Journal Pre-proof

Implied volatility forecast and option trading strategy

Dehong Liu, Yucong Liang, Lili Zhang, Peter Lung, Rizwan Ullah

PII: S1059-0560(20)30251-3
DOI: https://doi.org/10.1016/j.iref.2020.10.023
Reference: REVECO 2083

To appear in: International Review of Economics and Finance

Received Date: 21 March 2020

Revised Date: 18 October 2020
Accepted Date: 25 October 2020

Please cite this article as: Liu D., Liang Y., Zhang L., Lung P. & Ullah R., Implied volatility forecast
and option trading strategy, International Review of Economics and Finance (2020), doi: https://
doi.org/10.1016/j.iref.2020.10.023.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.

© 2020 Published by Elsevier Inc.

 Implied Volatility Forecast and Option Trading Strategy

Dehong Liu
School of Economics and Management
Beijing Jiaotong University
China
[email protected]

Yucong Liang
School of Economics and Management

of
Beijing Jiaotong University
China

ro
[email protected]

-p
re
Lili Zhang*
School of Economics and Management
lP

Beijing Jiaotong University

China
[email protected]
na

Peter Lung
ur

Reiman School of Finance

University of Denver
Jo

United States
[email protected]

Rizwan Ullah
School of Economics and Management
Beijing Jiaotong University
China
[email protected]

* Corresponding author, School of Economics and Management, Beijing Jiaotong University No.3
Shangyuancun Haidian District Beijing 100044 P. R. China
 Implied Volatility Forecast and Option Trading Strategy

Abstract

We examine the implied volatility derived from an improved Artificial Bee

Colony with Back Propagation (BP) neural network model that is Artificial Bee
Colony-Back Propagation (ABC-BP) neural network model. We find that the
improved model can better predict the implied volatility than basic BP neural network
model and Monte Carlo simulation. Nevertheless, the option price derived from the
Monte Carlo simulation is more efficient when we apply the simulation to the option
straddle trading strategy. Additionally, in a robustness test we find that our proposed
neural network model performs better than the traditional GARCH model in building

of
up option trading strategies.

ro
Keywords: ABC-BP neural network; Implied volatility; Options trading strategy
-p
re
lP

JEL classification: C6; D9; G1

na

Highlights
ur

• We use the artificial bee colony algorithm and neural network model to replace the
Jo

traditional volatility prediction model.

• The implied volatility calculated by the new model is used to predict the option price
through Monte Carlo simulation.
• Combined with three options trading strategies, including bottom straddle strategy,
butterfly strategy, and calendar spread strategy, the accuracy of volatility prediction
under the new model is explored.

1
I. Introduction

Literature has well documented that the estimation of implied volatility

(hereafter, IV) is crucial in risk management, derivatives pricing (i.e., Muzzioli, 2010).

When all other option parameters are known, there is a one-to-one relationship

between option prices and the underlying expected asset volatility. This yields the

so-called implied volatility (Giot, 2009). If option markets are informationally

efficient and the option pricing model is correct, implied volatility should subsume

of
ro
the information contained in other variables in explaining future volatility (Jiang and

Tian, 2005). -p
re
Implied volatility is often referred to as the market's volatility forecast and is said
lP

to be forward looking as opposed to historical methods which are by definition

backward looking (Koopman et al. 2005). And studies on IV estimation have three
na

major strands: backwardation derivation according to derivatives prices, time-series

ur

volatility models, and the artificial neural network models. For the backwardation
Jo

method, Ser-Huang and Clive (2005) and Harvey and Whaley (1992) find IV contains

the predictability of realized volatility in the future. Current literature also shows that

model-free implied volatility (hereafter, MFIV) can yield the best forecasting

performance both during normal and extreme market conditions (Markose et al. 2012).

MFIV subsumes information contained in the Black-Scholes (hereafter, BS) and past

realized volatility (Jiang and Tian, 2005; Huang and Zheng, 2009). But there has been

a considerable debate on the use of the BS implied volatility as an unbiased forecast

for future volatility of the underlying asset (Jiang and Tian, 2005).

