitation. For the former, we specify the latent factor

to be generated by a STAR model. For the latter, we set the innovation of the factor to be correlated

with the innovation of the observed time series in the previous period.

18
Chapter 2

Regime Switching Models with Multiple Dynamic Factors

2.1 Introduction

Regime switching models have been widely used and proven fruitful in many different contexts.

Hamilton (1989a) introduced the autoregressive model with regime dependent mean, which is

switched between high and low levels by a Markov chain independent of the innovation in the

underlying autoregressive model. The model was analyzed subsequently by Kim (1994). Since

then, the literature on regime switching models has been extended to a broader class of models,

particularly by introducing endogeneity in model variables and/or regime switchings, and to allow

for time-varying transition probabilities. See Kim (2004, 2009), Diebold et al. (1994), Chib (1996),

Chib and Dueker (2004), Kim et al. (2008), Bazzi et al. (2014), Kang (2014), Kalliovirta et al.

(2015), and Chang et al. (2017), among others. For an overview of early developments in regime

switching models, the reader is referred also to the monograph by Kim and Nelson (1999). Various

statistical properties of regime switching models are studied in Hansen (1992), Hamilton (1996),

Garcia (1998), Timmermann (2000), and Cho and White (2007).

In this paper, we introduce a new class of regime switching models, where regimes are deter-

mined by a latent multivariate state variable generated as a first order vector autoregressive process,

depending upon whether each state variable takes a value above or below some threshold. There

19
are several important novel features in our model. First, it allows for the existence of multiple

unsynchronized switchings, each of which is driven by the corresponding state variable. The state

variables may interact with each other dynamically via their vector autoregressive structure, as well

as statically due to nonzero correlation present in their innovations. Second, the innovations of state

variables in the current period are specified as being correlated with the innovations of observed

variables in previous period. This creates a channel for an endogenous feedback effect on regime

switchings in our model. Finally, the latent state variable in our model not only determines the

regimes but also designates the strength of regimes.

All novel features discussed above are not present in the conventional regime switching model.

In the model, regimes are determined by Markov chains that are independent of all other parts of the

model, and therefore, there is no endogeneity in regime switchings. Regimes are switched entirely

exogenously. Moreover, introducing multiple unsynchronized switchings in the conventional regime

switching model is not meaningful at all. In its framework, multiple unsynchronized switchings

by multiple Markov chains are observationally equivalent to switchings by a single Markov chain

with a properly expanded state space. For instance, the set of two Markov chains generating two

unsynchronized switchings between two states has the same implication on likelihood function as a

single Markov chain generating switchings among four states. Therefore, it is clear that we cannot

analyze the interactions of Markov chains driving those multiple unsynchronized switchings in any

meaningful manner. Needless to say, there is no information on the strength of regimes reflected in

the conventional regime switching model.

The main feature of our model is also not shared by other existing regime switching models.

The existing literature on regime switching models introduce endogeneity either by defining a static

latent variables correlated with the innovations of observed variables as in Kim et al. (2008), or by

20
specifying transition probabilities directly as some ad hoc functions of an endogenous variable as

in Diebold et al. (1994). In particular, even if we expand them to allow for the presence of multiple

unsynchronized switchings, none of them has any structure that could possibly be informative about

the interactions of multiple unsynchronized switchings, neither dynamic nor static. Therefore, our

model is very unique in that it can be used to study the interactions among multiple unsynchronized

switchings, especially in a very general and realistic setting. We believe that our model is widely

useful in analyzing the interactions of policies or cycles, such as the interactions of monetary-fiscal

policies and business-financial cycles, which may be thought of being generated by switching

regimes.

As an illustration, we analyze an extension of Chang and Kwak (2017) to study U.S. monetary-

fiscal policy interactions. On top of rediscovering substantial size of within-equation feedback,1 we

report evidence of non-trivial synchronization between monetary and fiscal factors, and non-trivial

cross-equation feedback. Specifically, we report three primary results. First, the impact of fiscal

factor on monetary factor is twice in size compared to the impact of monetary factor on the fiscal

factor. Second, a common factor that drives the shocks of both policy rules. Third, the common

factor is fed mainly to next period monetary factor innovation, whereas the fiscal factor is mainly

influenced by its equation-specific policy shocks. We also find the extracted latent factors from our

model to be highly positively correlated over the course of history with occasional exceptions in

the 1950s, 1970s and early 2000s. The coherence of extracted latent factors suggests a substantial

correlation between policy factors over the typical business cycle frequency. These results indicate

1Chang et al. (2017) document ubiquitous and robust feedback in both macro and financial time-series. Chang
and Kwak (2017) also report strong feedback in the series of monetary and fiscal policy instruments.

21
a tighter connection between the monetary and fiscal regimes than those reported by Chang and

Kwak (2017).

We develop a modified Markov switching filter, which can be used to estimate our regime

switching model by the maximum likelihood (ML) method. The filter yields the ML estimates of

parameters and extracts the latent factors in our model. It is simple to implement and computa-

tionally efficient. Through simulation, we compare our approach with that of Chang et al. (2017),

which may be regarded as a single switching version of ours assuming no interactions among

multiple nonsynchronized switchings. More explicitly, in a bivariate regression model with two

unsynchronized regime switchings, one for each regression, we compute the bias and inefficiency

incurred by the use of a single equation method, which neglects the interaction of the switching

in the equation with the other switching in the other equation. Negligence of interactions between

cross equation switchings inflicts substantial bias particularly for the threshold values, the size of

interaction and feedback channels, and the correlation of shocks.2 We also find that standard errors

of parameter estimates are in general reduced, substantially in some cases, if we take the interactions

of two switchings into consideration and estimate two equations jointly.3 In sum, our simulation

shows that it is important to take into consideration the interactions of multiple unsynchronized

switchings in modeling and estimating regime switching models.

The rest of the paper is organized as follows. Section 2.2 introduces our model with regime

2On the other hand, relatively small bias is observed for the regression coefficients, and the regimes implied by the
extracted latent factors do not change much in timing and length, even if interactions are neglected.

3We obtain standard errors of parameter estimates by simulation since it is computationally challenging to evaluate
Hessian matrix of the log-likelihood function.

22
switching, and presents some probabilistic characteristics implied by our regime switching rule,

which are necessary to get the transition probabilities and likelihood functions given by our model.

Section 2.3 develops a modified Markov filter and obtains the likelihood function of our model. In

Section 2.4, we discuss the identifiability of our model. Section 2.5 examines, as an empirical illus-

tration, a reduced form bivariate system of monetary and fiscal policy rules and analyze the policy

regime interactions. Section 3.4 reports our simulation results and supplements empirical section

by evaluating the performance of our approach considering interactions of multiple unsynchronized

switchings relative to an approach neglecting such interactions. Finally, Section 2.7 concludes the

paper. Appendix collects computational details, omitted proofs, and additional tables and figures.

2.2 The Model and Preliminaries

This section introduces a model with multiple unsynchronized regime switchings. It also derives

the transition probabilities of our regime switchings and shows how to compute the transition

probabilities numerically.

2.2.1 A Model with Multiple Unsynchronized Regime Switchings

We consider a bivariate regression model

y1t = x 01t β1t + σ1u1t

y2t = x 02t β2t + σ2u2t

23
with exogenous variables (x t ), x t = (x 1t, x 2t )0, error terms (ut ), ut = (u1t, u2t )0, and switching

parameters

βit = β (1 − sit ) + βi sit

i

σit = σi (1 − sit ) + σi sit

for i = 1, 2, where (sit ) for i = 1, 2 are binary state variables taking values 0 and 1, signifying

respectively the low and high states. The state variables (sit ) are generated as

sit = 1{wit ≥ τi }

with some thresholds τi for i = 1, 2, and latent factors (wt ), wt = (w1t, w2t )0, evolving as a VAR(1),

i.e.,

wt = Awt−1 + vt

with error terms (vt ), vt = (v1t, v2t )0, and autoregressive coefficient matrix

α11 α12 +
A = *. /.
, α21 α22 -

We assume that both of the eigenvalues of A lie inside unit disk, so that (wt ) becomes stationary.

Finally, we further specify

(ut0, vt+1
0
)0 =d N(0, P)

24
with
*. 1 +/
Puu Puv +/ = ... ρu2 u1
. //
1
P = *. // ,
, Pvu Pvv - .. ρv1 u1 ρv1 u2 1
. //
. //
, ρv2 u1 ρv2 u2 ρv2 v1 1 -

and assume that (ut ) and (vt ) are i.i.d. normals that are are independent of each other at all other

leads and lags.

If Pvu , 0, we say that there is endogenous feedback in our regime switching. Otherwise,

the regime switching is said to be exogenous. Note that ut is correlated with vt+1 , not vt , for all

t = 1, 2, . . .. In our specification of P, we set the variance of (uit ) and (vit ) to be unity. This is

necessary to identify our model, since in particular the pairs (τ, Pvv ) and (cτ, cPvv ) for any c > 0

imply identical state transition probabilities of (st ).

2.2.2 Transition Probabilities and Their Computations

If we let

zt = vt − Pvu Puu
−1
ut−1,

it follows that zt is independent of ut−1, ut−2, . . ., and of vt−1, vt−2, . . ., and therefore, it is also

independent of wt−1, wt−2, . . .. Moreover, we have

zt =d N (0, Pvv·u ) ,

where

Pvv·u = Pvv − Pvu Puu

−1
Puv, (2.2.1)

25
is the conditional variance of vt+1 given ut for each t = 1, 2, . . .. Therefore, we may deduce that

P wt < τ wt−1, yt−1, x t−1



( )
= P zt < τ − Pvu Puu
−1
ut−1 − Awt−1 wt−1, yt−1, x t−1
 
= Φv|u τ − Pvu Puu
−1
ut−1 − Awt−1 , (2.2.2)

where Φv|u is the distribution function of bivariate normal distribution with covariance matrix Pvv·u

defined in (2.2.1).

It follows from (2.2.2) that

P wt < τ wt−1 < τ, yt−1, x t−1


,
= P {wt < τ, wt−1 < τ|yt−1, x t−1 } P wt−1 < τ yt−1, x t−1


"Z τ   #,
= Φv|u τ − Pvu Puu
−1
ut−1 − Awt−1 φw (wt−1 )dwt−1 Φw (τ), (2.2.3)
−∞

where Φv|u is defined as above in (2.2.1), and φw and Φw are respectively the density and distribution

functions of (wt ), which is bivariate normal distribution with covariance matrix Pww , which is given

by

vec Pww = (I − A ⊗ A) −1 vec Pvv,

where vec(·) is the operator stacking rows of a matrix and transforming it into a column vector.

The transition probability (2.2.3) is time-invariant if Pvu = 0. In this case, our model reduces to

the conventional Markov switching model after relabeling the states (0, 0)0, (0, 1)0, (1, 0)0 and (1, 1)0

to be 1, 2, 3 and 4.

26
 For the actual computations of transition probabilities, we need to calculate

Z c2 Z c1
N ( A, b, c, Σ1, Σ2 ) = Φ1 (b − Ax) φ2 (x)dx 1 dx 2 (2.2.4)
−∞ −∞

for various sets of values of 2 × 2 matrix A, and two dimensional vectors b and c, where Φ1 and φ2

are respectively the bivariate normal distribution and density functions with covariance matrices

Σ1 = Pvv·u and Σ2 = Pww . For this purpose, we note that

 * X1 + * c1 + * Y1 +

 

N ( A, b, c, Σ1, Σ2 ) = P  , b AX ,

. / ≤ . / . / ≤ −
 X2
- , c2 - , Y2 -
 

, 

where X = (X1, X2 )0 is a bivariate normal random variate with mean zero and covariance matrix

given by Σ2 , and Y = (Y1, Y2 )0 is also a bivariate normal random variate, independent of X, with

mean zero and covariance matrix Σ1 . However, we may rewrite

 * X1 + * c1 + * Y1 +

 

, b AX  = P{X ≤ c, Y + AX ≤ b},

P . / ≤ . / . / ≤ −
X c Y

 

, 2 - , 2 - , 2 - 

and therefore, it follows that

Z b2 Z b1 Z c2 Z c1
N ( A, b, c, Σ1, Σ2 ) = p A,Σ1,Σ2 (z1, z2, z3, z4 )dz1 dz2 dz3 dz4,
−∞ −∞ −∞ −∞

where p A,Σ1,Σ2 is the density function of a four-dimensional normal random variate with zero mean

and covariance matrix

*. Σ2 Σ2 A0 +/
, AΣ2 Σ1 + AΣ2 A -
0

27
for given Σ1 and Σ2 .

Note that

Z ∞ Z c1
Φ1 (b − Ax) φ2 (x)dx 1 dx 2
c2 −∞

c1 +
= N ( A, b, *. / , Σ1, Σ2 ) − N ( A, b, c, Σ1, Σ2 ),
, ∞ -
Z c2 Z ∞
Φ1 (b − Ax) φ2 (x)dx 1 dx 2
−∞ c1

∞ +
= N ( A, b, *. / , Σ1, Σ2 ) − N ( A, b, c, Σ1, Σ2 ),
, c2 -
Z ∞Z ∞
Φ1 (b − Ax) φ2 (x)dx 1 dx 2
c2 c1

∞ + ∞ + c1 +
= N ( A, b, *. / , Σ1, Σ2 ) − N ( A, b, *. / , Σ1, Σ2 ) − N ( A, b, *. / , Σ1, Σ2 )
, ∞ - , c2 - , ∞ -

+ N ( A, b, c, Σ1, Σ2 ),

which can also be easily obtained, once we compute the integral in (2.2.4).4

2.3 A Modified Markov-Switching Filter

In this section, we characterize the likelihood function and extract latent factors by developing a

modified Markov-switching filter.

4Calculation of (2.2.4) involves non-trivial randomness, since efficient implementations such as the Matlab (Genz,
1992) employ Monte-Carlo integration techniques. The issue can be practically resolved by fixing the random seed of
Monte-Carlo integration. The disadvantage is that we lose control of precision. But our limited experience suggests
the resulting precision is roughly to the level of 10−4 .

28
2.3.1 Likelihood Function

Let information set F0 = ∅ and Ft = σ y1:t, x 1:T for t = 1, · · · , T and suppress x t in all future

notation for it is exogenous and will not change our arguments. The likelihood function for a vector

of parameters θ ∈ Θ takes form

YT X
`(θ) = p(Y1:T |θ) = *. p(yt |st, Ft−1 )p(st |Ft−1 ) +/ (2.3.1)
t=1 , st -

Equation (2.3.1) can be evaluated sequentially with following predict-update recursion. An equiv-

alent but parallelizable algorithm is proposed in Appendix Section 2.A.

Algorithm 2.3.1 (Modified Markov-Switching Filter)

1. Initialization. If t = 0, set p(st ) to be the unconditional state probabilities such that

P {s0 = (0, 0)0 } = P w1,0 < τ1, w2,0 < τ2 , in which (w1,0, w2,0 )0 ∼ N(0, Σww ).


2. Recursion. If t ≥ 1, repeat steps (a) - (c). Stop if t = T + 1.

(a) Forecasting. Predict the distribution of st given information set Ft−1

X
p(st |Ft−1 ) = p(st |st−1, Ft−1 )p(st−1 |Ft−1 )
st−1

with time-varying transition probability p(st |st−1, Ft−1 ) characterized by (2.2.3).

(b) Evaluation. Calculate the conditional density of yt given information set Ft−1

X
p(yt |Ft−1 ) = p(yt |st, Ft−1 )p(st |Ft−1 ).
st

29
 (c) Updating. Update the distribution of st given information set Ft

X
p(st |Ft ) = p(st, st−1 |yt, Ft−1 )
st−1
X p(yt |st, Ft−1 )p(st |st−1, Ft−1 )p(st−1 |Ft−1 )
= .
s
p(yt |Ft−1 )
t−1

Several remarks are in order. First, immediately useful byproducts of Algorithm 2.3.1 are the

time-varying transition probabilities p(st |st−1, Ft−1 ) and the filtered state probabilities p(st |Ft ),

with which we may draw inferences of current and future states. Second, an optional step

parallel in sequence to Updating can be easily implemented to extract conditional moments of

latent factor wt given information Ft in each iteration of Recursion. In particular, we show in

Subsection 2.3.2 the conditional density p(wt |Ft ) is characterized by densities and distributions

computed in previous iterations. Finally, Algorithm 2.3.1 can be extended to admit regime switch-

ing autoregression of order m by replacing conditional probabilities p(st |Ft−1 ), p(st |st−1, Ft−1 ),

p(st−1 |Ft−1 ) and p(yt |st, Ft−1 ) for each t by p(st−m:t |Ft−1 ), p(st |st−m−1:t−1, Ft−1 ), p(st−m−1:t−1 |Ft−1 )

and p(yt |st−m:t, Ft−1 ), respectively. Accordingly, the summations in Forecasting and Updating are

to be adjusted to marginalize out st−m−1 , and the summation in Evaluation marginalizes out st:t−m .

The characterization of transition probability is invariant to this extension by Theorem 3.1 in Chang

et al. (2017).

30
2.3.2 Extraction of Latent Factor

We now characterize the conditional density p(wt |Ft ), with which the extraction of, for example,

E(wt |Ft ) becomes a standard integration exercise. Applying Bayes formula to write

X p(wt |st−1, Ft−1 )p(yt |st, Ft−1 )p(st−1 |Ft−1 )

p(wt |Ft ) = , (2.3.2)
s
p(yt |Ft−1 )
t−1

in which all densities but p(wt |st−1, Ft−1 ) have been specified above.

For st−1 = (0, 0)0, there is

  −1/2
p(wt |st−1 = (0, 0)0, Ft−1 ) = det Ω−1 Q−1 Pvv·u Σww
 
ΦQ τ − Q A0 Pvv·u
−1 (w − P P−1 u
t vu uu t−1 )
×
ΦΣww (τ)
 
×φ wt ; Pvu Puu
−1
ut−1, Ω (2.3.3)

in which

Q = ( A0 Pvv·u
−1
A + Σww ) ,
−1 −1

Ω = (Pvv·u
−1
− Pvv·u
−1
AQ A0 Pvv·u
−1 −1
) ,

with generic notation Φ M (·) denoting zero-mean bivariate normal distribution function with co-

variance matrix M, and φ(·; µ, Ω) denoting bivariate normal density with mean µ and covariance

Ω. For a proof of (2.3.2) and (2.3.3), see Appendix 2.C.1. Similar calculation easily delivers the

conditional densities p(wt |st−1, Ft−1 ) in which st−1 = (0, 1)0, (1, 0)0 and (1, 1)0.

The calculation of E(wt |Ft ) is computationally demanding due to both the non-standard shape

31
of underlying distribution and high dimension. Many methods are available for efficient calculation.

For simplicity, we apply self-normalized importance sampler. The detail of our sampler is reported

in Appendix 2.B.1 as the treatment is now standard in literature.

2.4 Identification

In this section, we first bridge the standard theory of mixture distribution to show identification of

parameters in state transition. Then we show the parameters characterizing feedback are identified

through the interaction between transition and measurement. Finally, the parameter identification

in the measurement equation follows trivially from the Gaussian regression errors.

2.4.1 Identification of Mixture Distribution

Following standard terminologies of mixture distribution, the likelihood function

X
p(y1, y2, · · · , yT ) = p(y1, y2, · · · , yT |s1, s2, · · · , sT )p(s1, s2, · · · , sT )
s1,··· ,sT
X YT
= * p(yt |st, Ft−1 ) + p(s1, s2, · · · , sT ) (2.4.1)
s1,··· ,sT , t=1 -

is a finite mixture of normal densities with mixing distribution p(s1, s2, · · · , sT ) and mixing com-

p(yt |st, Ft−1 ).

Q
ponents t

The joint distribution p(st−1, st ) is identifiable following a two-step argument.5 First, the finite

mixture
X
p(yt |st, Ft−1 )q(st ),
st

5See Krolzig (1997), Section 6.2, for a relevant discussion

32
with generic mixing distribution q(st ) and mixing components p(yt |st, Ft−1 ), are identifiable by

Proposition 1 of Teicher (1963) provided a total ordering for the mixing components. Following

proof of Proposition 1, we may impose a lexicographical order such that

p(yt |(a1, a2 )0, Ft−1 ) ≺ p(yt |(b1, b2 )0, Ft−1 )

if and only if a1 < b1 or a1 = b1 but a2 < b2 where ai, bi ∈ {0, 1} for i = 1, 2 by assuming

σ , σ and/or x 0 β < x 0 β almost surely X in each equation. Second, the mixture (2.4.1) is
QT
identifiable by Theorem 1 of Teicher (1967) since its mixing component t=1 p(yt |st, Ft−1 ) is

a T-fold product of p(yt |st, Ft−1 ). Then by Definition 12.4.3 of Cappé et al. (2005), the mixing

distribution p(s1, s2, · · · , sT ) of (2.4.1) is identifiable. It follows immediately that for each st , mixing

p(yt |st, Ft−1 ) is identifiable. By normality of the mixing component, parameters in

Q
component t

our measurement equations, ( β, Puu ), are identifiable.

Our argument may be extended to include Markov switching autoregression of order m with a

total ordering of mixing components such that

a1,1 + am,1 + b1,1 + bm,1 +

p *. yt *. / , · · · , *. / , Ft−1 +/ ≺ p *. yt *. / , · · · , *. / , Ft−1 +/
, , a1,2 - , am,2 - - , , b1,2 - , bm,2 - -

if and only if (a1,1, a1,2, · · · , am,1, am,2 )0 ≺ (b1,1, b1,2, · · · , bm,1, bm,2 )0.

2.4.2 Identification of State Transition

We can characterize parameters of regime switching, (τ, A, Pvv ), in terms of joint state distribution

p(st, st−1 ). Note the transition of latent factor per se is a homogeneous Markov process, thereby the

33
sequence of state variables (st ) by itself is an exogenous homogeneous Markov chain. Therefore the

state transition probability p(st |st−1 ) carries the same information of the joint probability p(st, st−1 ).

Identification of Thresholds and Factor Correlation

From state distribution p(st ), we identify threshold parameters τ up to constant multiples, as well

as the correlation matrix between latent factors.

Let Dw be the diagonal of Σww such that

σ1
Dw1/2 = *. +/ ,
, σ2 -

in which σ1, σ2 are reparameterizations of ( A, Pvv ) given by

f
σ12 = − α12
3
α21 + α12
2
(1 + α11 α22 ) + (1 − α11 α22 )(1 − α22
2
) − α12 α21 (1 + α22
2
)
g.
+2α12 (α11 − α22 det A) ρv1,v2 π(λ 1, λ 2 ), (2.4.2)
f
σ22 = α11
3
α22 − α11
2
(1 + α12 α21 ) + (1 − α12 α21 )(1 + α21
2
) − α11 α22 (1 − α21
2
)
g.
+2α21 (α22 − α11 det A) ρv1,v2 π(λ 1, λ 2 ), (2.4.3)

with

π(λ 1, λ 2 ) = (1 − λ 1 λ 2 )(1 − λ 21 )(1 − λ 22 ),

where λ 1, λ 2 are eigenvalues of A. Note the stability of A guarantees π(λ 1, λ 2 ) , 0. The distri-

bution of Dw−1/2 wt is standard bivariate normal with correlation ρ12, whose density is denoted by

34
φ ρ12 (z1, z2 ). Then the state probability

( )
P(st = (0, 0)0 ) = P Dw−1/2 wt < Dw−1/2 τ

by construction.

Marginalizing p(st ) along s2 to have p(s1,t ) and

τ1
= Φ−1
1 (P(s 1,t = 0))
σ1

1 (·) is the inverse of standard normal distribution function. Then the normalized
in which Φ−1

thresholds are
1 (P(s 1,t = 0))
Φ−1
Dw−1/2 τ = *. +/ .
, 1 (P(s 2,t
Φ−1 = 0)) -
( )
For each Dw−1/2 τ, the function Ψ( ρ12 ) = P Dw−1/2 wt < Dw−1/2 τ is strictly increasing in ρ12 because

Z τ2 /σ2 Z τ1 /σ1
dΨ( ρ12 ) ∂
= φ ρ12 (z1, z2 )dz1 dz2
dρ12 ∂ ρ12 −∞ −∞
Z τ2 /σ2 Z τ1 /σ1
∂2
= φ ρ (z1, z2 )dz1 dz2
−∞ −∞ ∂z1 ∂z2 12
= φ ρ12 (τ1 /σ1, τ2 /σ2 ) > 0.

And it follows that

ρ12 = Ψ−1 P(st = (0, 0)0 ) .

This result establishes global identification for the stationary case of Chang et al. (2017) as a
√
corollary: recovered from p(st, st−1 ), the scaled threshold parameter τ 1 − α 2 , and the correlation

35
coefficient α for (wt, wt−1 )0.

Identification of Transition Equation

Extending results in the preceding section, we may identify all parameters in the state transition.
0 , w0 ) 0 is identified through the joint state distribution
We first show the correlation matrix of (wt−1 t

p(st, st−1 ). Then we characterize ( A, Pvv ) in terms of the identified correlation matrix.
0 , w0 ) 0 ∼ N(0, Σ) such that
To begin with, observe (wt−1 t

Σww Σww A0 +
Σ = *. /.
, AΣww Σww -

We may write
1/2 1/2
Σ = Dww Γ(wt−1, wt )Dww ,

1/2
 
such that Dww = diag Dw1/2, Dw1/2 and

*. 1 +/
0 ρ12
. //
Γ0 Γ1 + .. 1
Γ(wt−1, wt ) = *. / = .. // ,
, Γ1 Γ0 - .. ρ13 ρ23 1
//
. //
, ρ14 ρ24 ρ34 1 -

0 , w0 ) 0 with ρ −1/2
denoting the correlation matrix of (wt−1 t 12 = ρ34 . As is clear, (Dw τ, ρ12 ) is

identified through p(st−1 ). With identical arguments, the correlation matrix Γ(wt−1, wt ) is identified

through p(st, st−1 ). Specifically, we obtain the marginal distributions p(s1,t−1, s1,t ), p(s1,t−1, s2,t ),

p(s2,t−1, s1,t ) and p(s2,t−1, s2,t ) and identify ρ13, ρ14, ρ23 , and ρ24 , respectively.

The correlation matrix Γ(wt−1, wt ) provides 5 equations, i.e. ρ12, ρ13, ρ14, ρ23 and ρ24 , for 5

36
unknowns (α11, α12, α21, α22, ρv1,v2 ). We solve this system in part easily using the block matrices

of Σ. Note that

AΣww = Dw1/2 Γ1 Dw1/2,

with

Σww = Dw1/2 Γ0 Dw1/2 .

Therefore,

ρ13 − ρ23 ρ12

α11 =
1 − ρ212
ρ23 − ρ13 ρ12 σ2
α12
∗
= , α12
∗
= α12
1 − ρ212 σ1
ρ14 − ρ24 ρ12 σ1
α21
∗
= , α21
∗
= α21
1 − ρ212 σ2
ρ24 − ρ14 ρ12
α22 =
1 − ρ212

Note α11 , α22 and α12 α21 are identified, as well as the signs of α12 and α21 , followed by trA and

det A. In particular, α12, α21 are directly identified if at 0.

We may substitute out α21 , since the only non-trivial case concerns α12 α21 , 0. Hence the

remaining parameters (α12, ρv1 v2 ) are characterized by the solution of the bivariate system regarding

37
 σ12
ρ12 and α12
∗ . Note that ρ =
12 σ1 σ2 , in which

f
σ12 = − α11
2
α12 α22 + α12 (1 + α21
2
)α22 + α11 α21 (1 + α12
2
− α22
2
)
g.
+(1 − α11
2
− α22
2
+ (α11 α22 ) 2 − (α12 α21 ) 2 ) ρv1,v2 π(λ 1, λ 2 ).

It is easily solved that

v
u
tf g
(α12 α21 ) 2 − (1 − α22
2 )α ∗ 2 (1 + trA + det A) + 2α α ∗ ρ α α
12 22 12 12 12 21
α12 = sign(α12 ) 2 − 1 + α ∗ 2 )(1 + trA + det A) + 2α α ∗ ρ
,
(α11 12 11 12 12
f  
ρv1,v2 = α12
2
(1 + det A)(α21 − α12 ) + (1 − det A)(1 − α11
∗ 2 2
) − 2α12 α21
g
∗ 2
−α12 (1 − α12 α21 − α22 (trA − α22 det A))
.f  g
∗ 2 ∗ 2
2α12 α11 (1 − α22
2
)α12 − α22 (1 − α11
2
− α12 )(α12 α21 ) − α11 (α12 α21 ) 2 .

Finally, we conclude that τ1 and τ2 are identified since σ12, σ22 are specified by (2.4.2) and

(2.4.3).

2.4.3 Identification of Endogenous Feedback

The identification of state transition implies the identification of endogenous feedback. By the

theory of mixture distribution, the joint density

p(y1, · · · , yT , s1, · · · , sT ) = p(y1, · · · , yT |s1, · · · , sT )p(s1, · · · , sT ),

is identifiable, followed by
p(st, st−1 )
p(st |st−1 ) =
p(st−1 )

38
and time-varying transition

p(y1, · · · , yt−1, st, st−1 )

p(st |st−1, Ft−1 ) = .
p(y1, · · · , yt−1, st−1 )

On the one hand, all parameters except for Pvu are identifiable through the unconditional

transition probability p(st |st−1 ). On the other hand, suppose there exist Pvu and Pvu
∗ such that both

induce the same time-varying transition probability p(st |st−1, Ft−1 ), then by (2.2.3), it must be true

that

Pvu Puu
−1
ut−1 =a.s. Pvu Puu ut−1,
∗ −1

followed by Pvu = Pvu

∗ .

2.5 Empirical Illustration

In this section, we examine a vector model for U.S. monetary and fiscal policy rules with regime

switching and feedback in policy regimes and analyzing the unobserved regime interactions.

Main theoretical works on policy interactions include Sargent and Wallace (1981), Wallace

(1981), Aiyagari and Gertler (1985) and Leeper (1991). The overriding message is that monetary

and fiscal policies together stabilizes real government debt and determines price level. There are

two main competing theories. The conventional theory describes an active monetary authority,

who systematically raise nominal interest rate more than one-for-one in response to current inflation

while the fiscal authority adjusts tax and spending passively to maintain solvency. The alternative

theory describes an active fiscal authority who spend on its agenda, and a passive monetary authority

generates seigniorage to maintain solvency.

39
 Empirical works devote much attention to the dynamics of policy interactions. For example,

Favero and Monacelli (2005) consider regime switching policy rules and find little evidence of

synchronization in regime-switching; Bianchi and Ilut (2017a) embed regime-switching policy

rules into a New-Keynesian dynamic stochastic general equilibrium (DSGE) model and explain

inflation drop in the 1980s concerning a policy shift. Two works are more closely related to our

exercise. Davig and Leeper (2006) study monetary and fiscal policy rules jointly with exogenous

Markov regime process. Moreover, Chang and Kwak (2017) examine these rules separately and

demonstrate the importance of allowing for endogenous feedback to the latent regime factors, and

in which it is argued that latent factor provide a more plausible interpretation of sudden policy

shifts because rarely do policymakers choose to shift discretely to a new regime.

The primary advantage of our approach is that we provide a consistent framework for analyzing

policy interaction with the presence of feedback because the channels of interaction are explicitly

modeled. In what follows, we refer to the direct channel as the factor loading A of latent factors, and

the indirect channel as the feedback to regime factors from a joint policy innovation, characterized

by correlation matrix P.

2.5.1 Interactions of U.S. Monetary and Fiscal Policies

We consider regime-switching monetary and fiscal policy rules

it = α c (stM ) + α π (stM )πt + σ M utM (2.5.1)

τt = βc (stF ) + βb (stF )bt−1 + βg (stF )gt + σ F utF (2.5.2)

40
with latent regime factor

wt = Awt−1 + vt

sit = 1{wti ≥ τi }

for i = M, F such that wt = (wtM , wtF )0, vt = (vtM , vtF )0. In the policy rules, it and πt represent

nominal interest rate and inflation rate at time t; τt, bt−1 and gt represent revenues net of transfer

payments, government spending and debt held by public. The main model we consider assumes

both general stable A and unrestricted correlation matrix P, each characterizing a channel for policy

interactions. In what follows, we use “main" and “unrestricted" interchangeably.

We hereby view regression error ut = (utM , utF )0 as fundamental shocks or functions of variables

omitted by econometrician. Within-equation feedback thus arises naturally when policy regime

responds systematically to information available to policymakers. In such case, size of responses are

functions of both historical policy disturbances and exogenous shocks. Also, since both authorities

routinely project variables in the information set of the other branch, the cross-equation feedback

arises naturally as well.

Following Leeper (1991), regimes for monetary policy and fiscal policy depend upon the

parameter values in monetary and fiscal policy rules. The monetary policy is said to be active if

α π > 1, i.e., the policy rate respond more than one for one to inflation, and passive if 0 ≤ α π < 1.

The fiscal policy is said to be active if βb is less than the real interest rate, and passive if otherwise.

We follow Leeper (1991) and call Regime M and Regime F to be “active-passive" and “passive-

active" combinations of monetary and fiscal regimes, respectively. Both regimes imply the existence

of a determinate bounded rational expectations equilibrium.

41
 Additional identification issues may arise, because, in addition to Regime M/F, our model

allows both doubly active and passive regimes. Leeper (1991) argues that doubly active regime

leads to the nonexistence of a money-growth process that ensures consumer will hold government

debt unless policy innovations are correlated; and doubly passive regime leads to multiple money-

growth processes satisfying equilibrium conditions and, hence, indeterminate pricing function.

Nonetheless, our model assumptions are different. First, we allow shocks (utM , utF )0 to be correlated

and assume independence over time, whereas Leeper (1991) assumes uncorrelated innovations

across equation, and AR(1) specification for utF . Second, provided a passive fiscal rule, Davig and

Leeper (2007) show that monetary policy can satisfy the Taylor principle in the long run, even while

deviating from it substantially for brief periods or modestly for prolonged periods. Empirically, we

find doubly passives regimes to be all relatively short-lived. We thus argue indeterminacy (hence,

identification) is not a major concern for our empirical exercise.

Davig and Leeper (2006) include output gaps and switching volatility in the regression, but

can otherwise be considered as a special case of our model since they estimate policy rules with,

equivalently,
α11
A = *. +/ , P = I4 .
, α22 -

Chang and Kwak (2017) is also a special case of our model since separate estimation of (2.5.1) and

(2.5.2) is equivalent to assuming only within-equation feedback, characterized by

*. 1 +/
α11
.. //
0 1
A = *. +/ , P = ... // .
α22 - .. ρv1 u1 0 1
//
, //
.
, 0 ρv2 u2 0 1 -

42
In what follows, we call Chang and Kwak (2017) the “restricted" model.

2.5.2 Data and Identified Regimes

We use quarterly U.S. data from 1949:2 to 2014:2 for the empirical exercise. Figure 2.1 plots

these time-series for policy instruments, on which we superimpose the Regime M and Regime F

identified using extracted latent factors E(wt |Ft ) and estimated threshold τ from the main model.

The regimes are primarily M or F over the course of history with a brief doubly active period in the

1980s and multiple brief doubly passive epochs in 1950s and 2000s.

Figure 2.1: Historical Data and Estimated Policy Regimes

15
it

t
10

1950 1960 1970 1980 1990 2000 2010

0.7
t
0.15
b t-1 0.6
gt
0.1 0.5

0.4
0.05 0.3

0.2
0
1950 1960 1970 1980 1990 2000 2010
Time

Note: Blue indicates estimated Regime M, androse indicates estimated Regime F. The series for
bt−1 corresponds to right vertical axis in the lower panel. All other series correspond to left vertical
axis.

For monetary policy (2.5.1), we define πt to be inflation rate over contemporaneous and prior

43
three quarters as in Taylor (1993) and obtain inflation each period as log difference of GDP deflator.

We use three-month Treasury bill (T-bill) rate in the secondary market for nominal interest rate it .

We choose T-bill rate over federal funds rate (FFR) mainly due to its short series. On the other hand,

T-bill rate is highly correlated with FFR with sample correlation 0.988 over the period 1954:1-

2014:2, and is available since 1949:1. Using T-bill rate allows us to study meaningful regime

changes in monetary and fiscal policy rules before 1954 which include critical historical episodes

such as Treasury Accord of March 1951 leading to passive monetary policy and the wartime fiscal

financing for Korean war leading to active fiscal policy.

For fiscal policy (2.5.2), all variables are for the federal government only. τt is federal tax

receipts net of total federal transfer payments to GDP ratio, and bt−1 is the market value of gross

marketable federal debt held by the public to GDP ratio and gt is federal government consumption

plus investment expenditures to GDP ratio. Finally, we use average debt-output ratio over previous

four quarters as a measure of bt−1 .

2.5.3 Estimation and Results

We perform ML estimation for a sequel of specifications with a detailed description of optimization

routine in Appendix 2.B.2 and results in Table 2.1. Standard errors are reported for only the major

cases, and are obtained by simulation since the hessian of log-likelihood is difficult to compute. The

regime-switching coefficients for both policies appear to be identified in all specifications. In the

unrestricted model, the monetary rule responds actively to inflation with coefficient 1.049(0.147)

and passively with coefficient 0.640(0.044), while the fiscal rule responds actively to debt with

coefficient -0.029(0.004) and passively with coefficient 0.050(0.011).

In estimation, we propose an error component parametrization for (ut0, vt+1

0 ) 0 with exactly 6

44
parameters to identify structure of shocks and give proper correlation matrix P. Let

q
uit = λ i ξt + 1 − λ i2 ζit (2.5.3)
q
vi,t+1 = φi ξt + ψi ζit + 1 − φi2 − ψi2  it (2.5.4)

for i = 1, 2 such that (ξt ), (ζit ) and ( it ) are independent standard normal, and −1 < λ i, φi, ψi < 1.

We further assume λ 1 ≥ 0 and φi2 + ψi2 < 1 for each i = 1, 2. The former restriction is necessary

since (ξt ) paired with (λ i, φi ) gives identical transition as by (−ξt ) paired with (−λ i, −φi ). Channels

for endogenous feedback are characterized by λ i, φi and ψi for i = 1 and 2. The within-equation
q
feedback of equation i is determined by λ i φi +ψi 1 − λ i2 for i = 1 and 2, whereas the cross-equation

feedback is characterized by products λ 1 φ2 and λ 2 φ1 .

We report two main findings from ML estimation in addition to the improved efficiency of the

unrestricted model compared to the restricted one. On the one hand, the likelihood ratio test gives

evidence that restricted model is misspecified, and the estimates suggest that the misspecifications

lie in both channels of policy interaction we consider. Data favors the direct channel of interaction,

characterized by off-diagonal entries of A. Notably, the impact of fiscal factor on monetary factor,

α12 , is twice in size compared to the effect in the opposite direction, α21 . Data also favors a

non-trivial positive correlation between fiscal and monetary information, ρu1,u2 , and positive cross-

equation feedback from fiscal information to monetary regime factor, ρu2,v1 , but not the opposite

direction of cross-equation feedback, ρu1,v2 . Both channels are essential in the sense that severing

one leads to non-rejection in likelihood ratio test (see (3) in Table 2.1). On the other hand, however,

coefficient estimates for the regime-switching policy rules are similar across all specifications

considered.

45
 Table 2.1: Maximum Likelihood Estimation (1949:Q2-2014:Q2)

Parameter (1)-Unres. S.E. (2) -Res. S.E. (3) (4)

Thresholds
τm 0.435 1.705 -0.389 1.648 0.802 0.533
τf -0.582 1.407 -0.600 1.551 -0.026 -0.462
Transition of Latent Factor
α11 0.956 0.248 0.984 0.420 0.975 0.961
α21 0.023 0.161 - - - 0.012
α12 0.056 0.249 - - - 0.050
α22 0.938 0.177 0.968 0.283 0.961 0.953
Endogenous Feedback
ρu 1 u 2 0.178 0.066 - - 0.180 0.146
ρu1 v1 0.997 0.418 0.999 0.470 0.943 0.982
ρu2 v1 0.165 0.206 - - 0.322 -
ρu1 v2 0.000 0.243 - - 0.225 0.136
ρu2 v2 0.970 0.242 0.999 0.368 0.988 0.996
ρv1 v2 0.000 0.245 - - 0.403 -
Regime-Switching Monetary Policy
α c (s m = 0) 0.533 0.265 0.443 0.529 0.456 0.528
α c (s m = 1) 2.524 0.477 2.601 0.607 2.531 2.564
α π (s m = 0) 0.640 0.044 0.661 0.095 0.658 0.647
α π (s m = 1) 1.049 0.147 1.039 0.183 1.049 1.047
σm 1.310 0.064 1.306 0.068 1.309 1.309
Regime-Switching Fiscal Policy
βc (s f = 0) -0.028 0.001 -0.028 0.003 -0.028 -0.028
βc (s f = 1) 0.011 0.003 0.014 0.011 0.011 0.012
βb (s f = 0) -0.029 0.004 -0.033 0.009 -0.028 -0.029
βb (s f = 1) 0.050 0.011 0.052 0.020 0.054 0.051
βg (s f = 0) 1.016 0.024 1.020 0.051 1.017 1.018
βg (s f = 1) 0.644 0.056 0.603 0.107 0.634 0.637
σf 0.014 0.001 0.014 0.001 0.014 0.014
log-likelihood 275.137 270.409 273.332 274.629
p-value (vs (2)) 0.051* - 0.211 0.077
Note:* indicates d f = 4 in the likelihood ratio test against restricted (Res.) model.
Missing values are 0 in the shaded rows, and undefined if otherwise.

Granger causality test is easily constructed for the direct channel of factor interaction. The fiscal

(monetary) factor Granger causes shifts in monetary (fiscal) factor if α12 , 0 (α21 , 0). Since

the sampling distribution for our model is unclear, we formally test the null hypothesis α12 = 0

(α21 = 0) with a bootstrap test. Data, however, generates insufficient evidence for interaction

46
between policy factors (see Table 2.2).

Interestingly, the estimated main model suggests that fiscal information leads monetary regime

determination. Nonetheless, data is ambiguous about the leading role of monetary and fiscal

information since the monetary leading model also fits significantly better than the restricted model

(see (4) in Table 2.1). We formalize the statement by testing null hypothesis that ρu2 v1 = ρv1 v2 = 0

and monetary information leads using a bootstrap test. The test result is reported in Table 2.2.