1
 For time-series volatility models, GARCH model is one of the most popular

approaches to predict volatility. Zhang et al. (2009), for example, use GARCH family

models to forecast the volatility of crude oil. Engle and Sokalska (2012) use similar

time-series model to predict intraday volatility in the US equity market and Fang et al.

(2017) apply GARCH to analyze the volatility in gold futures market.

According to the artificial neural network models, Andreou et al. (2006) show

that neural network model with Huber function outperforms optimized least square

of
ro
models. Sermpinis et al. (2013) suggest that higher-order neural networks (HONNs)

-p
show a good performance in forecasting the future realized volatility of the FTSE 100
re
futures index. Jiang (2002) proposes a parsimonious GMM（Generalized method of
lP

moments） estimation for continuous time option pricing models with stochastic

volatility, random jump and stochastic interest rate. Bollen and Whaley (2004)
na

examined the relation between net buying pressure and the shape of the implied
ur

volatility function (hereafter, IVF) for index and individual stock options and verified
Jo

that implied volatility of stock options is dominated by call option demand. More

recently, Xu (2017) analyzes the key determinants of S&P 500 volatility at both

monthly and daily frequencies. Therefore, there are many applications of neural

models to solve financial problems, but there are few papers to solve the volatility

problem.

In sum, IV estimation approaches have been widely applied to many option

trading strategies for both speculation and risk management. However, current

literature has not examined the performance of the Artificial Bee Colony-Back

2
Propagation (hereafter, ABC-BP) neural network model in IV estimation. The model

takes a part of the training value in the data range as the experimental object and uses

another part of the measured value as the test object. In the current paper, we study

the performance of this relatively new model in IV predictability. Our results suggest

that ABC-BP neural network model can more efficiently and effectively measure IV.

More importantly, the model is applicable to various option trading strategies.

of
The structure of this paper is as follows: Part II introduces the BP neural

ro
network model, ABC-BP neural network model, Monte Carlo simulation and

-p
traditional volatility calculation models, that is, BS and GARCH models. Part III
re
outlines the data source and data error. The prediction results of implied volatility
lP

under the ABC-BP neural network model are also processed in Part III, and the option
na

strategy is introduced. And finally, Part IV concludes the paper.

ur

II. Models
Jo

Ⅱ.1 Back Propagation (BP) neural network model

The back propagation (hereafter, BP) neural network model is a multilayer

feed-forward network trained according to the error backpropagation algorithm and is

one of the most widely applied neural network models (Li et al. 2012). BP neural

network model is a three-layer feed-forward network composed of the input layer, the

hidden layer, and the output layer. Each layer contains several disconnected neurons

nodes, and the adjacent nodes are connected according to certain weight values. The

direction of the transfer of information is from the input layer to the hidden layer to

the output layer. If the difference between the actual output and the expected output

3
cannot meet the required error, the error value will be fed back layer by layer along

the network path, and the connection weights and thresholds of each layer will be

corrected.

In BP type models, the input vector is ( , … , , … , ), and the output vector

is ( ,…, ,…, ). The weight of the input node and the hidden node is ,

and the weight of the output hidden node and the output node is , if the given

= {( ( ), ( )}, where ( ), ( ) are the

of
training data is input value and the

ro
expected output value in the training set respectively, and the entire training is

-p
continuously adjusting the parameters of the neural network. The error is gradually
re
reduced until the required conditions are satisfied. According to the principle of
lP

supervised learning with neural network, the training error can be described as
na

follows:

"
( ) = ∑$# ∑ # ( )− ( ( )
| )!
ur

(1)

% and & respectively represent the number of data in the neural network training set
Jo

and the number of neurons in the output layer. After training the neural network,

arbitrarily input a data set ($' ) to obtain the corresponding predicted

data ( $' | ) .The number of hidden units is calculated according to the following

formula:

() * = + + -( ) * + (.* * , (2)

+ is a constant between 1 and 10.