Table 2.2: Leading FP vs. Leading MP

Parameter Leading FP 90% Band Leading MP 90% Band

α11 0.956 0.230 0.987 0.961 0.031 0.993
α21 0.023 -0.087 0.085 0.012 -0.066 0.134
α12 0.056 -0.008 0.158 0.050 -0.029 0.209
α22 0.938 0.814 0.981 0.952 0.812 0.992
ρu 1 u 2 0.178 0.069 0.288 0.146 0.025 0.253
ρu1 v1 0.997 -0.186 1.000 0.983 -0.305 0.997
ρu2 v1 0.165 -0.205 0.519 - -0.554 0.751
ρu1 v2 - -0.393 0.465 0.137 -0.035 0.470
ρu2 v2 0.970 0.390 0.996 0.996 0.488 1.000
ρv1 v2 - -0.455 0.421 - -0.562 0.747
Note: The Granger causality tests are constructed using α21 and α12 . All missing values
are 0 unless noted otherwise. And rose is for bootstrap test with H0 : ρu2 v1 = ρv1 v2 = 0
and ρu1 v2 = 0.137.

We also report the extracted latent factors for different specifications. Following Chang and

Kwak (2017), the monetary regime is defined to be active if wtM ≥ τm , whereas the fiscal regime

is active if wtF < τ f , and the inferred regimes are determined by their estimates and extracted

conditional expectation. Figure 2.2 presents the extracted latent dynamic factors with inferred

active regimes indicated by the shaded area. The result suggests implied regime strength are

different between specification (1) and (2). However, nonetheless, the implied regimes are similar

in timing and length.

Without a suitable framework, Chang and Kwak (2017) use the extracted latent factors to

47
 Figure 2.2: Extracted Latent Factor (MLE)

(1) - Joint Monetary Policy Factor

5
0
-5
1950 1960 1970 1980 1990 2000 2010
(2) - Single

5
0
-5
1950 1960 1970 1980 1990 2000 2010
Fiscal Policy Factor
5
(1)

0
-5
1950 1960 1970 1980 1990 2000 2010

5
(2)

0
-5
1950 1960 1970 1980 1990 2000 2010
Time

Note: The extracted factor (solid line) E(wt |Ft ) and threshold (dashed line) τ defines the active
regime (shade). In the panels of monetary factor, the regime is identified as active if E(wtM |Ft ) ≥ τM .
In the panels of fiscal factor, active regime is identified if E(wtF |Ft ) < τF .

perform a second-stage inference on policy interactions by fitting them to a time-varying coefficient

vector autoregression (TVC-VAR). Their approach, however, is inconsistent with the notion of

interaction because none is actually accounted for. Figure 2.3 reports the impulse responses of

regime factors corresponds to MLE and second-stage inference, respectively. There is notable

difference between MLE implied impulse responses and estimated response from latent factors in

the restricted model suggesting such practice is more appropriate in our model since the interaction

is reflected in the extracted latent factor.

The correlation of extracted factors is 0.43 for the restricted model, whereas it is 0.78 for the

48
 Figure 2.3: Impulse Responses of Factors

(a) (b) (c) (d)

1 1 1 1

MLE - Single
MLE - Joint
0.8 0.8 0.8 0.8
0.6 0.6 0.6 0.6
0.4 0.4 0.4 0.4
0.2 0.2 0.2 0.2
0 0 0 0
20 40 20 40 20 40 20 40

(e) (f) (g) (h)

1 1 Monetary 1 1
Fiscal

VAR - Single
VAR - Joint

0.5 0.5 0.5 0.5

0 0 0 0

-0.5 -0.5 -0.5 -0.5

20 40 20 40 20 40 20 40

Note: Panels (a) - (d) plot impulse responses of factors implied by MLE. And panels (e) - (h) plot
estimated impulse responses implied by extracted latent factors. Panels (a), (c), (e) and (g) assume
1 standard deviation shock to monetary factor. The rest assume 1 standard deviation shock to fiscal
factor.

unrestricted model (see Table 2.D.1). Figure 2.4 examines the pattern of comovements for extracted

factor by comparing the 6-year rolling window correlation. This result shows the regime factors

move more closely in the unrestricted model than in the restricted model. In terms of historical

events, the restricted model suggests non-cooperative regime factors in the 1950s and 1980s. In

contrast, the unrestricted model presents evidence of non-trivial cooperation through out the whole

sample period with several exceptions. Nonetheless, It must be noted such cooperation does not

guarantee synchronization in the policy regimes.

We also characterize correlations of data and factors in frequency domain. Figure 2.5 compares

49
 Figure 2.4: Comovements and Correlation of Extracted Latent Factors

5
Joint

0
MP
-5 FP

1950 1960 1970 1980 1990 2000 2010

1
6-Yr Corr

-1
1950 1960 1970 1980 1990 2000 2010

5
Single

0
-5
1950 1960 1970 1980 1990 2000 2010
1
6-Yr Corr

-1
1950 1960 1970 1980 1990 2000 2010
Time

Note: The panels demonstrate the comovement between extracted monetary factor (blue solid) and
fiscal factor (red dashed). “6-Yr Corr" reports rolling window correlation between extracted latent
factors with window size 24. The upper panels pertain to the unrestricted model, and the lower
panels to the restricted model.

the coherence of policy rates and those implied by filtered latent factors in frequency domain for

different models, from which we summarize two results. First, our unrestricted model implies

strong coherence between factors in business cycle frequency. In contrast, policy rates display

strong correlation at lower frequencies while restricted model attribute rather weak correlation

across the entire spectrum.

Second, combining with parameter estimates in Table 2.2, there is evidence of direct but weak

interaction between factors. In addition, our unrestricted model identifies shocks that propagate

50
 32Q 6Q
1
Joint
Single
0.9

0.8
Magnitude-Squared Coherence

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized Frequency ( )

Note: Magnitude-squared coherence at frequency λ is defined to be | f xy (λ)| 2 /[ f x x (λ) f yy (λ)],

with f xy (λ) denoting the cross-spectral-density, and estimated using Welch’s method with window
size 24. The dashed vertical lines indicate the normalized frequencies associated with 6Q and 32Q.
Figure 2.5: Magnitude-Squared Coherence of Extracted Latent Factors

Table 2.3: Selected Parameter Estimates for Error Component Model (1949:Q2-2014:Q2)

Parameter Unres. S.E. Res. S.E.

λ1 0.744 0.234 - -
λ2 0.239 0.190 - -
φ1 0.691 0.469 - -
φ2 0.000 0.368 - -
ψ1 0.723 0.462 0.999 0.470
ψ2 0.999 0.283 0.999 0.368

very differently to monetary and fiscal factors. It is tempted to regard (λ i, φi, ψi ) in Equation

2.5.3 and 2.5.4 as relative importance of components identified in shocks. The common factor

of shocks accounts for a large proportion in monetary rule and is much weaker in fiscal rule

(Table 2.3). It is then largely propagated to monetary factor in next period, whereas its impact

51
to fiscal factor is negligible. Fiscal factor innovation, on the other hand, is mainly driven by last

period equation-specific factor to fiscal rule.

2.6 Simulation

2.6.1 Simultaneity Bias Reduction

Simultaneity bias is well expected if the conventional approach is applied when the data generating

process (DGP) entails interacting latent factors, correlated error terms or cross-equation feedback.

In this section, we draw a comparison between the unrestricted model and the restricted model by

simulation. Using this result, we supplement the empirical exercise in the preceding section.

We simulate data from our empirical model (2.5.1) and (2.5.2) at its ML estimates. And as

before, the correlation matrix P is parameterized by error component model (2.5.3) and (2.5.4).

The exogenous variables are fixed to be the full sample data with sample size is 262. The number

of iteration is set to be 1,000.

Table 2.4 reports the simulation results. We measure bias reduction with the restricted model

as benchmark using a simple metric

%Bias Reduction = |%Bias Res. | − |%BiasUnres. |.

Column 5 of Table 2.4 shows that the estimates in the restricted model suffer heavy simultaneity

bias when ignoring the off-diagonal elements of A and the non-trivial cross-equation endogenous

feedback, and the correlation in the shocks. Individually, ignoring cross-equation interaction

inflicts (i) sizable bias for threshold values, (ii) substantial bias in the size of interaction and

feedback channels, (iii) substantial bias in the correlation of shocks, but lead to (iv) relatively small

52
 Table 2.4: Relative Bias against DGP at Joint MLE

Parameter DGP % Bias %Bias Reduction

(1)-Unres. (2)-Res.

Thresholds
τm 0.435 -28.627 74.016 45.389
τf -0.582 -1.589 -133.315 131.726
Transition of Latent Factor
α11 0.956 -7.210 -15.945 8.734
α21 0.023 6.138 -100.000 93.862
α12 0.056 79.508 -100.000 20.492
α22 0.938 -2.641 -6.339 3.699
Endogenous Feedback
ρu 1 u 2 0.178 -0.504 -100.000 99.496
ρu1 v1 0.997 -27.336 -28.332 0.996
ρu2 v1 0.165 -5.163 -100.000 94.837
ρu1 v2 0.000 - - -
ρu2 v2 0.970 -14.159 -19.089 4.930
ρv1 v2 0.000 - - -
Regime-Switching Monetary Policy
α c (s m = 0) 0.533 11.008 25.428 14.419
α c (s m = 1) 2.524 -2.889 -6.482 3.593
α π (s m = 0) 0.640 0.401 1.238 0.837
α π (s m = 1) 1.049 -1.906 -3.524 1.618
σm 1.310 -1.227 -1.159 -0.068
Regime-Switching Fiscal Policy
βc (s f = 0) -0.028 -0.188 -0.107 -0.080
βc (s f = 1) 0.011 0.048 -1.540 1.492
βb (s f = 0) -0.029 -0.443 1.516 1.073
βb (s f = 1) 0.050 -0.284 -0.938 0.654
βg (s f = 0) 1.016 0.382 1.233 0.852
βg (s f = 1) 0.644 -1.062 -2.581 1.519
σf 0.014 -1.213 -1.238 0.026
Note: We suppress results for ρu1,v2 and ρv1,v2 for their relative biases are not defined.
The % bias reduction is defined as the difference of absolute value of the relative bias
between (1) and (2).

bias for the coefficients in the measurement equation.

The implication is twofold. First, one should consider unrestricted model whenever possible

since, otherwise, the channels for interaction may be distorted. In consequence, making an inference

53
of interaction by examining latent factors from individual rules is inconsistent with the notion of

interactions. Second, estimating a single equation model may be worthwhile because the parameters

in the measurement equation are not very sensitive to the dynamics of latent factors and it is much

less computationally intensive than estimating the unrestricted model. An efficient strategy is to

learn from data incrementally by starting from estimating the model equation by equation, followed

by a joint estimation with the first-stage ML estimates as the initial guess.

2.7 Conclusion

We have shown the regime-switching model driven by latent VAR(1) factors is a powerful tool. We

allow rich temporal dynamics of factors and both within-equation and between-equation feedback,

so a shock to one observed time series or factor may transmit to both regime factors. Our model

provides a proper treatment for empirical works in which we explicitly account for channels of

regime interactions. In our exercises, we find evidence suggests the presence of regime factor

coordination and cross-equation feedback in U.S. monetary and fiscal policy regimes. Moreover,

our simulation makes it clear that neglecting simultaneity in regime-switching incurs substantial

bias for parameters of regime switching, but relatively small bias for parameters in the measurement

equation.

54
Appendix

Appendix 2.A A Parallelizable Algorithm for Likelihood Evaluation

With 2 regimes for each equation, the computational complexity for Equation (2.3.1) is O(42T ).

However, practical issue arises from the calculation of (2.2.4). Its time-varying nature entails

heavy and repeated evaluations of 4-dimensional multivariate normal distribution. This motivates

a simpler and parallelizable characterization of the log-likelihood function. Here, we devise

a parallelizable algorithm of complexity O(42T ), in which calculations of conditional density

p(yt |Ft−1 ) for each t can be distributed to separate CPUs.

Note the likelihood function can be written in form

YT X
`(θ) = *. p(yt |st, st−1, Ft−1 )p(st, st−1 |Ft−1 ) +/ (2.A.1)
t=1 , st ,st−1 -

with p(yt |st, st−1, Ft−1 ) = p(yt |st, Ft−1 ) by construction. From Equation (2.A.1),

Algorithm 2.A.1 (Parallelizable Modified Markov-Switching Filter)

1. Initialization. For t = 0, set p(st, st−1 ) to be the unconditional state probabilities such that

P{s0 = (0, 0)0, s−1 = (0, 0)0 } = P w1,0 < τ1, w2,0 < τ2, w1,−1 < τ1, w2,−1 < τ2


55
 in which
Σww Σww A0 ++
(w00 , w−1
0 0
) ∼ N *.0, *. //
, , AΣ ww Σ ww --

2. Parallelization. For each t ≥ 0, calculate p(st, st−1 |Ft−1 ) using (2.2.4) for each possible

realization of (st, st−1 ), and evaluate p(yt |st, Ft−1 ) for each possible realization of st . Then

X
p(yt |Ft−1 ) = p(yt |st, Ft−1 )p(st, st−1 |Ft−1 )
st ,st−1

3. Combination. Combining conditional density for each t ≥ 0 to have the log-likelihood

T
X X
log `(θ) = log *. p(yt |st, Ft−1 )p(st, st−1 |Ft−1 ) +/
t=1 ,st ,st−1 -

Appendix 2.B Computation

2.B.1 Computation of Latent Factor

The computation of latent factor wt is costly by integration techniques such as Newton-Cotes

method (e.g. Riemann sum) due to curse of dimensionality. A standard alternative is Monte Carlo

integration by importance sampling. The extracted latent factor at time t can be written as

p(wt |Ft )
Z
E(wt |Ft ) = wt q(wt )dwt
R2 q(wt )

where q(wt ) is any proposal such that p(wt |Ft ) > 0 implies q(wt ) > 0. An obvious candidate for

q(wt ) is the stationary distribution of wt . The self-normalized importance sampling estimator of

56
E(wt |Ft ) is thus given by

Pn
i=1 wti ωti
E(wt |Ft ) =
D Pn i
i=1 ωt

where (wti ) is a sequence of draws generated from the proposal q(wt ), and ωti = p(wti |Ft )/q(wti )

is the importance weight associated with the i-th draw. It is obvious that the p(wt |Ft )/q(wt ) is a

proper density with respect to the measure q(wt )dwt . We use self-normalized estimator to control

the finite sample behavior. The conditions above are sufficient to imply E(w
D t |Ft ) →a.s. E(wt |Ft ).

E(wt |Ft ). We thus can measure the approximation precision of

Central limit theorem holds for D

E(wt |Ft ) by constructing a confidence interval around it using approximated variance

D

Pn i2 i
i=1 ωt (wt E(wt |Ft ) 2 )
1
1 n −D
var(
L D E(wt |Ft )) = .
n 1 n
 P
i
2
n i=1 ωt

Note this confidence interval is different in concept to the confidence interval for parameter infer-

ence. It merely provides a measurement of approximation precision.

2.B.2 Optimization

In ML estimation, we propose the use of pattern search optimization. There are three arguments

that favors pattern search over derivative-based methods (who are usually more efficient).

1. The numerical likelihood surface is expected to be rough. This roughness renders derivative-

based optimizer less robust. In contrast, patterns search is documented to have more robust-

ness since it is derivative-free.

2. We can easily impose nonlinear constraints such as k Ak < 1 in pattern search. This is not so

easy in other methods.

57
 3. Pattern search is a global method as the likelihood surface is known to have multiple local

maxima.

Our exercises as of now suggests that pattern search can robustly reproduce estimates in Chang

and Kwak (2017). It also appears that pattern search is more viable than other methods we

considered in joint estimation. For instance, our exercise shows joint estimation with pattern search

yields estimates close to those of the single equation models, whereas derivative-based methods

failed to find any local optimum.

We also proposed a strategy to reduce the chance of being stranded at a local minimum.

Specifically, pattern search is called recursively with initial guess being the last stage local minimum

θ̂ n and stopping criteria that sup-norm k θ̂ n − θ̂ n+1 k <  for some prescribed  . In our exercises, we

set  = 10−5 .

Appendix 2.C Omitted Proofs

2.C.1 Extraction of Latent Factor

Apply Bayes formula to p(wt |Ft ) and have

p(wt |Ft ) = p(wt |yt, Ft−1 )

p(yt |st (wt ), Ft−1 ) st−1 p(wt, st−1 |Ft−1 )
P
=
p(yt |Ft−1 )
X p(yt |st (wt ), Ft−1 )p(wt |st−1, Ft−1 )p(st−1 |Ft−1 )
= .
s
p(yt |Ft−1 )
t−1

58
It thus amounts to specify conditional density function p(wt |st−1, Ft−1 ). By the decomposition of

vt ,

p(wt |st−1 = (0, 0), Ft−1 ) = p(wt |st−1 = (0, 0), ut−1 )
"Z τ #,
= φv|u (wt − Pvu Puu ut−1 − Awt−1 )φ(wt−1 )dwt−1 Φ(τ)
−1
−∞
"Z τ
= (2π) −1 (det Pvv·u ) −1/2
−∞
 1 
× exp − (wu,t − Awt−1 )0 Pvv·u
−1
(wu,t − Awt−1 )
2 ! #,
1 0 −1
×(2π) (det Σww )
−1 −1/2
exp − wt−1 Σww wt−1 dwt−1 Φ(τ)
2
= (2π) −2 (det Pvv·u Σww ) −1/2
"Z τ  1 
× exp − (wt−1 − Q A0 Pvv·u −1
wu,t )0Q−1 (wt−1 − Q A0 Pvv·u
−1
wu,t )
−∞ 2
1  0 −1 ! #,
exp − wu,t Pvv·u wu,t − wu,t Pvv·u AQ A Pvv·u wu,t dwt−1 Φ(τ)
0 −1 0 −1
2
! 1/2
det Ω−1 Q−1
= φ(wt ; Pvu Puu
−1
ut−1, Ω)
det Pvv·u Σww
"Z τ #,
× φQ (wt−1 − Q A Pvv·u wu,t )dwt−1 Φ(τ)
0 −1
−∞
! 1/2 "
det Ω−1 Q−1 ΦQ (τ − Q A0 Pvv·u
−1 w ) #
u,t
=
det Pvv·u Σww Φ(τ)
×φ(wt ; Pvu Puu
−1
ut−1, Ω)

(2.C.1)

59
where

wu,t = wt − Pvu Puu

−1
ut−1

Q = ( A0 Pvv·u
−1
A + Σww )
−1 −1

Ω = (Pvv·u
−1
− Pvv·u
−1
AQ A0 Pvv·u
−1 −1
)

and Φ is the unconditional distribution function for (wt ), whereas φQ (ΦQ ) denotes the zero-mean

bivariate normal density (distribution) function with covariance matrix Q. Similar results are easily

obtained for cases in which st−1 = (0, 1), (1, 0) and (1, 1).

Appendix 2.D Additional Tables and Graphs

Table 2.D.1: VAR(1) Fit for Extracted Latent Factor

Parameter (1)-Unrestricted (2)-Restricted

α11 0.989 (0.026) 0.975 (0.017)
α21 0.037 (0.020) 0.028 (0.014)
α12 -0.038 (0.036) -0.023 (0.025)
α22 0.921 (0.028) 0.942 (0.020)
corr b/w factors 0.780 0.428

60
Chapter 3

Expectation Effects of Switching Financial Frictions

3.1 Introduction

We investigate in this paper the effect of regime-switching in financial market conditions on

macroeconomic variables by extending the model of Christiano et al. (2014) (CMR) to allow for

switching in financial frictions arise from a costly state verification (CSV) problem as in Bernanke

et al. (1999) (BGG). Mounting evidence indicates that the spike in idiosyncratic uncertainty defined

as the cross-sectional productivity dispersion at the firm level and its tight link to the investment are

the leading sources of the sharp contraction and slow recovery in the U.S. macroeconomic activities

during the Great Recession and its aftermath. The elevation in uncertainty drives down investment

primarily through increases in the risk premium. The CSV problem is one of the standard and

empirically relevant mechanisms that generate such financial market friction (Meisenzahl, 2014).

Banks face an agency problem due to asymmetric information about the idiosyncratic efficiency

of the installed physical capital, and the verification of enterprise performance is costly.1 The

1The agency problem assumes a physical capital efficiency in production idiosyncratic to each entrepreneur
materializes after the extension of credits and the acquisition of physical capital. The efficiency follows a distribution
of the unit mean and is observable only to entrepreneurs. An entrepreneur defaults if the realized efficiency is too low
since the return on capital depends on the efficiency. Upon default, banks must acquire the level of individual efficiency

61
uncertainty level and the recovery rate of the defaulting loans determine the degree of financial

friction jointly. Holding the monitoring cost (uncertainty) constant, risk premium increases in the

level of uncertainty (monitoring cost).

The motivation for regime-switching is both anecdotal and quantitative. Anecdotally, the

regime-switching friction appears necessary since the post-1980 data presents significant and

recurrent swings in the corporate bond spread defined by the BAA-rating bonds over 10-year

treasury bills. The spread rises in recessions and declines in expansions, where rises are associated

to the tightening of credit conditions and declines correspond to loosening in credit conditions.

Many authors (e.g., Reinhart and Rogoff (2008)) have documented that credit conditions drastically

loosened leading up to the financial crisis. In BGG’s framework, it amounts to a drastic reduction

in idiosyncratic uncertainty. Quantitatively, Linde et al. (2016) show with an standard medium-

scale DSGE model that the smoothed risk-premium shocks and investment-specific shocks are

non-Gaussian with fat left tails closely related to recessions. We estimate the smoothed uncertainty

and seize the entrepreneur’s realized return on capital after paying a monitoring cost proportional to the realized return.
An increase in the monitoring cost decreases the recovery rate of defaulting loans. The aforementioned asymmetric
information and the costly-state-verification problem motivate banks to formulate financial contracts against the default
risk derived from the uncertainty defined as the dispersion of capital efficiency. With the contract, the equilibrium pins
a single pair of interest rate (price) and the leverage ratio (quantity, as a function of entrepreneurial net worth) because
risk-neutral entrepreneurs care only about their expected return. The uniqueness, in turn, implies a single efficiency
threshold for all entrepreneurs, a value below which default is optimal. Holding constant the threshold, an increase in
the uncertainty (dispersion of efficiency) leads to higher default risk. The risk forces banks to charge a higher interest
rate to guarantee a non-negative profit, thereby depressing the aggregate investment by making credit and investment
overall more expensive. The risk-free rate corresponds to the case in which uncertainty is zero. Therefore, banks must
demand a higher price for the risky bonds compared to the risk-free bonds to compensate the loss derived from the
idiosyncratic uncertainty.

62
(Figure 3.1) from the core model of CMR2, and demonstrate that the model without time-varying

financial conditions is inadequate to explain the dynamics of spread. The estimated model under

a fixed regime of financial condition reveals a two-decade disconnect between credit spreads and

uncertainty in the mid-1990s and mid-2000s, where the model appears to overstate the level of

uncertainty by a significant margin. The heightened uncertainties in early 1980s also raise questions

in explaining the faster recovery rate of investment relative to the aftermath of the Great Recession.3

A potential answer to the slow rebound rate lies in the agents’ expectations. With a bleak outlook

of the financial condition, agents substitute out future investment for current investment since the

credit appear expensive going forward, resulting in slow future growth.

Regime-switching is a convenient device to account for the recurrent alterations in the degrees

of financial frictions manifest in the corporate bond spread. Moreover, the transition probability

matrix naturally characterizes agents’ outlook of the financial market nd influences the pathes

of macroeconomic variables. In this paper, we consider a two-regime synchronized switching

between two levels of monitoring costs and two steady-state means of the uncertainty process

as in a conventional Markov switching model (Hamilton, 1989a; Kim and Nelson, 1999). The

switching monitoring cost follows a two-state Markov chain, and the log uncertainty process

2Compared to CMR, the model assumes away term structure, news shocks, and all distortionary taxation. The
data set contains eight standard macro variables and three financial variables spanning from 1985q1 to 2010q2, and is
inherently a subset of CMR.

3Figure 3.C.1 of Appendix 3.C reports the smoothed uncertainty given CMR’s posterior mode and compares it to
our estimates. The patterns of the two series are consistent: a disconnect between uncertainty and risk premium in
the mid-1990s and mid-2000s; spikes in the 1982 and the 2008 recessions. The smoothed series are of different sizes
because the posterior estimate of uncertainty shock volatility σe,σ is 0.0369 which halves the CMR value 0.0072.

63
 Figure 3.1: Credit Spread Against Idiosyncratic Uncertainty

BAA-10YTB
3.5 Uncert. (Fix Regime)
3

2.5
2

1.5 1

0.5 0

-0.5 -1

-1

-2
1985 1990 1995 2000 2005 2010

Note: The dashed dark curve (right scale) plots the standardized smoothed uncertainty in natural
log produced at the posterior mode of the estimated quantitative model with financial friction in
fixed regime. The data spans from 1981:Q1 to 2010: Q2. The blue curve (left scale) plots the
standardized credit spread in natural log defined by the spread of BAA rating corporate bonds
over the U.S. 10-year treasury bill in constant maturity. The credit spread is observable. The plot
differentiate left and right scale because the spread data is heavily skewed toward right, while the
estimated uncertainty is relatively symmetric.

follows a stationary autoregression with switching in the means. We associate the Regime-1 to

low uncertainty and Regime-2 to high uncertainty, and allow data to determine the associated

monitoring cost under each regime. Our consideration encompasses the standalone switching

in uncertainty as a special case since we do not require monitoring costs to be different across

regimes. Under a fixed regime, the model is in the vein of CMR. With switching regimes, the

model is related to lhuissier and Tripier (2016) and Linde et al. (2016) (LSW), where the former

64
identifies a RS-DSGE model from an estimated MS-SVAR model, and the latter analyzes the effect

of switching monitoring costs in GDP forecast.

We emphasize the expectation effect arises from the changes in the state transition probability

matrix. To understand the expectation effect, we first document the different effects of uncertainty

shocks, steady state uncertainty level and monitoring cost under a fixed regime, then analyze the

differences between the generalized impulse responses perturbing the transition probabilities. In

these simulation exercises, all structural parameters are estimated for the U.S. data in the fixed-

regime model, the switching steady-state uncertainties, monitoring costs calibrated to resemble

the bank loans’ default rate and corresponding estimates of CMR and LSW, and we consider

regime-specific generalized impulse responses at the impact date. The simulation results suggest,

an increase in the transition probability to the low uncertainty regime increases the rebound rate of

investment after an adverse uncertainty shock. A reinforcing feedback that increases the transition

probability to the low friction regime further mitigates the impact of the adverse shock.

Motivated by the simulation exercises, we further examine the empirical relevance of time-

varying transition by estimating the regime switching model of constant transition probability on

samples excluding and including the Great Recession (1990q1-2005q2 and 1990q1-2010q2). The

estimates on these periods show comparable structural parameters, except for those of the transition

probabilities. The shift of transition probabilities on different samples is economically significant.

In particular, the pre-2005 estimate is associated to a significantly more optimistic outlook of the

financial market compared to the estimate of the 1990-2010 sample. The change in transition

probabilities over time suggests that constant transition probability is empirically inadequate.

Systematic analysis of the time-varying transition probabilities requires an extension to the

switching model with constant transition matrix. To this end, we adopt the feedback mechanism

65
of Chang et al. (2017), where the transition probabilities are written as functions of the linear

combination of historical structural shocks and an exogenous innovation. It is indeed natural to

expect regime dynamics of financial conditions tied to economic fundamentals. The regime switch

is driven by a stationary AR(1) regime factor and a threshold parameter. It is Regime-1 if the

latent factor is below the threshold, and Regime-2 if otherwise. The innovation of the latent factor

can be decomposed to be a linear combination of past fundamental shocks and an exogenous

shock. The vector of coefficients for the linear combination of fundamental shocks has unit length,

hence the coefficients can naturally be interpreted as the percentage contributions of each shock

to the regime shifts. We estimate the model on 1981q1-2010q2 and identify two regimes with

Regime-1 characterized by low uncertainty and high monitoring cost. The estimated uncertainty

better explains the drops in corporate bond spreads observed in the mid-1990s and mid-2000s,

compared to the fixed-regime model. We] identify the contributions of each shock to the change in

agents’ outlook of financial market and find the structural shocks contribute to more than 99% of

the variation in the regime factor. In particular, most positive supply shocks push the economy to

the high uncertainty regime with the exception of persistent technological shock, while the pattern

of demand shocks’ effects remains unclear.

The paper is planned as follows. Section 3.2 relates this paper to the literature. Section

3.3 details the benchmark model, and the solution method. Section 3.4 discusses the effects of

switching financial conditions and reports simulation results. Section 3.5 presents the empirical

evidence of time-varying outlook of the financial conditions. Section 3.6 provides a systematic

analysis of the regime-switching model of time-varying transition and identify the sources of the

time-varying outlooks, followed by the concluding remarks in Section 3.7.

66
3.2 Literature

This paper is related to a vast body of literature concerning uncertainty. The relationship between

uncertainty and output growth in the business cycle frequency and the transmission mechanism

from one to the other has received substantial attention since the influential work of Bloom (2009).

Uncertainty is a key suspect of the depth of the Great Recession. Proxies of uncertainty rise sharply

with the NBER-based recessions (Bloom, 2009; Bloom et al., 2018; Baker et al., 2016; Jurado et al.,

2015). Evidence from a dynamic factor model suggests two classes of highly correlated shocks,

namely, credit supply and financial/political uncertainty, were the primary drivers of variation in

growth of detrended GDP and employment during the Great Recession (Stock and Watson, 2012).

Also, heightened uncertainty can, in theory, exert adverse effects on output growth through channels

of real-option, risk aversion, and risk premia; see Bloom (2014) for a general survey.4

This paper relates to works concerning policy effectiveness in an economy with uncertainty.

On the one hand, the real-option effect in uncertain times implies temporarily less effective fiscal

policies because firms are more cautious in responding to price changes due to irreversible invest-

ment (Bloom et al., 2018). On the other hand, in line with the real-option theory, evidence from

structural VARs suggests dampened effects of monetary policy shocks in more uncertain times

(Aastveit et al., 2017). Additionally, the effect of monetary policy instruments on macro variables

is at best indirect (Bernanke and Kuttner, 2005). Monetary policy actions more directly influence

financial markets by affecting asset prices and returns. Empirically, monetary policy surprises are

4As discussed in Bloom (2014), heightened uncertainty can also potentially exert positive effects on output through
growth option and Oi-Hartman-Abel effects. Quite plausible is that a combination of all channels was effective during
recessions. For the Great Recession, however, the adverse effects likely dominate.

67
shown to significantly impact on asset prices, which is primarily associated with the changes to

risk premia (Bernanke and Kuttner, 2005; Drechsler et al., 2018).

Models on the real-option channel typically predict a relatively quick overshoot of output several

periods after the initial reduction due to high uncertainty. The slow recovery following the great

slump thus presents a challenge as Bachmann et al. (2013) in a VAR analysis with standard recursive

identification strategy detect no evidence suggesting quick rebound and overshooting effect in the

U.S. data. Gilchrist et al. (2014) provide strong VAR evidence suggesting credit spread as a critical

conduit of uncertainty transmission in the U.S. during 1963Q3 - 2012Q3, and theorize that increases

in firm risk lead to a rise in bond premia and the cost of capital which, in turn, triggers a prolonged

decline in investment activity. Christiano et al. (2014) demonstrate fluctuations in uncertainty

can generate sizable and persistent reductions in output and argue it is the determinant of the

US business cycle using an estimated medium-scale DSGE model with BGG financial friction,

a sharply different result concerning the contribution of a conventional set of structural shocks

compared to the estimated model of Christiano et al. (2005). Moreover, Caldara et al. (2016) argue

with a VAR model identified via the penalty function approach that the Great Recession is likely a

consequence of the interaction of acutely elevated uncertainty and tightened financial constraint.

This paper also relates to the literature on the expectation formation effects. Liu et al. (2011)

examine the importance of the expectation formation effects of regime switching in monetary

policy in a DSGE model. They show that the possibility of regime shifts in policy can significantly

influence agents’ expectation formation and equilibrium dynamics. Bianchi (2013) estimates a

DSGE model with switching monetary policy regimes and finds that if agents in the 1970s had

anticipated a more aggressive response to inflation by Federal Reserve, inflation would have been

lower. Bianchi and Ilut (2017b) extend this work by allowing a mixture of monetary-fiscal policy

68
regimes.

On the topic of uncertainty and time-varying financial friction, this paper is related to Linde

et al. (2016). They demonstrate the macroeconomic implications of financial market friction in

a medium-scale DSGE model with BGG financial accelerator and regime-switching monitoring

cost, without further discussion of the expectation effect. A medium-scale model with financial

accelerator in their setting does not generate quantitatively important amplification effect of other

macroeconomic shocks, and that the Baa-Aaa spread is mostly driven by financial shocks. Con-

sequently, the model is incapable of predicting the output growth in 2008:Q4 with reasonable

predictive density. By modeling switching monitoring cost, they generate non-zero predictive

density on observing the output growth in 2008:Q4. Extending upon this work, lhuissier and

Tripier (2016) estimate a similar model by minimizing the distance between the impulse responses

of an identified Markov-switching VAR (MS-VAR), and those of the structural model. Driven by

the VAR evidence, they conclude that as the amplification effect diminishes as agents grow more

confidence in the outlook of the financial market. There are three essential differences between my

approach and theirs. First, we directly illicit the expectation formation effect from the structural

model, while theirs relies heavily on the specification of the MS-VAR. Second, they implicitly

assume a first-order log-linear approximation of the DSGE solution around a regime-independent

steady state, whereas we solve the model around two regime-specific steady states. Third, and

more importantly, the expectation effect carries different meaning in their work. They focus on

the amplification of adverse shocks when agents are less confident. we focus on the slow recovery

arises from the intertemporal substitution effect induced by the pessimistic outlook.

69
3.3 Model

The model features the financial accelerator and regime switching parameters in a Smets and

Wouters (2007) medium-scale DSGE model. The model under fixed regime is a special case of

CMR. Specifically, the model assumes away term structure, news shocks and distortionary taxation.

The idiosyncratic uncertainty in the financial accelerator mechnism is defined as the time-varying

standard deviation of the idiosyncratic efficiency for the physical capital. Bloom et al. (2018)

defines idiosyncratic uncertainty to be the standard deviation of the idiosyncratic shocks to the total

factor of production, which is similar to the definition we adopt.

The primary agents in this model are entrepreneurs and households. Entrepreneurs issues

one-period corporate bonds to the banks to purchase physical capital and service effective capital

to production firms. Entrepreneurial profit is non-zero in the bond market for there is asymmetric

information between entrepreneurs and banks; banks do not observe the idiosyncratic efficiency

level unless a monitoring cost is paid. The profit contributes to the next period’s net worth of en-

trepreneurs. Households consume, save and produce differentiated labor and homogeneous physical

capital. To introduce expectation formation effect, we consider a regime switching autoregressive

process of uncertainty, as well as the regime switching monitoring costs for banks. The switching

of uncertainty and monitoring cost can be interpreted as the jumps in the fundamentals of the goods

market and the depth of the agency problem in the financial market.

In this section, we first provide details of the quantitative model and the regime dynamics that

allows time-varying transition through feedback, as well as nests the conventional RS-DSGE model

as a corollary. Then we discuss the assumption over agents’ information set within the RS-DSGE

framework. Finally, we present details of the solution method.

70
3.3.1 The Real Sector

Household

There is a representative household choosing {Ct, K̄t+1, It, Bt+1 } to solve for problem

∞ 1
ht (i) 1+σ L
( Z )
t
X
max E0 β ζ c,t log(Ct − bCt−1 ) − ψ L di (3.3.1)
t=0 0 1 + σL

with β ∈ (0, 1) the discount factor, b ∈ [0, 1) the habit formation parameter, σ −1
L the Frisch

elasticity of labor hours and ψ L the labor disutility parameter. In the household’s preference, ζ c,t is

a preference shock, Ct denotes the per capita consumption, and ht (i) is the differentiated labor. At

each period t, the household faces budget constraint

Pt
!
Pt Ct + Bt+1 + t It + Q K̄,t (1 − δ) K̄t
Υ µΥ,t
Z 1 (3.3.2)
= Wt (i)ht (i)di + Rt Bt + Q K̄,t K̄t+1 + Πt
0

in which Pt is the nominal price for the consumption good, Bt is a one-period nominal bond with

rate of return Rt , It is the investment good, K̄t is the physical capital with market price Q K̄,t , K̄t+1

is the end-of-period physical capital, Wt (i) is the wage for the differentiated labor ht (i), and Πt is a

lump-sum transfer of dividend payment after taxation.

Equation (3.3.2) indicates that the household is the physical capital producer. After the produc-

tion of final goods in period t, the representative household produces the physical capital K̄t+1 at

the end of the period t, with technology

It
!!
K̄t+1 = (1 − δ) K̄t + 1 − S ζ I,t It (3.3.3)
It−1

71
where 0 < δ < 1 denotes the depreciation rate of the physical capital, and S(·) is an adjustment cost

to investment. The household has access to a technology that translates one unit of the homogeneous

consumption goods Ct into Υt µΥ,t units of investment good It with a constant growth rate Υ > 1

and a investment technology shock µΥ,t . The relative prices of the investment good in terms of the

consumption good is 1/(Υt µΥ,t ).

The investment adjustment cost function S(·) is an increasing and convex function of form

 √ √ 
S(x t ) = e S (x t −x ss ) + e− S (x t −x ss ) − 2 2
00 00
(3.3.4)

in which x t = ζ I,t It /It−1 , x ss is the corresponding steady state value, and ζ I,t is the shock to the

marginal efficiency of investment (MEI). The curvature parameter S00 characterizes the cost of

(dis)investing away from the steady state.

Goods Market and Labor Market

A representative and competitive final good packer combines the intermediate goods Yt ( j) for

j ∈ [0, 1] to produce homogeneous good Yt with the following technology

"Z 1 # λ f ,t
Yt = Yt ( j) 1/λ f ,t
(3.3.5)
0

where λ f ,t ≥ 1 is the price markup shock. The j-th intermediate good is produced by a monopolist

with production function

Yt ( j) = max 0,  t Kt ( j) α (zt l t ( j)) 1−α − Φzt∗ .

( )
(3.3.6)

72
The shock to the total factor of production is dichotomized into a stationary shock  t and a shock

of stationary growth zt . In the production function, Kt ( j) represents the effective capital which is

a constant multiple of the physical capital K̄t ( j). The l t ( j) is the total amount of homogeneous

labor employed by the producer. There is a fixed cost Φzt∗ to ensure zero long-run profit in the

intermediary good market and to preclude entry and exit at the steady state. For the existence of a

balance growth path, zt∗ = zt Υ(α/(1−α))t .

We adopt the Calvo’s pricing scheme and let j-th intermediate good producer reoptimize the

price Pt ( j) with probability 1 − ξ p . And with probability ξ p , the producer set the price following

Pt ( j) = π̃t Pt−1 ( j) with indexation factor π̃t = (πt∗ ) ι (πt−1 ) 1−ι . The parameter ξ p characterizes the

price rigidity of the intermediary good market. The inflation rate of the final good Yt is defined to

be πt = Pt /Pt−1 , and πt∗ denotes the inflationary target in the monetary policy rule.

A representative and competitive labor packer demands differentiated labor service ht (i) for

i ∈ [0, 1] and combines them into homogeneous bundles of labor with technology

"Z 1 # λw
lt = ht (i) 1/λ w
di , (3.3.7)
0

with wage markup parameter λ w ≥ 1. The labor packer then sells l t to the intermediate good

producers for nominal wage Wt . The differentiated labor suppliers are assumed to adopt Calvo-

style friction as well. With probability 1 − ξ w , the i-th labor supplier reoptimize the wage rate Wt (i).

If otherwise, the labor supplier follows indexation rule Wt (i) = (µ z∗,t ) ιµ (µ z∗ ) 1−ιµ π̃w,t , where µ z∗ is

the growth rate of zt∗ in the deterministic steady state, and π̃w,t = (πt∗ ) ιw (πt−1 ) 1−ιw . The parameter

ξ w characterizes the wage rigidity in the differentiated labor market.

73
3.3.2 The Financial Sector

Entrepreneur and Financial Friction

Entrepreneurs are risk-neutral. After the production in period t, an entrepreneur with net worth

N ≥ 0 borrows Bt+1 (N ) from the banks to purchase physical capital K̄t+1 (N ) from households

following

Q K̄,t K̄t+1 (N ) = N + Bt+1 (N ). (3.3.8)

The physical capital is then turned into effective capital Kt+1 (N ) = ω K̄t+1 (N ) in production.

Following Bernanke et al. (1999), the efficiency level of capital is distributed as

σω,t
2
ωt ∼ log-normal *− , σω,t
2 +
(3.3.9)
, 2 -

to ensure a unit mean. Here, σω,t denotes a stochastic process of the level of idiosyncratic

uncertainty.

Upon realization of aggregate rates of return, prices and the efficiency shock to the physical

capital, this entrepreneur chooses the utilization rate ut+1 of the effective capital to maximize the
k . The ex post rate of return of this entrepreneur
return of capital for a competitive market rate r k+t

is given by

k − a(u
[ut+1rt+1 t+1 )]Υ t+1 + (1 − δ)Q K̄,t+1
−(t+1) P
k
Rt+1 = (3.3.10)
Q K̄,t

in which Q K̄,t denotes the nominal price of physical capital at period t. This equation means that

74
entrepreneurs receive income by servicing effective capital after an adjustment cost a(ut+1 ) and

reselling the depreciated physical capital back to households. The adjustment cost a(·) to variable

utilization is an increasing and convex function of form

a(u) = r k eσa (u−1) − 1 σa .

f g.
(3.3.11)

The curvature parameter σa > 0 characterizes the cost of capital utilization and r k is the steady

state rental rate in the model.

At each period, a portion of the entrepreneurs default when the realized idiosyncratic shocks to

physical capital fall too low. The financial market friction emerges from the asymmetric information

between banks and the entrepreneurs. Specifically, banks do not observe the idiosyncratic shocks

to the physical capital, and they must pay a monitoring cost µ proportional to the net worth of each

borrowing entrepreneur to acquire the realized efficiency level. Let ω̄t+1 denote the threshold that

divides the repaying entrepreneurs and the defaulting ones. The banks must demand a rate of return

Zt to loans from those repaying entrepreneurs such that

k
Rt+1 ω̄t+1 Q K̄,t K̄t+1 (N ) = Bt+1 (N ) Zt+1 . (3.3.12)

For each entrepreneur of net worth N at period t, the law of motion of net worth after receiving

transfer W e follows

Nt+1 (N ) = γt Rtk Q K̄,t−1 K̄t (N ) − Zt (Q K̄,t−1 K̄t (N ) − N ) + W e

f g
(3.3.13)

Banks receive zero profit after diversification in equilibrium. Given price and transfer, the repre-

75
sentative entrepreneur choose ω̄t+1, K̄t+1 to optimize expected return

k
( )
max Et [1 − Γt (ω̄t+1 )] Rt+1 Q K̄,t K̄t+1 (3.3.14)

subject to the bank’s zero-profit condition

 k
Γt (ω̄t+1 ) − µGt (ω̄t+1 ) Rt+1 Q K̄,t K̄t+1 = Rt+1 Bt+1 .

(3.3.15)

Following the notation of Bernanke et al. (1999), the expected monitoring cost of banks and the

expected gross share of profit going to the banks are, respectively,

µGt (ω̄t+1 ) = µΦ(mt − σω,t ) (3.3.16)

Γt (ω̄t+1 ) = Gt (ω̄t+1 ) + ω̄t+1 (1 − Φ(mt )) (3.3.17)

where Φ(·) is the CDF of a standard normal distribution and

!
1 2 .
mt = log ω̄t+1 + σω,t σω,t . (3.3.18)
2

Banks’ zero-profit condition implies the leverage ratio

k −1
Q K̄,t K̄t+1  Rt+1 
= 1 − Γt (ω̄t+1 ) − µGt (ω̄t+1 )  .
  
  (3.3.19)
Nt Rt+1
 

76
And the shadow price Λt at optimal choice ω̄t+1 is

Γt0 (ω̄t+1 )
Λt = . (3.3.20)
Γt0 (ω̄t+1 ) − µG0t (ω̄t+1 )

Therefore all entrepreneurs choose identical ω̄t+1 such that

k
Rt+1  R k 
 t+1 (Γt (ω̄t+1 ) − µGt (ω̄t+1 )) − 1 

( ω̄ )] + Λt  = 0.

Et  [1 − Γt t+1 (3.3.21)
 Rt+1  Rt+1  
 

The law of motion of total net worth after aggregation is

Nt+1 = γt (1 − Γt−1 (ω̄t ))Rtk Q K̄,t−1 K̄t + W e

f g
(3.3.22)

in which K̄t is the aggregated physical capital.

3.3.3 Policy Rules and Aggregate Resource Constraint

We consider the monetary policy rule in linearized form to be

" #
1 1
Rt − R = ρ p (Rt−1 − R) + (1 − ρ p ) α π (πt+1 − πt∗ ) + α∆y (∆yt − µ z∗ ) + σe,p e p,t (3.3.23)
4 400

p
where ρ p is the smoothing parameter,  t is the monetary policy shock, R is the steady state

quarterly interest rate, πt∗ is the inflation target, ∆yt is the quarterly growth in GDP and µ z∗ is the

corresponding steady state. The fiscal policy rule of government expenditure follows

Gt = zt∗ gt (3.3.24)

77
with gt an exogenous process, and Yt /zt∗ converges to a constant in the deterministic steady state.

The aggregate resource constraint is given by

Yt = Ct + It /(Υt µΥ,t ) + Gt (3.3.25)

+µGt−1 (ω̄t )(1 + Rtk )Q K̄,t−1 K̄t /Pt + a(ut ) K̄t Υ−t .

3.3.4 Shocks Processes

Table 3.1 gives a complete list of fundamental shocks. In addition, we consider a two-state switching

Table 3.1: List of Fundamental Shocks

Shocks Label
εt Transitory Tech. Growth Shock
µ z∗,t Persistent Tech. Growth Shock
gt Government Spending Shock
e p,t Monetary Policy Shock
πt∗ Inflation Target Shock
µΥ,t Investment-Specific Shock
γt Equity Shock
λ f ,t Price Markup Shock
ζ c,t Preference Shock
ζi,t Marginal Efficiency of Investment Shock
σω,t Uncertainty Shock

process for the monitoring cost following Linde et al. (2016),

µt = (1 − st ) µ + st µ. (3.3.26)

The underlying regime indicator st follows a Markov chain with a 2×2 transition matrix T. Governed

by the same chain of regime indicators, the idiosyncratic uncertainty follows a regime-switching

78
autorgressive process

log(σω,t /σ ss,t ) = ρσ log(σω,t−1 /σ ss,t−1 ) + σe,σ eσ,t, (3.3.27)

σ ss,t = (1 − st )σ ss + st σ ss, σ ss < σ ss . (3.3.28)

Note we do not impose an order for µ and µ and identify the ordering from data in the empirical

sections.

The inflation target is assumed to evolve following

log(πt∗ /πss ) = ρπ∗ log(πt−1

∗
/πss ) + σe,π∗ eπ∗,t . (3.3.29)

Following Christiano et al. (2014), we fix ρπ∗ = 0.975 and σe,π∗ = 0.0001 to accommodate the

downward trend of inflation in the data. The government spending shock follows

log{gt /[η g (Css + Iss )/(1 − η g )]} = ρg log{gt−1 /[η g (Css + Iss )/(1 − η g )]} + σe,g eg,t . (3.3.30)

The log-deviation to steady state of all other fundamental shocks, ζ c,t, ζi,t, γt,  t, λ f ,t, µΥ,t and µ z∗,t ,

follow a standard AR(1) process of form

log(ε x,t /ε x,ss ) = ρ x log(ε x,t−1 /ε x,ss ) + σe,x e x,t . (3.3.31)

In particular, the coefficients ρ p = 0 and ργ = 0 for the monetary policy innovation and the equity

shock, respectively.

79
 Table 3.2: List of Feedback Channels

Parameters Label
ρv,z persist. technological shock
ρv, transitory technological shock
ρv,γ equity shock
ρv,µΥ investment technology shock
ρv,ζi MEI shock
ρv,σ risk shock
ρv,λ f price markup shock
ρv,g government spending shock
ρv,p MP shock
ρv,π∗ inflation target shock
ρv,ζc preference shock

3.3.5 Regimes, Regime Factor and Feedback

We let stationary autoregressive process

wt = αwt−1 + νt, νt ∼ N(0, 1) (3.3.32)

to drive the regime following

st = 1 + 1{wt ≥ τ}, (3.3.33)

where τ is a threshold parameter. And following Chang et al. (2017), we consider intertemporal

correlation between the column vector of all historical structural shocks εt−1 and the regime factor

innovation νt of form

*. et−1 +/ ∼ N *.0, *. I ρε,ν ++ 0

// , ρε,ν ρε,ν < 1 (3.3.34)
ν
, t - ρ
, , ε,ν
0 1 --

where ρε,ν is the column vector of correlation coefficients between each structural shock and the

regime factor innovation. See a summary of feedback coefficients in Table 3.2.

80
 Four notes are in order. First, the feedback mechanism is a convenient way of introducing flexible

time-varying transition probabilities. The time-varying transition matrices are characterized by the

transition probabilities of stay in Regime-1, P1|1,t and the probabilities of switching from Regime-2

to Regime-1, P1|2,t ,

Z τ√1−α2w !
αw w
Φ ρε,ν τ − √ 2 − ρ0ε,ν εt dΦ(w)
−∞ 1−α w
P1|1,t = q (3.3.35)
Φ(τ 1 − α 2w )
Z ∞ !
αw w
√ Φ ρε,ν τ − √ − ρ0ε,ν εt dΦ(w)
τ 1−α 2w 1−α 2w
P1|2,t = q (3.3.36)
1 − Φ(τ 1 − α 2w )

with Φ(·) be CDF of standard normal and

 q 
Φ ρε,ν (w) = Φ w/ 1 − ρ0ε,ν ρε,ν . (3.3.37)

Second, P1|1,t and P1|2,t are time-invariant if ρ,ν = 0. The regime process (st ) is Markovian

in this case. Therefore, the conventional RS-DSGE models with constant transition probability is

encompassed in our framework as a corollary. Furthermore, Chang et al. (2017) shows the pair of

parameters (α, τ) maps one-to-one to a 2 × 2 transition matrix characterized by P1|1 and P1|2 .

Third, we may project innovation vt onto the space span by εt−1 and decompose

q
vt = ρ0ε,ν εt−1 + 1 − ρ0ε,ν ρε,ν ε v,t (3.3.38)

into a feedback term ρ0ε,ν εt−1 and an innovation term 1 − ρ0ε,ν ρε,ν ε v,t , where εt−1 and ε v,t are
p

81
independent standard normal random variables. The unit variance of vt has decomposition

11
X
var(vt ) = ρε,ν 2(i) + (1 − ρ0ε,ν ρε,ν ) (3.3.39)
i=1

in which ρε,ν (i) is the i-th element of the vector ρε,ν . Since the regime factor is exclusively driven

by its innovation vt , it is thus natural to interpret ρε,ν 2(i) × 100 as the percentage contribution of i-th

structural shock to the regime shift. Likewise, ρ0ε,ν ρε,ν × 100 is the total contribution of structural

shocks to regime shift in percentage term.

Fourth, there is an MA(∞) representation of wt by the stationarity assumption

∞
X
wt = φ k νt−k (3.3.40)
k=0
X∞  q 
= φ k ρ0ε,ν εt−1−k + 1 − ρ0ε,ν ρε,ν η t−k ,
k=0

with φ0 = 1. Hence the latent factor wt is a linear function of the historical structural shocks and

the factor innovations to the infinite past. In the case of 1 − ρ0ε,ν ρε,ν ≈ 0,
p

∞
X
wt ≈ φ k ρ0ε,ν εt−1−k . (3.3.41)
k=0

In this case, the latent factor wt can be approximated by a linear function of the historical structural

shocks, and is easily written into a linear function of historical state variables.

82
3.3.6 Agents’ Information Set

We confine the analysis to the framework of RS-DSGE models, and assume agents’ information set

exclude the regime factor. Indeed, agents in a RS-DSGE model require only the information of the

current regime and the transition probabilities to make fully informed forward-looking decisions.

It is intuitive, rather than technical, to call an agent “optimistic" when the transition probability

place more weight on the low uncertainty regime in the future regardless of the current regime. We

use "optimistic" and "pessimistic" to highlight the difference in the transition probabilities instead

of reflecting agent’s subjective view of the future financial market.

3.3.7 Solution Method

For st = i, st+1 = j, and i, j ∈ {1, 2}, we look for regime-dependent policy functions

Xt = Ti (Xt−1, εt ), (3.3.42)

that solve the system of equations of first order conditions and constraints

 2 
X
0 = Et  pi, j (εt ) f i (T j (Ti (Xt−1, εt ), εt+1 ), Ti (Xt−1, εt ), Xt−1, εt )  .

(3.3.43)
 j=1 | {z } | {z } 
 X t+1 Xt 

We apply the 1-st order perturbation method proposed by Maih and Waggoner (2018), which

features state-dependent policy functions perturbed around state-dependent steady states x̄ i and

perturbation parameter σ in the time-varying transition matrix pi, j (εt ). Specifically, we consider

the solution

Xt = Ti (Xt−1, σ, εt ), i = 1, 2 (3.3.44)

83
of perturbation parameter σ ∈ [0, 1] such that

Ti (Xt−1, 1, εt ) = Ti (Xt−1, εt ) , (3.3.45)

Ti x i, 0, 0 = x̄ i,


(3.3.46)

to the system of equations

 X

0 =Et  pi, j (σ, εt ) f i T j (Ti (x t−1, σ, εt )
 j=1,2
# (3.3.47)


+ (1 − σ) x j − Ti x i, 0, 0 , σ, σεt+1 ), Ti (x t−1, σ, εt ) , x t−1, εt ) ,

where

 σpi, j (εt ) for i , j
pi, j (σ, εt ) = 

(3.3.48)
 1 − σ 1 − pi,i (εt )


for i = j

Note the perturbed system of equations (3.3.47) reduces to

0 = f i ( x̄ i, x̄ i, x̄ i, 0), (3.3.49)

at the steady state x̄ i when σ = 0, εt = 0 given (3.3.46), and is equivalent to (3.3.43) when σ = 1.

Moreover, the transition probability at the expansion point is an identity matrix. The approximate

solutions at expansion points are the state-dependent policy functions assuming the regime is fixed.

Finally, the feedback effects disappear in the policy functions by 1-st order solution, and show up

in the transition probability functions governing the regime dynamics. We leave the details of the

perturbation procedure in Appendix Section 3.A.

84
3.4 Effects of Switching Financial Conditions

Christiano et al. (2014) show that the uncertainty shock σ is the most important driver of the U.S.

business cycle. We follow this route and study the impact of uncertainty shock to macroeconomic

variables when there is a possibility of switching fundamentals of financial friction.

We begin by simulating the impulse responses of these variables under a fixed regime with

different levels of steady state uncertainty, monitoring cost, and shock size. It is essential to under-

stand the different effects of the uncertainty shock under fixed but different financial conditions. We

show that, depending on the specifics of the financial condition, a shock to uncertainty generates

different effects on key macroeconomic variables such as consumption, investment, labor hours,

and output.

We then extend the exercise to obtain generalized impulse responses of macroeconomic variables

under switching regimes, assuming rational agents with information of current regime but uncertain

about the future regimes, and compare the responses under “optimistic" and “pessimistic" outlook

of the financial market. In this exercise, we stress the expectation effect resulted from different state

transition probability matrices. Our simulation demonstrates quantitatively significant expectation

effect under a calibration realistic to the U.S. data.

3.4.1 Effects of Financial Friction in Fixed Regime

The degree of friction in the CSV problem at a steady state is determined by the level of idiosyncratic

uncertainty and the monitoring cost jointly. The degree of financial friction increases in the level of

uncertainty after a shock, as well as of the two parameters characterizing the steady state uncertainty

and the monitoring cost, holding the other parameter constant. However, these three factors exert

85
different effects on the macroeconomic variables. We document these differences in a simulation

exercise in which parameters are either calibrated or estimated to fit the U.S. data. Under this

simulation setup, we also rank the quantitative importance of these three factors.

The simulation model is of a fixed regime and a corollary of the benchmark model introduced

in Section 3.3. Specifically, we fix steady state uncertainty σ ss = σ̄ ss = σ ss , state-dependent

monitoring cost µ = µ̄ = µ and feedback coefficients ρε,ν = 0. The simulation model is also a

simplified version of CMR, which assumes away the long-term bond, news shocks, and taxation. In

the simulation setup, we follow CMR in the calibration of a subset of deep parameters and estimate

the linearized model for the rest of the parameters with data and priors identical to CMR. Table 3.1

presents a list of the calibrated parameters. Also, as in CMR, we do not directly estimate the steady

state uncertainty. Instead, since there is a one-to-one mapping between the steady state default rate

of corporate bonds F (ω̄) and the steady state uncertainty σ ss , we estimate the bond default rate and

back out the uncertainty. Table 3.2 reports the priors and posterior modes of the parameters. The

posterior modes are, by and large, similar to the posterior mode in the benchmark case of CMR.

Table 3.C.1 in Appendix 3.C compares the CMR posterior modes and ours.

Table 3.1: Calibrated Parameters

Parameter Label Value

β discount rate 0.9987
σL curvature, disutility of labor 1.0000
ψL disutility weight on labor 0.7705
λ w,ss s.s. markup, labor 1.0500
µz growth rate of economy 0.4100
Υ trend of investment technology 0.4200
δ capital depreciation rate 0.0250
α capital share 0.4000
λ f ,ss s.s. markup, intermediate good 1.2000
γss s.s. survival rate of entrepreneurs 0.9850
We transfer to entrepreneurs 0.0050
ηg s.s. spending-to-gdp ratio 0.2000
π∗ s.s. inflation target 2.4300

86
 Table 3.2: Priors and Posterior Mode: Fixed Regime

Prior Pmode
Parameter Label Dist. Mean SD 1-Regime
b consumption habit B 0.5 0.1 0.7746
F (ω̄) probability of default B 0.007 0.0037 0.0145
µ monitoring cost B 0.275 0.15 0.1838
σa curvature, utilization cost N 1 1 1.8454
S00 curvature, invest. adjust. cost N 5 3 12.0885
απ MP weight on inflation N 1.5 0.25 1.0818
α∆y MP weight, output growth N 0.25 0.1 0.3620
ρp MP smoothing B 0.75 0. 1 0.8481
ξp price rigidity B 0.5 0.1 0.7981
ι price index B 0.5 0.15 0.8710
ξw wage rigidity B 0.75 0.1 0.8243
ιw wage index, inflation target B 0.5 0.15 0.4862
ιµ wage index, persist tech. growth B 0.5 0.15 0.9333
σe,λ f stddev price markup invg2 0.002 0.0033 0.0116
σe,µΥ stddev investment price invg2 0.002 0.0033 0.0040
σe,g stddev government spending invg2 0.002 0.0033 0.0253
σe,µz stddev persistent technological growth invg2 0.002 0.0033 0.0073
σe,γ stddev equity invg2 0.002 0.0033 0.0039
σe, stddev transitory technology invg2 0.002 0.0033 0.0047
σe,p stddev MP invg2 0.002 0.0033 0.5049
σe,ζc stddev consumption preferece invg2 0.002 0.0033 0.0259
σe,ζi stddev MEI invg2 0.002 0.0033 0.0209
σe,σ stddev unanticipated uncertainty invg2 0.002 0.0033 0.0369
ρλ f AR price markup B 0.5 0.2 0.9959
ρ µΥ AR price of investment good B 0.5 0.2 0.9928
ρg AR government spending. B 0.5 0.2 0.9111
ρ µz AR persistent technological growth B 0.5 0.2 0.1035
ρ AR transitory technology B 0.5 0.2 0.9928
ρσ AR uncertainty B 0.5 0.2 0.8977
ρζc AR preference B 0.5 0.2 0.9830
ρ ζi AR MEI B 0.5 0.2 0.4051

Consider first the effect of uncertainty shocks to the consumption, investment, labor hours, and

GDP. We simulate impulse responses of these four variables under a normal and a large uncertainty

shock. The steady state default rate is set to be 1%, and the monitoring cost is set to be 21% since

our posterior estimate of monitoring cost is 18%, which is slightly below 20%, the empirically

relevant lower bound for the U.S. data in the literature. Figure 3.1 compares the IRFs in which the

solid lines represent the responses to a normal shock, and the dashed lines represent the responses

87
to a large shock. The IRFs are reported in levels. The effect of an uncertainty shock is linear and

in its size since the solution is a 1-st order approximation, and it does not affect the steady state of

these macroeconomic variables.

Figure 3.1: Normal vs. Large Uncertainty Shock in Fixed Regime

1.2

1.1
I

1
0 5 10 15 20 25 30 35 40

1
Y

0.99

0.98
0 5 10 15 20 25 30 35 40

1.9
C

1.85
0 5 10 15 20 25 30 35 40
1.25
H

1.2 1 5

0 5 10 15 20 25 30 35 40

Note: The figure reports the generalized impulse response functions with respects to a one standard
deviation uncertainty shock under high-uncertainty steady state. The solid curve presents the
benchmark low-friction case, and the dashed curve presents the high-friction case.

To gain additional insight, consider next the effect of steady state uncertainty level to the same

set of macroeconomic variables. We compare IRFs under steady state corporate bond default rates

1% and 2% to a one standard deviation uncertainty shock. The associated steady state uncertainty

levels are 0.24 and 0.40, respectively. The setup is empirically relevant since the estimated default

rate is 1.4% under a fixed regime, and 1% and 2% under switching regimes. The monitoring cost is

set to be 21%. Figure 3.2 compares the IRFs with the solid lines representing the low friction case

88
and the dashed lines representing the high friction case. The steady state uncertainty levels enter

into the steady state calculation and affect the IRFs of all macro variables in a persistent and non-

linear manner. The negative effect of higher steady state uncertainty emerges in all horizon for all

variables except for the GDP. The intertemporal substitution effect plays an important role; agents

in a persistently high-uncertainty economy trades more future investment for current investment

since the future investments are more expansive than the ones in a low-uncertainty economy.

Figure 3.2: Low vs. High Steady State Uncertainty in Fixed Regime

1.1
I

1.05

0 5 10 15 20 25 30 35 40

1
Y

0.998
F( ) = 0.1 F( ) = 0.2
0 5 10 15 20 25 30 35 40

1.9
C

1.88
0 5 10 15 20 25 30 35 40

1.22
H

1.21

1.2
0 5 10 15 20 25 30 35 40

Note: The figure reports the generalized impulse response functions with respects to a one standard
deviation uncertainty shock under high-uncertainty steady state. The solid curve presents the
benchmark low-friction case, and the dashed curve presents the high-friction case.

Moreover, we consider the effect of different monitoring costs. With the steady state default

rate equals 1%, we compare the IRFs under 21% and 27% monitoring costs in Figure 3.3. The

solid lines represent the IRFs of low monitoring cost, and the dash lines represent the ones of the

89
high cost. The monitoring cost within the “empirically relevant" range according to the simulation

exercise exerts a small effect on the level of GDP, and shift the steady states of the macroeconomic

variables almost exclusively without too much effect on the dynamics.

Figure 3.3: Low vs. High Monitoring Cost in Fixed Regime

1.2

1.15
I

1.1

0 5 10 15 20 25 30 35 40

1
Y

0.998
= 0.21 = 0.27
0 5 10 15 20 25 30 35 40

1.92
C

1.91
1.9

0 5 10 15 20 25 30 35 40
1.24
H

1.22

0 5 10 15 20 25 30 35 40

Note: The figure reports the generalized impulse response functions with respects to a one standard
deviation uncertainty shock under high-uncertainty steady state. The solid curve presents the
benchmark low-friction case, and the dashed curve presents the high-friction case.

Finally, the solid blue lines in Figure 3.1, Figure 3.2, and Figure 3.3 are identical and work as a

benchmark to compare the quantitative importance of the three factors in this simulation exercise.

Clear it is that the steady state uncertainty level is the more important factor in generating negative

impact with non-trivial effects on the dynamics of macroeconomic variables in all horizons, except

for the GDP. The uncertainty shock is more important than other factors in generating short-

horizon declines in the GDP. The IRFs in deviation of the steady states are reported in Figure 3.C.2,

90
Figure 3.C.3, and Figure 3.C.4 of Appendix 3.C.

3.4.2 Switching Financial Conditions and Expectation Effect

Anecdotal and quantitative evidence motivate the regime-switching modeling. An essential set of

parameters in the regime-switching model are the state transition probabilities. Rational agents form

the expectation of future states based on the transition matrix. If assuming time-varying transition,

the proper definition of the total effect of switching regimes should combine a direct effect caused

by a shock, and the expectation effect caused by a change in the transition probabilities. We have

discussed the direct effect of tightening financial condition in the preceding section. In this section,

we explore the expectation formation effect in the regime-switching model and demonstrate the

significance of expectation formation effect.

We consider without feedback the generalized impulse response function (GIRF) of a generic

macroeconomic variable X to one standard deviation shock of uncertainty eσ,t under transition

probability matrix P

GI XP (st, eσ,t, k) = E(Xt+k |Ft−1, st, eσ,t ; P) − E(Xt+k |Ft−1, st, ; P), k ≥ 0. (3.4.1)

The information set Ft−1 contains all shocks and states up to time t − 1. The impulse response

traces the evolution of X around its steady state under a realized regime at time t with unknown

future regimes. For different state transition probability matrices P1 and P2 , we define the k-th

period expectation effect of P1 on generic variable X relative to P2 to be

Expectation EffectPX,P
1
2
(st, eσ,t, k) ≡ GI XP1 (st, eσ,t, k) − GI XP2 (st, eσ,t, k). (3.4.2)

91
Note this definition of expectation effect deviates from the one adopted by Leeper and Zha (2003).

Table 3.3 provides a detailed setup for the simulation. The steady state probability of default

under high and low friction regimes are set to be 2% and 1% to echo the simulation exercises

in the preceding section. The monitoring costs are chosen to be 0.21 in the low friction regime

and 0.27 in the high friction regime to reflect the range of estimates in empirical studies. The

benchmark transition is set to have α = 0.9 and τ = 0 to represent relatively more uncertain

future. In particular, the unconditional high uncertainty regime probability under this setup is 0.5.

In the alternative setup, the transition is set to have α = 0.9 and τ = 1.2 to reflect a more and

optimistic outlook since the implied unconditional low uncertainty regime probability is 0.7. All

other parameters, if not specified in Table 3.1 and Table 3.3, are fixed at the estimated posterior

mode reported in Table 3.2.

Table 3.3: Simulation Setup of Expectation Effect

Parameter Label Pessimistic Optimistic

F (ω̄)1 Low probability of default 0.01 -
F (ω̄)2 High probability of default 0.02 -
µ1 Low monitoring cost 0.21 -
µ2 High monitoring cost 0.27 -
α Regime factor persistence 0.9 -
τ Threshold 0 1.2

Note: The “-" denotes the same value as in the “Pessimistic" case to highlight the difference.

Figure 3.4 reports the generalized impulse responses of macroeconomic variables to an adverse

uncertainty shock in the high uncertainty regime. The solid lines present the benchmark "pes-

simistic" case, and the dashed lines present the "optimistic" case. The GIRFs are reported as the

deviation away from steady states. The steady states are identical in both cases since they are state-

dependent. Clear is that as transition probability shift from “pessimistic" to “optimistic," the size

92
of the impulse response to uncertainty shocks change substantially. When agents are pessimistic of

the future state, the adverse effect of uncertainty is very persistent, and the size of the effect is two

times larger compared to the case in which agents are more optimistic.

Figure 3.4: Pessimistic vs. Optimistic Outlook in Switching Regimes

-0.1
I

-0.2

0 6 12 18 24 30 36
10 -3
0
-5 Pessimistic
Y

-10
Optimistic
-15
0 6 12 18 24 30 36
10 -3
0
-1
C

-2
-3

0 6 12 18 24 30 36

-0.05
H

-0.1
0 6 12 18 24 30 36

Note: The figure reports the generalized impulse response functions with respects to a one standard
deviation uncertainty shock under high-uncertainty steady state. The solid curve presents the
benchmark "pessimistic" case, and the dashed curve presents the "optimistic" case. The GIRFs are
computed from a simulation of 500 iterations.

3.5 Time-Varying Outlook of Financial Market

The key finding of this article is the significant expectation effect induced from the time-varying

outlook of the switching financial conditions. The empirical relevance of this finding must be

evaluated. In this section, we provide empirical evidence of time-varying transition probabilities

93
through the lens of a conventional RS-DSGE model with constant transition probability. Specifi-

cally, we show with an estimated model on sub-samples including and excluding the most recent

financial crisis, rational agents hold significantly different outlook of the financial market, while

the rest of the parameters are stable across sub-samples. In what follows, we first describe the data,

then provide details of the estimation procedure and results.

3.5.1 Data Set

The data contains quarterly observations of eleven variables spanning from 1981q1 to 2010q2. All

of the variables are in real and per capita terms. There are eight standard aggregate variables in

empirical analyses: GDP, consumption, investment, inflation, real wage, relative price of investment

goods, labor hours, and federal funds rate. Among these, the consumption, investment, credit, GDP,

net worth, the relative price of investments, and real wage are demeaned first order log-difference.

GDP is deflated by its implicit price deflator; real household consumption is the sum of

household purchases of non-durable goods and services, each deflated by its own implicit price

deflator; investment is the sum of gross private domestic investment plus household purchases

of durable goods, each deflated by its own price deflator. The per capita terms are divided by

the population over 16. Annual population data is obtained from the Organization for Economic

Cooperation and Development were linearly interpolated to obtain quarterly frequency. The real

wage is hourly compensation of all employees in non-farm business divided by the GDP implicit

price deflator. The short-term risk-free interest rate is the three-month average of the daily effective

federal funds rate. Inflation is measured as the logarithmic first difference of the GDP deflator. The

relative price of investment goods is measured as the implicit price deflator for investment goods

divided by the implicit price deflator for GDP. The labor hours are in log (per capita) levels, net of

94
the sample mean. Inflation is measured in level.

There are also three financial variables: credit to non-financial firms, the net worth of en-

trepreneurs, and credit spread. The credit to non-financial firms is taken from the flow of funds

dataset constructed by the U.S. Federal Reserve Board. The entrepreneurial net worth is the Dow

Jones Wilshire 5000 index. The credit spread is the difference between the interest rate on BAA-

rated corporate bonds and the rate of the 10-Year U.S. Treasury bill in constant maturity. The credit

spread and the risk-free rate are measured in level. The dataset is a subset of Christiano et al.

(2014). We exclude the term-spread data because the term-structure is not explicitly modeled in

our setup.

3.5.2 Sub-Sample Evidence of Time-Varying Outlook of Financial Market

In this section, we consider the regime-switching model with constant transition probability by

fixing the feedback parameters ρε,ν = 0. We estimate the model on two sub-samples. The first

sub-sample spans from 1981q1 to 2005q2 to account for the incoherent comovement between credit

spread and uncertainty implied by the fixed regime model. The second sub-sample spans from

1990q1 to 2010q2 to include the most recent financial crisis episode.

The prior distribution is an adaptation to the fixed regime prior and is adjusted based on

fixed regime estimates. The transition probabilities, regime-dependent default rates, and regime-

dependent monitoring costs are new parameters in the regime-switching model. We consider Beta

priors for these parameters. The 5th and 95th percentiles of the priors for the switching probability

from low to high friction regime P2|1 are set to be 0.001 and 0.1, while the corresponding values for

P1|2 are set to be 0.001 and 0.5. The 5th and 95th percentiles of the priors for monitoring costs are

set to be the empirically relevant lower and higher bound, 20% and 36%, in the literature. Moreover,

95
we fix the investment adjustment cost and capital utilization cost parameters to be the fixed regime

posterior modes, 12.089 and 1.845, respectively, because these parameters are relatively unstable

in sub-sample estimations if not fixed.5 Indeed, we expect similar estimates of these adjustment

costs on the two sub-samples because their data overlaps to a large extent. Table 3.1 reports the

prior and posterior mode estimates for both sub-samples. We leave a detailed description of the

Bayesian estimation in Appendix 3.B.1.

A sequel of notes concerning the posterior modes is in order. The pre-2005 and post-2005 esti-

mations yield similar estimates except for the transition probabilities of the regimes and the monetary

policy responses to inflation and output gap. The pre-2005 estimates suggest a less aggressive mon-

etary policy compared to the sub-sample, which includes the crisis episode. The posterior modes

of monitoring costs are (0.0996, 0.1184) in the pre-2005 sub-sample, and (0.1,0.1149) otherwise.

Both pairs are admittedly small regarding CMR’s comment on the empirical relevance. However,

these values are similar in magnitude compared to Linde et al. (2016)’s 0.03 and 0.08. The au-

toregressive coefficient of government spending shock is close to unity in both samples, suggesting

that it a unit-root process. Lastly, the autoregressive coefficients of uncertainty shock are 0.4324

and 0.3774 in the regime-switching model. These values are significantly smaller compared to the

fixed regime model since the persistence of σω,t is partially accounted for by the persistence of the

regimes characterized by transition probabilities.

The posterior modes of the transition probabilities from the high-friction to the low-friction

regime forms a sharp contrast in the two sub-samples, indicating a significant difference in agents’

5Both of the curvature parameters S 00 and σ a are estimated to be significantly larger in the pre-2005 sub-sample
than those of the later sub-sample. We report the estimation results in Table 3.C.2 and Table 3.C.3 of Appendix 3.C.

96
 Table 3.1: Priors and Posterior Mode: Switching Regimes, Constant Transition

Prior Posterior Mode

Parameters Dist Mean SD 1985-2005 1990-2010
b B 0.7 0.1 0.1440 0.9584
ξp B 0.8 0.1 0.8021 0.8630
αp N 2.5 0.25 3.8766 3.0903
ρp B 0.75 0.1 0.9536 0.9410
ι B 0.5 0.15 0.9875 0.9960
ιw B 0.5 0.15 0.0969 0.9154
ιµ B 0.5 0.15 0.7616 0.1244
α∆y N 0.25 0.1 0.0458 0.4850
ρλ f B 0.9 0.2 0.9990 0.9977
ρΥ B 0.9 0.2 0.9173 0.7585
ρg B 0.9 0.2 0.9445 1.0000
ρ µ∗ B 0.1 0.2 0.2810 0.0002
ρε B 0.9 0.2 0.8345 0.6844
ρσ B 0.9 0.2 0.6986 0.5736
ρζc B 0.9 0.2 0.1127 0.8004
ρ ζi B 0.9 0.2 0.9608 0.1289
σe,σ IG 0.05 0.04 2.3927 0.8332
Prior
Dist. Q0.05 Q0.95
ξw B 0.7 0.9 0.9099 0.5435
σe,λ f IG 0.0005 0.0015 0.0101 0.0195
σe,Υ IG 0.002 0.006 0.0047 0.0070
σe,g IG 0.001 0.0033 0.0322 0.0601
σe,µ∗ IG 0.003 0.01 0.0188 0.0734
σe,γ IG 0.003 0.01 0.0441 0.0318
σe,ε IG 0.003 0.01 0.0792 0.0813
σe,p IG 0.01 1 0.7066 0.4616
σe,ζc IG 0.003 0.01 0.1900 0.1532
σe,ζi IG 0.003 0.01 0.1331 0.0336
P2|1 B 0.001 0.1 0.0281 0.3017
P1|2 B 0.001 0.5 0.2271 0.0029
F (ω̄)1 B 0.003 0.01 0.0047 0.0030
F (ω̄)2 B 0.01 0.02 0.0067 0.0032
µ1 B 0.2 0.36 0.0695 0.0996
µ2 B 0.2 0.36 0.1260 0.1187

Note: The posterior modes correspond to estimations in which investment adjustment cost
parameter S00 and capital utilization parameter σa are calibrated at 12.0885 and 1.8454,
respectively. Both values are posterior modes of the fixed regime model estimated on sample
spanning from 1981:Q1 to 2010:Q2.

97
outlook of the financial market. Specifically, the transition probabilities are (0.065,0.232) and

(0.302, 0.003) in the pre-2005 sample and the post-2005 sample, respectively, suggesting a higher

chance of getting out of a financially stressful regime before 2005, and a higher chance of moving

into a stressful regime after 2005.

Figure 3.1 compares the standardized smoothed uncertainty produced by the fixed regime model

and the regime-switching model with constant transition probability on the 1990-2010 sub-sample.

The switching model with a constant transition on the sub-sample generates uncertainty process

with dynamics that better aligns with the credit spread series compared to the fixed regime model,

but it is challenging to produce uncertainty process that corresponds the data well on the full sample.

We report the full sample filtering result in Figure 3.C.5 of Appendix 3.C.

3.6 Sources of Time-Varying Uncertainty Outlook

The simulation results and sub-sample evidence made clear of the empirical significance of the

time-varying outlook of the uncertainty regimes. However, the conventional regime-switching

model suffers from numerous problems. Technically, a model allowing only constant transition

matrix is inadequate for a systematic analysis of the effect of time-varying transition matrices,

provided evidence of such time-variability exists. Economically, natural it is to associate the

agents’ outlook of the financial market with historical and current economic fundamentals.

Our framework with feedback allows for the systematic analysis required. Also, a key advantage

of our approach over conventional methods is that we identify the sources of regime-switching by

their contribution to the regime factor. In this section, we provide a systematic analysis of the

time-varying transition and its effects by allowing non-trivial feedback from structural shocks

98
 Figure 3.1: Credit Spread Against Idiosyncratic Uncertainty

10 -3

BAA-10YTB
Uncert. (Fix Regime)
6 Uncert. (Switch Regime)
3

5
2

3 1

2
0

0 -1

-1
-2

1985 1990 1995 2000 2005 2010

Note: The dashed dark curve and the diamond red curve (right scale) plot the standardized smoothed
uncertainty in natural log produced at the posterior mode of the estimated quantitative model without
and with regime switching financial friction, respectively. The data spans from 1981q1 to 2010q2.
The blue curve (left scale) plots the standardized credit spread in natural log defined by the spread
of BAA rating corporate bonds over the U.S. 10-year treasury bill in constant maturity. The credit
spread is an observable in the dataset. The plot differentiate left and right scale because the spread
data is heavily skewed toward right, while the estimated uncertainty is relatively symmetric.

one-period ahead. To this end, we begin by describing the filtering method which allows for the

feedback mechanism, followed by the details of estimation and empirical results, where we find

the identified feedback channels accounts more than 99% variations of the regime shifts. To show

empirical fitness, we also produce a more reasonable uncertainty-spread correlation compared to

the filtering results of Christiano et al. (2014) and the results by a conventional RS-DSGE model.

99
3.6.1 Filtering Method

The time-varying transition matrix in a DSGE model complicates the filtering problem since it,

in general, requires to keep track of a complete history of regimes. For a two-regime model,

the computation quickly becomes non-trivial as the history unfolds because the total number of

paths to track is 2T with T the length of data. Chang et al. (2018) propose a Kim (1994)-style

regime-switching filter to overcome this issue.

The filter assumes a state space model (SSM) of form

yt = Dst + Z st x t + Fst zt + Q st ut (3.6.1)

x t = Cst + G st x t−1 + Est zt + Rst  t (3.6.2)

with st specified by

wt = αwt−1 + νt (3.6.3)

st = 1 + 1{wt ≥ τ} (3.6.4)

allowing correlation between νt and  t−1 with vector of correlation coefficients ρ. Equation (3.6.1)

describes the relationship between data and the state variables in the model. Equation (3.6.2) is

the system of policy functions. The SSM is completed by the time-varying transition probabilities

characterized by (3.6.3) and (3.6.4). The filter performs the predict-update recursion and takes a

marginalizing-collapsing approach to approximate the likelihood function p(Y1:T |θ). For notation,

we do not differentiate the exact and the approximate likelihood, acknowledging that the likelihood

100
is always approximated in the empirical analysis. We leave a detailed description of the filtering

algorithm in Appendix 3.B.3.

3.6.2 Priors and Posterior Modes

We extend our analysis to three specifications on the full sample: regime-switching with constant

transition regime-switching with feedback, regime-switching with time-varying transition and con-

stant monitoring cost. In what follows, we use “exogenous switching" and “endogenous switching"

to describe the models with constant transition and time-varying transition, respectively.

We report in Table 3.1 and Table 3.2 the posterior modes of the constant parameters and

the regime-switching parameters, respectively. The priors are identical to the previous regime-

switching estimations with several exceptions. To allow for feedback, we consider uniform[−1, 1]

priors for all feedback coefficients ρε,ν , Beta prior for regime factor persistence α with Q0.05 = 0.5

and Q0.95 = 0.95 and Normal prior for threshold parameter τ with Q0.05 = 0 and Q0.95 = 1. Note

the unconditional low friction probability, (Φ(τ(1 − α 2 ))), decreases in α and increases in τ. The

priors over (α, τ) effectively put a prior on the unconditional low friction probability with Q0.05 = 0

and Q0.95 = 0.8.6

The regime-independent parameters are comparable across all three cases, but we identify

different estimates of parameters pertinent to the regime-switching. On the one hand, we identify

regime one in the endogenous model to be a mixture of low uncertainty and high monitoring cost,

and regime two a mixture of high uncertainty and low monitoring cost. On the other hand, we

find regime one in the exogenous model to be a mixture of low uncertainty and low monitoring

6Easy to check Φ(0) = 0.5, and Φ(1 − 0.52 ) = 0.8.

101
cost. The steady state levels of uncertainty are significantly different across regimes in all cases

considered. Nevertheless, the estimated regime-dependent monitoring costs are similar across

regimes in both the endogenous and exogenous switching models. By assuming a non-switching

µ, we still find regimes very similar to the exogenous and endogenous switching models. However,

the data decisively favors the models with switching uncertainty and switching µ, albeit the small

differences in µ and µ across regimes, because with switching monitoring cost, the natural log of

marginal data density of endogenous switching model and exogenous switching model increases

by 62.91 and 36.44, respectively.

The total contribution of feedback to regime factor innovation is more than 99%, indicating

the historical structural shocks almost exclusively drive the regime dynamics. This result further

supports the idea of endogenously evolving regimes. In our specification, a positive shock with

positive feedback further increase the regime factor, pushing the economy toward the high uncer-

tainty regime. We rank the feedback channels by their signs and sizes in Table 3.2. On the supply

side, a positive shock to the transitory TFP, equity, investment, marginal efficiency of investment,

or uncertainty results to an expectation of higher uncertainty in the market, while a positive shock

to persistent TFP reduces the expectation of uncertainty. On the demand side, a positive shock to

the price markup, monetary policy, or preference reduces the expectation of high uncertainty, while

a positive shock to the fiscal policy or inflation target leads to an expectation of higher uncertainty.

The positive feedback from the fiscal policy shock is close to zero and weak in size relative to all

other channels. On the other hand, the feedback channels of monetary policy shocks and inflation

target shocks are moderate in size and opposite in sign, indicating non-trivial expectation effects of

the monetary policy instruments on the macro-economy.

102
 Table 3.1: Priors and Posterior Mode: Endogenous Switching

Parameters Label Prior Pmode

Structural Endo Exo Const. µ
ξw wage rigidity B 0.9107 0.8549 0.8265
ξp price rigidity B 0.7103 0.7450 0.7769
b consumption habit B 0.9013 0.8730 0.8534
απ MP weight on inflation N 1.0340 1.0004 1.0841
α δy MP weight on output growth N 0.3018 0.2995 0.2873
ρp MP smoothing B 0.9148 0.8587 0.8462
ι price indexation B 0.2350 0.4491 0.6055
ιw wage indexation on inflation target B 0.2350 0.3684 0.6282
ιµ wage indexation on presiste tech. growth B 0.2361 0.7802 0.7973
ρ λ, f AR price markup B 0.7080 0.8759 0.8517
ρ µΥ AR investment specific technology B 0.9870 0.9704 0.9928
ρg AR government spending B 0.9207 0.9245 0.9021
ρ µz AR persistent technological growth B 0.0648 0.0689 0.0809
ρε AR transitory technological growth B 0.9928 0.9844 0.8713
ρσ AR risk B 0.9770 0.9827 0.9737

103
ρζc AR preference B 0.9774 0.8391 0.7834
ρ ζi AR marginal efficiency of investment B 0.6716 0.7754 0.7001
σe,λ, f std. dev. Price markup IG 0.0166 0.0108 0.0116
σe,µΥ std. dev. Investment specific technology IG 0.0039 0.0039 0.0040
σe,g std. dev. Government spending IG 0.0227 0.0221 0.0229
σe,p std. dev. MP IG 0.5656 0.6346 0.5815
σe,µz std. dev. Persistent technological growth IG 0.0078 0.0076 0.0073
σε std. dev. Transitory technological growth IG 0.0051 0.0047 0.0047
σe,γ std. dev. Equity IG 0.0074 0.0145 0.0050
σe,σ std. dev. Risk IG 0.0432 0.0826 0.1151
σe,ζc std. dev. Preference IG 0.0486 0.0310 0.0259
σe,ζi std. dev. Marginal efficiency of investment IG 0.0259 0.0209 0.0299
log-MDD 4021.8751 3995.4050 3958.9613

Note: The posterior modes correspond to estimations in which investment adjustment cost parameter S00 and capital utilization
parameter σa are calibrated at 12.0885 and 1.8454, respectively. Both values are posterior modes of the fixed regime model
estimated on sample spanning from 1981:Q1 to 2010:Q2. Log-Marginal Data Density is estimated using Laplace approximation.
 Table 3.2: Priors and Posterior Mode: Endogenous Switching, Cont’d

Parameters Label Prior Pmode

Switching Endo Exo Const. µ
α persist. of regime factor B 0.8709 0.9531 0.8131
τ threshold of regime factor N 0.7994 0.2495 0.0141
ρv,ε transitory tech. shock U 0.5469 - -0.5767
ρv,µΥ investment specific tech. shock U 0.4219 - 0.5767
ρv,ζi MEI shock U 0.3142 - 0.0211
ρv,π∗ inflation target shock U 0.2188 - 0.0002
ρv,γ equity shock U 0.1105 - 0.0215
ρv,σ uncertainty shock U 0.0979 - 0.0080
ρv,g gov. spending shock U 0.0511 - -0.0075
ρv,λ f price markup shock U -0.2121 - -0.5767

104
ρv,µz persist. tech. shock U -0.2422 - 0.0028
ρv,p MP shock U -0.2500 - 0.0049
ρv,ζc preference shock U -0.4312 - -0.0218
F (ω̄)1 default prob. (regime 1) B 0.0100 0.0100 0.0100
F (ω̄)2 default prob. (regime 2) B 0.0197 0.0200 0.0200
µ1 monitoring cost (regime 1) B 0.1212 0.0884 0.1258
µ2 monitoring cost (regime 2) B 0.1116 0.0999 0.1258
log-MDD (Laplace) 4021.8751 3995.4050 3958.9613

Note: The posterior modes correspond to estimations in which investment adjustment cost parameter S00 and capital utilization
parameter σa are calibrated at 12.0885 and 1.8454, respectively. Both values are posterior modes of the fixed regime model
estimated on sample spanning from 1981:Q1 to 2010:Q2.
3.6.3 Close the Gap between Uncertainty and Spread

The model with feedback provides a new prospective in reconciling the difference between corporate

bond spread and model implied uncertainty, as well as the different recovery rates between 1980s

recessions and the 2008 recession. Figure 3.1 plots the smoothed log uncertainty and smoothed

regime one probability (right axis) against the spread (left axis) from 1981 to 2010. Both series

of uncertainty and spread are standardized. First, the filtered uncertainty appears to fluctuate with

the spread more closely in all periods except for the mid-1990s. Second, Figure 3.2 report the

smoothed regime factors and high-uncertainty regime probabilities, respectively.7 The filtering

result reveals that the early 1980s recession and the 2008 recession are in different regimes albeit

the hikes in spread are similar in both periods.

7See Figure 3.C.6 in Appendix 3.C for a comparison between smoothed high-uncertainty regime probabilities and
NEBR recessions.

105
 Figure 3.1: Credit Spread Against Uncertainty, Full Sample Endogenous Switching

3.5

3
3

2.5
2

1
1.5

1
0
0.5

0 -1

-0.5
-2
-1

-3
1985 1990 1995 2000 2005 2010

Spread FS Uncert. RS Uncert.

Note: The dashed dark curve and the diamond red curve (right scale) plot the smoothed regime
probability of low uncertainty and the standardized smoothed uncertainty in natural log produced
at the posterior mode of the estimated quantitative model with regime switching financial friction
of time-varying transition, respectively. The data spans from 1981q1 to 2010q2. The blue curve
(left scale) plots the standardized credit spread in natural log defined by the spread of BAA rating
corporate bonds over the U.S. 10-year treasury bill in constant maturity. The credit spread is an
observable in the dataset. The plot differentiate left and right scale because the spread data is
heavily skewed toward right, while the estimated uncertainty is relatively symmetric. The shaded
areas mark the NBER recessions.

106
 Figure 3.2: Regime Factor and Identified High Uncertainty Regimes

-1

-2

-3
1985 1990 1995 2000 2005 2010

wt

Note: The diamond red curve plot the smoothed regime factor at the posterior mode of the estimated
quantitative model with regime switching financial friction of time-varying transition. The data
spans from 1981q1 to 2010q2. The dashed dark line plots the regime threshold. The shaded area
mark the high uncertainty regimes identified by the relationship between the regime factor and the
regime threshold.

107
3.7 Conclusion

Expectation effect of switching financial market exerts a significant impact on the macro-economy.

This paper examines the effect of expectation concerning the future financial market by extending

the DSGE framework to include regime-switching degrees of financial friction. Compared to a more

confident outlook, the expectation effect originated from a bleak outlook of the financial market

generates a more profound and more persistent decline of investment upon an adverse uncertainty

shock and a highly uncertain state. Empirical evidence suggests that the agents’ outlook of the

future condition during the most recent financial crisis is significantly more pessimistic compared

to the pre-2005 episode, contributing to a deeper recession and slower recovery. The paper further

identifies the structural shocks’ contribution to the time-varying outlook by their signs and sizes

and find that the historical shocks account for more than 99% of the variations in the regime

shifts, suggesting new channels through which the conventional policy instruments can influence

the macro-economy.

108
Appendix

Appendix 3.A Perturbation Method

This section describes the procedure that solves the system of equations (3.3.47). Let

v = (x t+1, x t, x t−1,  t ), (3.A.1)

z = (σ, x t−1,  t ). (3.A.2)

Given the unknown policy function Ti, i ∈ 1, 2, let

ui, j (z) = (σ, Ti (z) + (1 − σ)( x̄ j − x̄ i ), σ t+1 ), (3.A.3)

vi, j (z) = (T j (ui, j (z)), Ti (z), x t−1 . t ), (3.A.4)


 σpi, j (z) i,j
qi, j (z) =  ,

(3.A.5)
 1 + σ(pi, j (z) − 1) i = j

X
Fi (z) = pi, j (z) f i (vi, j (z)), (3.A.6)
j=1,2
Fi (z) = Et Fi (z). (3.A.7)

109
The perturbation procedure aims to find Ti such that Fi (z) = 0 for i = 1, 2. To simplify notation,

write the total derivatives

f g
D f i (v) = D x t+1 f i (u) D x t f i (u) D x t−1 f i (u) D t f i (u) , (3.A.8)
f g
DFi (z) = Dσ Fi (z) D x t−1 Fi (z) D t Fi (z) , (3.A.9)
f g
DFi (z) = Dσ Fi (z) D x t−1 Fi (z) D t Fi (z) , (3.A.10)
f g
DTi (z) = Dσ Ti (z) D x t−1 Ti (z) D t Ti (z) . (3.A.11)

The first order expansion of Ti around z̄i = (0, x̄ i, 0) in terms of the total derivative decomposition

is

(3.A.12)

Ti ( z̄i ) + DTi z̄i )(z − z̄i ) = [Ti ( z̄i ) + Dσ ( z̄i )σ + D x t−1 Ti ( z̄i )](x t−1 − x̄ i ) + D t ( z̄i ) t .

Interchange derivatives and expectations to obtain

(3.A.13)

0 = DFi (z)

= Et DFi (z)
 X X 
= Et  qi, j (z)D f i (vi, j (z))Dvi,. j (z) − .
* pi, j (z) f i (vi, j (z)) − − f i (vi,i (z)), 0, 0/
+
 j=1,2 , j=1,2 -

110
where the total derivatives of vi, j and ui, j are

 DT j (ui, j (z)Dui, j (z))

 

DTi (z)
 
Dvi, j (z) =   , (3.A.14)
 0 Inx 0 
  
 

 0 0 In 
  
 
   
 0   1 0 0 

Dui, j (z) =  DTi (z)  +  x̄ i − x̄ j 0 0  . (3.A.15)
   
 0    t+1 0 0 

Combining all these terms to have

X 
D x t−1 Fi (z) = qi, j (z) D x t+1 f i (vi, j (z))D x t−1 T j (ui, j (z))D x t−1 Ti (z) (3.A.16)
j=1,2

+D x t f i (vi, j (z))D x t−1 Ti (z) + D x t−1 f i (vi, j (z)) ,
X 
D t Fi (z) = qi, j (z) D x t+1 f i (vi, j (z))D x t−1 T j (ui, j (z))D t Ti (z) (3.A.17)
j=1,2

+D x t f i (vi, j (z))D t Ti (z) + D t f i (vi, j (z)) ,
X 
Dσ Fi (z) = qi, j (z) D x t+1 f i (vi, j (z)) D x t−1 T j (ui, j (z)) (3.A.18)
j=1,2
(Dσ Ti (z) + x̄ i − x̄ j ) + Dσ T j (ui, j (z)) + D t Ti (ui, j (z)) t+1

 X
+D x t f i (vi, j (z))Dσ Ti (z) + pi, j (z) f i (vi, j (z)) − pi,i (z) f i (vi,i (z)).
j=1,2

111
Evaluating (3.A.16), (3.A.17) and (3.A.18) at z = z̄i , we have

0 = D x t−1 Fi ( z̄i ) = D x t+1 f i (v̄i,i )D x t−1 Ti ( z̄i )D x t−1 Ti ( z̄i ) (3.A.19)

+D x t f i (v̄i,i )D x t−1 Ti ( z̄i ) + D x t−1 f i (v̄i,i ),

0 = D t Fi ( z̄i ) = D x t+1 f i (v̄i,i )D x t−1 Ti ( z̄i )D t Ti ( z̄i ) (3.A.20)

+D x t f i (v̄i,i )D t Ti ( z̄i ) + D t f i (v̄i,i ),

0 = Dσ Fi ( z̄i ) = D x t+1 f i (v̄i,i ) D x t−1 Ti ( z̄i )Dσ Ti ( z̄i ) + Dσ Ti ( z̄i )

(3.A.21)
X
+D x t f i (v̄i,i )Dσ Ti ( z̄i ) + pi, j ( z̄i ) f i (v̄i, j ).
j=1,2

Note (3.A.19) gives a matrix quadratic in D x t−1 Ti ( z̄i ), and we can find all roots of this equation.

Provided a solution for D x t−1 Ti ( z̄i ), we can further solve for D t Ti ( z̄i ) and Dσ Ti ( z̄i ) from (3.A.20)

and (3.A.21), respectively. It is important to note that a major difference in this perturbation

method and others is equations (3.A.19) - (3.A.21) can be solved for each i independently. It is

also important to note that evaluating at z̄i suppresses the feedback effect in the approximate policy

functions since it assumes the shock term to be zero. The technique can be extended to higher

orders which will allow the feedback effect to show in the approximate policy functions, but we

suppress the introduction because it is not directly related to our application.

112
Appendix 3.B Estimation, Optimization and Filtering Method

3.B.1 Bayesian Estimation

We perform Bayesian estimation in two steps. The first step is often referred to as the quasi-Bayesian

estimation, in which we resort to numerical optimization and find

θ̂ = arg max log p(y1:T ; θ) + log q(θ) ,

 
(3.B.1)
θ∈Θ

with q(θ) the prior distribution, and p(y1:T ; θ) the likelihood of θ. θ̂ is the posterior mode. The

optimization above entails the evaluation of likelihood function. To this end, we take the following

steps:

1. For each θ, solve Xt = Ti (Xt−1,  t ; θ).

2. Stack observation equations, regime transitions and solutions to form a state space represen-

tation.

3. Apply Chang, Maih and Tan (2018) filter to obtain approximated p(y1:T ; θ).

We implement a standard Random-Walk Metropolis-Hasting sampler in the second step to draw

samples from the posterior distribution p(θ|Y1:T ). To this end, we draw a chain of {θ i } following

steps:

0. Use θ̂ as θ 1 .

1. Given θ i−1, p(Y |θ i−1 ) and q(θ i−1 ), draw ϑ = θ i−1 + η with η ∼ N(0, c2 Σ) and c a scale tuning

parameter.
 
p(ϑ|Y )
2. Let θi = ϑ with probability α = min p(θ i−1 |Y )
,1 , and θ i = θ i−1 otherwise.

For simplicity, we take Σ to be identity matrix.

113
3.B.2 Numerical Optimization

The estimation procedure entails numerical optimization over the posterior surface

p(θ|Y1:T ) ∝ p(Y1:T |θ)p(θ) (3.B.2)

where p(Y1:T |θ) is the likelihood function and p(θ) is the prior density function. We apply a DIRECT

search method with Matlab implementation“Pattern Search", among a forest of alternative methods

for its global convergence and superior numerical stability in our application. Alternative methods

can generally be dichotomized into local and global methods, as well as derivative-based and

derivative-free methods.

Pattern Search is a derivative-free global search method. At iteration i, it evaluates the objective

function on grid points round θi with distance (meshsize) di = d and moves to the point with the

highest evaluation in the next iteration, and set di+1 = 2d. If the it does not find a higher function

value, it generates a new set of grid points to evaluate by reducing the distance to di+1 = d/2, and

stop if di+1 <  tolerance that is small. In our application,  toleranbce = 1e − 5.

Pattern search is superior to local methods because it converges globally provided the objective

function has global maximum on a compact support. Its convergence rate is in general slower

than derivative-based methods because it does not take into account the slope information of the

objective function. However, it is numerically more stable because the numerical derivatives are

typically evaluated with non-trivial error.

Pattern search, to our experience, is more numerically more stable compared to other global

methods implemented in Matlab such as particle swarm and simulated annealing because it gener-

114
ates candidate points in a more controlled manner, whereas the alternatives often generate points

rendering the model unsolvable.

To guarantee we find a global maximum, we further modify the standard Matlab implementation

of pattern search to allow random meshsize d each time we optimize on the posterior surface, and

repeat this process 100 times before reporting the posterior mode estimates.

3.B.3 Chang, Maih and Tan Filter

This filter evaluates the approximate likelihood and states using “marginalization-collapsing" pro-
T
cedure. Note an exact filter requires to track the complete history of {st }t=1 ∈ {1, 2}T , hence is

costly to compute. The filter assumes a state space model (SSM) of form

yt = Dst + Z st x t + Fst zt + Q st ut (3.B.3)

x t = Cst + G st x t−1 + Est zt + Rst  t (3.B.4)

with st specified by

wt = αwt−1 + νt (3.B.5)

st = 1 + 1{wt ≥ τ} (3.B.6)

115
 allowing correlation between νt and  t−1 with vector of correlation coefficients ρ. Letting dt =  t ,

there is an equivalent SSM

 x
t +

yt = Dst + Fst zt + Z st 0 *. / +Q st ut (3.B.7)
| {z } | {z } d
D̃st , t -
Z̃st | {z }
ζt

*. x t +/ = *. Cst + Est zt +/ + *. G st 0 +/ *. x t−1 +/ + *. Rst +/  (3.B.8)

t
, dt - , 0 - , 0 0 -, d t−1 - , I -
| {z } | {z } | {z } | {z } | {z }
ζt C̃st G̃ st ζt−1 R̃st

Define the following notations

i, j
pt|t−1 = P(st = j, st−1 = i|Y1:t−1 )

i, j
pt|t = P(st = j, st−1 = i|Y1:t )

j
pt|t = P(st = j |Y1:t )

j
Xt|t = E(Xt |st = j, Y1:t )

j
Px,t|t = var(Xt |st = j, Y1:t ).

The filter is built on predict-update recursion with collapsing steps.

√
i , Pi ) using invariant distribution under regime i. Set p1 = Φ(τ 1 − α 2 ) and
Step 0. Initialize (ζ0|0 0|0 0|0

1 = 1 − p0 . (Note w ∼ N (0, 1/(1 − α 2 )))

p0|0 0|0 0

i
Step 1. Given inputs (ζt−1|t−1 i
, Pt−1|t−1 i
, pt−1|t−1 )i=1,2 ,

116
a. Forecast

(i, j) i
ζt|t−1 = C̃ j + G̃ j ζt−1|t−1

(i, j) i
Pζ,t|t−1 = G̃ j Pζ,t|t G̃0j + R̃ j R̃0j
Z ∞
(i, j)
pt|t−1 = P(st = j, st−1 = i| ρ0  t−1, Y1:t−1 )p( ρ0  t−1 |Y1:t−1 )dρ0  t−1
| −∞ {z }
with a trivariate normal CDF representation

Note by construction

P(st = 0, st−1 = 0| ρ0  t−1, Y1:t−1 ) = P(st = 0|st−1 = 0, ρ0  t−1 ) pt−1|t−1

0
| {z }
with exact representation

Approximate

p( ρ0  t−1 |Y1:t−1 ) ≈ N( ρ0  t−1 ; ρ0 ζ d,t−1|t−1

0
, ρ0 P0 ρ)
| {z d,t−1|t−1}
d section of the inputs

Taken together, we have for i = 1, 2

i
pt−1|t−1 Z ∞ Z (−1)i √ τ Z (−1)i τ√1−α2
(i,0) 1−ρ 0 ρ
pt|t−1 = √ p(x, y, z; µi, Σi )dxdydz
Φ(τ 1 − α 2 ) −∞ −∞ −∞

117
 where p(x, y, z; µ, Σ) is the pdf of a trivariate normal distribution and

 
 0 
 ρ0 ζ d,t−1|t−1
i
µi =  √ 0  ,

 1−ρ ρ
 ρ0 ζ i 
 d,t−1|t−1 
(−1) i α
√
 
1 0
 (1−ρ0 ρ)(1−α 2 ) 
 (−1) i α α2 ρ0 Pd,t−1|t−1
i ρ ρ Pd,t−1 |t−1 ρ 
0 i
Σi =  √ 1+ (1−ρ0 ρ)(1−α 2 )
+ 1−ρ0 ρ
√ 0  .
1−ρ ρ
 (1−ρ ρ)(1−α )
0 2
ρ0 Pd,t−1
i ρ

 0 √ |t−1 i
ρ Pd,t−1|t−1 ρ. 
0
 1−ρ0 ρ 

(i,1) i (i,0)
Note pt|t−1 = pt−1|t−1 − pt|t−1 .
b. Forecast

(i, j) (i, j)
yt|t−1 = D̃ j + Z̃ j ζt|t−1

(i, j) (i, j)
Py,t|t−1 = Z̃ j Pζ,t|t−1 Z̃ 0j + Q j Q0j

Evaluate conditional density

(i, j) (i, j) (i, j)

X
p(yt |Y1:t−1 ) = p(yt |yt|t−1, Py,t|t−1 ) pt|t−1
i, j | {z }
normal dist.

c. Update

(i, j) (i, j) (i, j)

(i, j)
p(yt |yt|t−1, Py,t|t−1 )pt|t−1 j
X
(i, j)
pt|t = , pt|t = pt|t
p(yt |Y1:t−1 ) i
(i, j) (i, j) (i, j) (i, j) (i, j)
ζt|t = ζt|t−1 + Pζ,t|t−1 Z̃ 0j (Py,t|t−1 ) −1 (yt − yt|t−1 )

(i, j) (i, j) (i, j) (i, j) (i, j)

Pζ,t|t = Pζ,t|t−1 − Pζ,t|t−1 Z̃ 0j (Py,t|t−1 ) −1 Z̃ j Pζ,t|t−1

118
 Collapse

(i, j) (i, j) (i, j) (i, j) ( j) (i, j) ( j) (i, j)

j
X pt|t ζt|t j
X pt|t [Pt|t + (ζt|t − ζt|t )(ζt|t − ζt|t )0]
ζt|t = j
, Pζ,t|t = j
i pt|t i pt|t

Step 2. Aggregate

T
Y
p(Y1:T ) = p(yt |Y1:T ).
1

119
Appendix 3.C Additions Figures and Tables

Figure 3.C.1: Filtered Idiosyncratic Uncertainty: CMR against Pmode

Reported -0.3
Estimated
-0.8

-0.9
-0.4

-1

-1.1 -0.5

-1.2

-1.3 -0.6

-1.4

-0.7
-1.5

-1.6
-0.8
-1.7

1985 1990 1995 2000 2005 2010

Note: The red stared curve (right scale, labeled as “Estimated") plots the filtered uncertainty in
natural log produced at the posterior mode of the estimated quantitative model without regime
switching. The blue curve (left scale, labeled as “Reported") plots the same filtered series at the
posterior mode of a the larger model reported in CMR. In both cases, the data spans from 1981:Q1
to 2010:Q2. The first 16 observations are considered pre-sample, and excluded from the likelihood
calculation.

120
 Figure 3.C.2: Normal vs. Large Uncertainty Shock in Fixed Regime, Deviation

-0.1
I

-0.2
0 5 10 15 20 25 30 35 40
10 -3
0
Y

-10

-20
0 5 10 15 20 25 30 35 40

0
C

-0.05

-0.1
0 5 10 15 20 25 30 35 40

0
-0.02
H

-0.04 1 5
0 5 10 15 20 25 30 35 40

Note: The figure reports the generalized impulse response functions with respects to a one standard
deviation uncertainty shock under high-uncertainty steady state. The solid curve presents the
benchmark low-friction case, and the dashed curve presents the high-friction case.

121
 Figure 3.C.3: Low vs. High Steady State Uncertainty in Fixed Regime, Deviation

0
-0.01
I

-0.02
-0.03

0 5 10 15 20 25 30 35 40
10 -3
0
-1
Y

-2
F( ) = 0.1 F( ) = 0.2
-3
0 5 10 15 20 25 30 35 40

0
C

-0.01

-0.02
0 5 10 15 20 25 30 35 40
10 -3
0
H

-5

-10
0 5 10 15 20 25 30 35 40

Note: The figure reports the generalized impulse response functions with respects to a one standard
deviation uncertainty shock under high-uncertainty steady state. The solid curve presents the
benchmark low-friction case, and the dashed curve presents the high-friction case.

122
 Figure 3.C.4: Low vs. High Monitoring Cost in Fixed Regime, Deviation

0
I

-0.02

-0.04
0 5 10 15 20 25 30 35 40
10 -3
0
-1
Y

-2
= 0.21 = 0.27
-3
0 5 10 15 20 25 30 35 40

0
C

-0.01

-0.02
0 5 10 15 20 25 30 35 40
10 -3
0
H

-5

-10
0 5 10 15 20 25 30 35 40

Note: The figure reports the generalized impulse response functions with respects to a one standard
deviation uncertainty shock under high-uncertainty steady state. The solid curve presents the
benchmark low-friction case, and the dashed curve presents the high-friction case.

123
Figure 3.C.5: Credit Spread Against Idiosyncratic Uncertainty, Full Sample Exogenous Switching

3.5

3
3

2.5
2

1
1.5

1
0
0.5

0 -1

-0.5
-2
-1

-3
1985 1990 1995 2000 2005 2010

Spread FS Uncert. RS Uncert.

Note: The dashed dark curve and the diamond red curve (right scale) plot the standardized smoothed
uncertainty in natural log produced at the posterior mode of the estimated quantitative model without
and with regime switching financial friction, respectively. The data spans from 1981q1 to 2010q2.
The blue curve (left scale) plots the credit spread in natural log defined by the spread of BAA
rating corporate bonds over the U.S. 10-year treasury bill in constant maturity. The credit spread
is an observable in the dataset. The plot differentiate left and right scale because the spread data is
heavily skewed toward right, while the estimated uncertainty is relatively symmetric. The shaded
areas mark the NBER recessions.

124
 Table 3.C.1: CMR vs. Fixed Regime

Prior Posterior mode

Parameter Label Dist. Mean SD CMR 1-Regime
b consumption habit beta 0.5 0.1 0.7358 0.7746
F (ω̄) prob. of default beta 0.007 0.0037 0.0056 0.0145
µ monitoring cost beta 0.275 0.15 0.2149 0.1838
σa curvature, utilization cost normal 1 1 2.5356 1.8454
S00 curvature, investment adjustment cost normal 5 3 10.7800 12.0885
απ MP weight on inflation normal 1.5 0.25 2.3965 1.0818
α∆y MP weight on output growth normal 0.25 0.1 0.3694 0.3620
ρp MP smoothing beta 0.75 0. 1 0.8503 0.8481
ξp price rigidity beta 0.5 0.1 0.7412 0.7981
ι price index beta 0.5 0.15 0.8974 0.8710
ξw wage rigidity beta 0.75 0.1 0.8128 0.8243
ιw wage index weight on inf. target beta 0.5 0.15 0.4891 0.4862
ιµ wage index weight on persist tech. growth beta 0.5 0.15 0.9366 0.9333
σe,λ f stddev price markup invg2 0.002 0.0033 0.0110 0.0116

125
σe,µΥ stddev investment price invg2 0.002 0.0033 0.0034 0.0040
σe,g stddev gov. spend. invg2 0.002 0.0033 0.0228 0.0253
σe,µz stddev persistent tech growth invg2 0.002 0.0033 0.0072 0.0073
σe,γ stddev equity invg2 0.002 0.0033 0.0081 0.0039
σe, stddev transitory tech invg2 0.002 0.0033 0.0446 0.0047
σe,p stddev MP invg2 0.002 0.0033 0.4893 0.5049
σe,ζc stddev consumption pref. invg2 0.002 0.0033 0.0233 0.0259
σe,ζi stddev marginal efficiency invest. invg2 0.002 0.0033 0.0550 0.0209
stddev net worth Weibull 0.01 5 0.1746 0.0856
σe,σ stddev unanticipated uncertainty invg2 0.002 0.0033 0.0700 0.0369
ρλ f AR price markup beta 0.5 0.2 0.9706 0.9959
ρ µΥ AR price of investment good beta 0.5 0.2 0.9870 0.9928
ρg AR gov. spend. beta 0.5 0.2 0.9427 0.9111
ρ µz AR perisistent tech growth beta 0.5 0.2 0.1460 0.1035
ρ AR transitory tech growth beta 0.5 0.2 0.8089 0.9928
ρσ AR uncertainty beta 0.5 0.2 0.9706 0.8977
ρζc AR preference beta 0.5 0.2 0.8968 0.9830
ρ ζi AR marginal efficiency of investment beta 0.5 0.2 0.9087 0.4051
 Table 3.C.2: Calibrated vs. Estimated Adjustment Costs

Prior Posterior Mode

Parameters Dist Mean SD 1990-2005* 1990-2005 1990-2010* 1990-2010
b beta 0.7 0.1 0.9910 0.9910 0.9584 0.8583
σa gamma 1 1 1.8454 19.6893 1.8454 11.1925
ξp beta 0.8 0.1 0.8664 0.8661 0.8630 0.8627
αp normal 2.5 0.25 2.2452 3.2297 3.0903 4.0903
ρp beta 0.75 0.1 0.9242 0.9400 0.9410 0.9334
ι beta 0.5 0.15 0.9960 0.9960 0.9960 0.9960
ιw beat 0.5 0.15 0.9662 0.9506 0.9154 0.7658
ιµ beta 0.5 0.15 0.2016 0.1946 0.1244 0.1541
α∆y normal 0.25 0.1 -0.0961 0.3328 0.4850 0.4052
ρλ f beta 0.9 0.2 0.9989 0.9983 0.9977 0.9963
ρΥ beta 0.9 0.2 0.9968 0.9959 0.7585 0.7626

126
ρg beta 0.9 0.2 1.0000 1.0000 1.0000 1.0000
ρ µ∗ beta 0.1 0.2 0.0000 0.0000 0.0002 0.0438
ρε beta 0.9 0.2 0.6222 0.6786 0.6844 0.6322
ρζc beta 0.9 0.2 0.1472 0.2677 0.8004 0.8591
ρ ζi beta 0.9 0.2 0.0012 0.0012 0.1289 0.0000
σe,σ invg 0.05 0.04 0.8051 0.4324 0.8332 0.2613

Note: The posterior modes highlighted by “*" correspond to estimations in which investment adjustment cost parameter S00 and
capital utilization parameter σa are calibrated at 12.0885 and 1.8454, respectively. Both values are posterior modes from the
fixed regime model estimated on sample spanning from 1985:Q1 to 2010:Q2. These constraints are relaxed to a set of priors in
the estimations corresponding to columns without “*".
 Table 3.C.3: Calibrated vs. Estimated Adjustment Costs, Cont’d

Prior Posterior Mode

Dist. Q0.05 Q0.95 1990-2005* 1990-2005 1990-2010* 1990-2010
ξw beta 0.7 0.9 0.5414 0.6004 0.5435 0.8906
S00 normal 5 15 12.0885 11.5569 12.0885 2.8281
σe,λ f invg 0.0005 0.0015 0.0195 0.0196 0.0195 0.0195
σe,Υ invg 0.002 0.006 0.0031 0.0031 0.0070 0.0070
σe,g invg 0.001 0.0033 0.0591 0.0530 0.0601 0.0603
σe,µ∗ invg 0.003 0.001 0.0571 0.0474 0.0734 0.0909
σe,γ invg 0.003 0.001 0.0339 0.0339 0.0318 0.0296
σe,ε invg 0.003 0.001 0.0598 0.0403 0.0813 0.0688
σe,p invg 0.01 1 0.6124 0.5819 0.8836 0.4616
σe,ζc invg 0.003 0.001 0.1892 0.1892 0.1532 0.0891
σe,ζi invg 0.003 0.001 0.0555 0.0556 0.0336 0.1299

127
T2|1 beta 0.001 0.1 0.0608 0.0648 0.3017 0.0720
T1|2 beta 0.001 0.5 0.2319 0.2315 0.0029 0.7920
F (ω̄)1 beta 0.003 0.01 0.0030 0.0029 0.0030 0.0030
F (ω̄)2 beta 0.01 0.02 0.0032 0.0032 0.0032 0.0031
µ1 beta 0.2 0.36 0.0994 0.0996 0.0996 0.1000
µ2 beta 0.2 0.36 0.1188 0.1184 0.1187 0.1149

Note: The posterior modes highlighted by “*" correspond to estimations in which investment adjustment cost parameter S00 and
capital utilization parameter σa are calibrated at 12.0885 and 1.8454, respectively. Both values are posterior modes from the
fixed regime model estimated on sample spanning from 1985:Q1 to 2010:Q2. These constraints are relaxed to a set of priors in
the estimations corresponding to columns without “*".
 Figure 3.C.6: High Uncertainty Regime Probability

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
1985 1990 1995 2000 2005 2010

Smoothed High Risk Prob. NBER recession

Note: The diamond red curve plot the smoothed high uncertainty regime probabilities at the
posterior mode of the estimated quantitative model with regime switching financial friction of
time-varying transition. The data spans from 1981q1 to 2010q2. The shaded area mark NBER
recessions.

128
 Figure 3.C.7: Uncertainty, Exogenous vs. Endogenous

-0.6

-0.8

-1

-1.2

-1.4

-1.6

-1.8

1985 1990 1995 2000 2005 2010

Uncert Exo Uncert Endo

129
Bibliography

Aastveit, K. A., G. J. Natvik, and S. Sola (2017). Economic uncertainty and the influence of

monetary policy. Journal of International Money and Finance 76, 50 – 67.

Aiyagari, S. R. and M. Gertler (1985). The backing of government bonds and monetarism. Journal

of Monetary Economics 16(1), 19–44.

Alspach, D. and H. Sorenson (1972). Nonlinear bayesian estimation using gaussian sum approxi-

mations. IEEE Transactions on Automatic Control AC-17(4), 439–448.

Bachmann, R., S. Elstner, and E. R. Sims (2013, April). Uncertainty and economic activity:

Evidence from business survey data. American Economic Journal: Macroeconomics 5(2),

217–49.

Baker, S. R., N. Bloom, and S. J. Davis (2016). Measuring economic policy uncertainty*. The

Quarterly Journal of Economics 131(4), 1593–1636.

Bazzi, M., F. Blasques, S. J. Koopman, and A. Lucas (2014). Time varying transition probabilities

for Markov regime switching models. Tinbergen Institute Discussion Paper TI 2014-072/III.

Bernanke, B. S., M. Gertler, and S. Gilchrist (1999). Chapter 21 the financial accelerator in a

quantitative business cycle framework. Volume 1 of Handbook of Macroeconomics, pp. 1341 –

1393. Elsevier.

130
Bernanke, B. S. and K. N. Kuttner (2005). What explains the stock market’s reaction to federal

reserve policy? The Journal of Finance 60(3), 1221–1257.

Bianchi, F. (2013). Regime switches, agents’ beliefs, and post-world war ii u.s. macroeconomic

dynamics. The Review of Economic Studies 80(2), 463–490.

Bianchi, F. and C. Ilut (2017a). Monetary/fiscal policy mix and agent’s beliefs. Review of Economic

Dynamics 26, 113–139.

Bianchi, F. and C. Ilut (2017b, October). Monetary/Fiscal Policy Mix and Agent’s Beliefs. Review

of Economic Dynamics 26, 113–139.

Bloom, N. (2009). The impact of uncertainty shocks. Econometrica 77(3), 623–685.

Bloom, N. (2014, May). Fluctuations in uncertainty. Journal of Economic Perspectives 28(2),

153–76.

Bloom, N., M. Floetotto, N. Jaimovich, I. Saporta-Eksten, and S. J. Terry (2018). Really uncertain

business cycles. Econometrica 86(3), 1031–1065.

Caldara, D., C. Fuentes-Albero, S. Gilchrist, and E. Zakrajsek (2016). The macroeconomic impact

of financial and uncertainty shocks. European Economic Review 88, 185 – 207. SI: The

Post-Crisis Slump.

Cappé, O., E. Moulines, and T. Rydén (2005). Inference in Hidden Markov Models. Spinger

Science+Business Media, Inc.

Chang, Y., Y. Choi, and J. Y. Park (2017). A new approach to model regime switching. Journal of

Econometrics 196(1), 127 – 143.

131
Chang, Y. and B. Kwak (2017). U. S. monetary-fiscal regime change in the presence of endogeneous

feedback in policy rules. Working Paper.

Chang, Y., J. Maih, and F. Tan (2018). State space models with endogenous regime switching.

Working Paper.

Chib, S. (1996). Calculating posterior distributions and modal estimates in Markov mixture models.

Journal of Econometrics 75, 79–97.

Chib, S. and M. Dueker (2004). Non-Markovian regime switching with endogenous states and

time-varying state strengths. Working Paper 2008-31, Federal Reserve Bank of St. Louis.

Cho, J. S. and H. White (2007). Testing for regime switching. Econometrica 75, 1671–1720.

Christiano, L. J., M. Eichenbaum, and C. L. Evans (2005). Nominal rigidities and the dynamic

effects of a shock to monetary policy. Journal of Political Economy 113(1), 1–45.

Christiano, L. J., R. Motto, and M. Rostagno (2014, January). Risk shocks. American Economic

Review 104(1), 27–65.

Davig, T. and E. M. Leeper (2006). Fluctuating Macro Policies and the Fiscal Theory. NBER

Macroeconomics Annual 2006 21, 247–316.

Davig, T. and E. M. Leeper (2007). Generalizing the Taylor Principle. American Economic

Review 97(3), 607–635.

Diebold, F., J.-H. Lee, and G. Weinbach (1994). Regime switching with time-varying transition

probabilities. In C. Hargreaves (Ed.), Nonstationary Time Series Analysis and Cointegration,

pp. 283–302. Oxford, UK: Oxford University Press.

132
Drechsler, I., A. Savov, and P. Schnabl (2018). A model of monetary policy and risk premia. The

Journal of Finance 73(1), 317–373.

Favero, C. A. and T. Monacelli (2005). Fiscal Policy Rules and Regime (In)Stability: Evidence

from the U.S. IGIER Working Paper (No. 282.).

Garcia, R. (1998). Asymptotic null distribution of the likelihood ratio test in Markov switching

models. International Economic Review 39, 763–788.

Genz, A. (1992). Numerical computation of multivariate normal probabilities. Journal of Compu-

tational and Graphical Statistics 1(2), 141–149.

Gilchrist, S., J. W. Sim, and E. Zakrajsek (2014, April). Uncertainty, financial frictions, and

investment dynamics. Working Paper 20038, National Bureau of Economic Research.

Hamilton, J. (1989a). A new approach to the economic analysis of nonstationary time series and

the business cycle. Econometrica 57, 357–384.

Hamilton, J. (1996). Specification testing in Markov-switching time-series models. Journal of

Econometrics 70, 127–157.

Hamilton, J. D. (1989b). A new approach to the economic analysis of nonstationary time series

and the business cycle. Econometrica 57(357-384).

Hansen, B. E. (1992). The likelihood ratio test under non-standard conditions. Journal of Applied

Econometrics 7, S61–82.

Jurado, K., S. C. Ludvigson, and S. Ng (2015, March). Measuring uncertainty. American Economic

Review 105(3), 1177–1216.

133
Kalliovirta, L., M. Meitz, and P. Saikkonen (2015). A gaussian mixture autoregressive model.

Journal of Time Series Analysis 36, 247–266.

Kang, K. H. (2014). State-space models with endogenous Markov regime switching parameters.

Econometrics Journal 17, 56–82.

Kim, C.-J. (1994). Dynamic linear models with Markov-switching. Journal of Econometrics 60,

1–22.

Kim, C.-J. (2004). Markov-switching models with endogenous explanatory variables. Journal of

Econometrics 122, 127–136.

Kim, C.-J. (2009). Markov-switching models with endogenous explanatory variables II: A two-step

MLE procedure. Journal of Econometrics 148, 46–55.

Kim, C.-J. and C. Nelson (1999). State-Space Models with Regime Switching. Cambridge, MA:

MIT Press.

Kim, C.-J., J. Piger, and R. Startz (2008). Estimation of Markov regime-switching regression

models with endogeneous switching. Journal of Econometrics 143, 263–273.

Krolzig, H.-M. (1997). Markov-Switching Vector Autoregressions: Modelling, Statistical Inference,

and Application to Business Cycle Anlysis. Springer.

Leeper, E. M. (1991, February). Equilibria under ’active’ and ’passive’ monetary and fiscal policies.

Journal of Monetary Economics 27(1), 129–147.

Leeper, E. M. and T. Zha (2003, November). Modest policy interventions. Journal of Monetary

Economics 50(8), 1673–1700.

134
lhuissier, S. and F. Tripier (2016). Do uncertainty shocks always matter for business cycles?

Working paper.

Linde, J., F. Smets, and R. Wouters (2016). Challenges for macro models used at central banks.

Volume 2 of Handbook of Macroeconomics, pp. 2185 – 2262. Elsevier.

Liu, Z., D. F. Waggoner, and T. Zha (2011). Sources of macroeconomic fluctuations: A regime-

switching dsge approach. Quantitative Economics 2(2), 251–301.

Maih, J. and D. F. Waggoner (2018). Perturbation methods for DSGE models with time-varying

coefficients and transition matrices. Mimeograph, Norges Bank.

Meisenzahl, R. R. (2014). Verifying the state of financing constraints: Evidence from u.s. business

credit contracts. Journal of Economic Dynamics and Control 43, 58 – 77. The Role of Financial

Intermediaries in Monetary Policy Transmission.

Reinhart, C. M. and K. S. Rogoff (2008, May). Is the 2007 us sub-prime financial crisis so different?

an international historical comparison. American Economic Review 98(2), 339–44.

Sargent, T. J. and N. Wallace (1981). Some unpleasant monetarist arithmetic. Quarterly Re-

view (Fall), v. 5, no. 3.

Smets, F. and R. Wouters (2007, June). Shocks and frictions in us business cycles: A bayesian dsge

approach. American Economic Review 97(3), 586–606.

Stock, J. H. and M. W. Watson (2012). Disentangling the channels of the 2007-09 recession.

Brookings Papers on Economic Activity 43(1 (Spring)), 81–156.

135
Taylor, J. B. (1993). Discretion versus policy rules in practice. Carnegie-Rochester Conference

Series on Public Policy 39, 195 – 214.

Teicher, H. (1963). Identifiability of Finite Mixtures. The Annals of Mathematical Statistics 34(4),

1265–1269.

Teicher, H. (1967). Identifiability of Mixture of Product Measures. The Annals of Mathematical

Statistics 38, 1300–1302.

Timmermann, A. (2000). Moments of Markov switching models. Journal of Econometrics 96,

75–111.

Wallace, N. (1981). A modigliani-miller theorem for open-market operations. The American

Economic Review 71(3), 267–274.

136
Shi Qiu

Research Econometrics, Macro-econometrics,

Interests Quantitative Macroeconomics, Computation

Education Indiana University, Bloomington, IN, U.S.

Ph.D., Economics, July 2019

Dissertation: Essays on Regime Switching Models with Endogenous

Feedback

Committee: Joon Y. Park, Yoosoon Chang, Grey Gordon, Ke-Li Xu

Syracuse University, Syrcause, NY, U.S.

M.A., Economics, May, 2012

Huazhong University of Science and Technology, Wuhan, Hubei, China

B.A., English Literature and Linguistics, May, 2010

B.A., Financial Management, May, 2010

Employment Fudan University, Shanghai, China

Assistant Professor, School of Economics August, 2019 – present

Publication 1. Gordon, G., and Qiu, S., A Divide and Conquer Algorithm for Exploiting

Policy Function Monotonicity, Quantitative Economics, 9:521-540, 2018

(lead article)
Work in 1. Chang, Y., Qiu, S., Expectation Effects of Switching Financial Frictions,
Progress Working Paper, 2019

2. Chang, Y., Kwak, B., and Qiu, S., U.S. Monetary-Fiscal Regime Changes

in the Presence of Endogenous Feedback in Policy Rules, 2019 (Journal

of Applied Econometrics, Under Revision)

3. Chang, Y., Park, J., and Qiu, S., Regime Switching Models with Multiple

Dynamic Factors, Working Paper, 2018

4. Chang, Y., Park, J., and Qiu, S., Regime Switching Model with Nonlinear

and Self-Excited Latent Dynamic Factor, Working Paper, 2017

Conference Asian Meeting of the Econometric Society, Xiamen, China 2019

Presentation Midwest Econometrics Group Meeting, Madison, WI, U.S. 2018

Intl. Assn. of Applied Econometrics Meeting, Montréal, Canada 2018

North American Summer Meeting of the Econometric Society, Davis, CA,

U.S. 2018

China Meeting of the Econometric Society, Shanghai, China 2018

Midwest Economics Association Meeting, Evanston, IL, U.S. 2018

Midwest Econometrics Group Meeting, College Station, TX, U.S. 2017

Midwest Macroeconomic Meeting, St. Louis, MO, U.S. 2015

Jordan River Conference, Bloomington, IN, U.S. 2015

Teaching Fudan University (2019-2020, Scheduled)

Macro Core I

Time-Series Analysis

Indiana University 2012-2019

Econometrics II: Regression and Time-Series (Graudate, TA)

Introduction to Microeconomics

Statistical Analysis for Business and Economics

Introduction to Applied Econometrics

Introduction to Microeconomics (Undergraduate, TA)

Introduction to Macroeconomics (Undergraduate, TA)

International Trade (Undergraduate, TA)

Introduction to Applied Econometrics (Undergraduate, TA)

Economics of Industry (Undergraduate, TA)

Honors and Alumni Associate Instructor Award for Excellence in Teaching 2018

Awards Daniel J. Duesterberg Award, Indiana University 2017,2018

College of Arts and Sciences Travel Award, Indiana University 2017

Economics Department Travel Award, Indiana University 2015

Graduate Assistantship, Indiana University 2012-2019

Maxwell School Dean’s Award, Syracuse University 2011

Language English (Fluent) Chinese (Native)

 Technical Skill Proficient in Fortran, Matlab, Stata, Mathematica, Linux

Reproduced with permission of copyright owner. Further reproduction prohibited without permission.