Ⅱ.2 Artificial Bee Colony Algorithm Optimization of BP Neural Network Model

Ⅱ.2.1 Principle of ABC algorithm

4
 The Artificial Bee Colony (hereafter, ABC) algorithm was proposed in 2005

by Dervis Karaboga, inspired by the bees' intelligent behavior. As an optimization

algorithm, it provides a population-based search process. Madeleine and Bin (2008)

revealed that bees rely on waggle dance to convey information about the source of

food. In the ABC algorithm, the position of the food represents a possible solution in

the optimization problem, and the amount of nectar represents the quality or fitness of

the corresponding solution.

of
First, the algorithm randomly generates the initial population, %/ solutions

ro
( X 1 , …, X N ) , it then sets the limit (0 1 2) and the maximum cycles number (34%).
-p
re
After initialization, the bees begin a cyclic search: employed bees carry out
lP

neighborhood search of old solutions using the greedy mechanism. If the fitness of the

new solution is greater than the fitness of the old solution, the bees will forget the old
na

solution and remember the new solution. And then calculate the possible values (5 )
ur

of these new solutions, and onlookers start to search for new solutions near this
Jo

possible value (5 ) and the solutions remembered by employed bees using the

greedy mechanism. If the latest solution obtained can no longer be updated (the

number of updates exceeded the limit), then the solutions will be abandoned by the

scouts and replaced with a new solution. This cycle is up to the maximum cycle

number. There is only one scout per cycle.

Ⅱ.2.2 BP neural network model based on ABC algorithm (ABC-BP)

In this paper, we integrate the ABC algorithm with BP neural network model.

We use this algorithm to find the optimal network weight and threshold. This paper

5
also use this algorithm which combines the generalization performance of neural

network, the global iteration, with local search ability of ABC algorithm.

We present the calculation method of the ABC-BP neural network model in

the appendix.

Ⅱ.3 Monte Carlo simulation

The principle of Monte Carlo simulation option pricing is: according to the

of
given price movement process of the underlying asset, simulate the price change of

ro
the underlying asset, and when the simulation times reach a certain number, obtain the
-p
expectation by taking the mean value. According to the law of large Numbers, the
re
Monte Carlo simulation results finally satisfy the convergence. Compared with BS
lP

model, this method has the following advantages: Firstly, it is flexible, easy to
na

implement and improve; secondly, the error and convergence speed of the simulation
ur

estimate are independent of the dimension of the problem solved, which can solve the
Jo

multi-asset option pricing and path dependent option pricing problems well. However,

in order to achieve high accuracy, thousands of simulations are generally required, so

the Monte Carlo method usually takes much more time than the BS model.

For comparison, this paper also examines the Monte Carlo method of option

pricing whose theoretical basis is risk neutral pricing: under the risk neutral measure,

the option price can be expressed as the expected value of the discount for its maturity

return, that is:

5= "
678(9:;) <(= , = , ⋯ , =; )!, (3)

6
where " represents the risk neutral expectation, ? is the risk-free rate, and is

the expiration date, <(= , = , ⋯ , =; ) is the expected return on the underlying asset.

Calculating the option price is to calculate an expected value, and the Monte

Carlo simulation method is used to estimate the expected value, that is Monte Carlo

simulation method of option pricing. The Monte Carlo simulation method for option

pricing includes the following four steps:

of
a) Divide the time interval into ( sub-intervals, and the discrete form of calculating

ro
the underlying asset price are:

@:9
A
D(
-p
EFA 9 E )'G - EFA 9 E HE
= (2 ' )6 BCB , I ~%(0,1)
re
(4)

= (2 ' ), = (2 ) represent the price of underlying asset at the time of 2 ' and
lP

2 receptively.
na

b) Calculate the return to maturity of option under this path and the discount of the
ur

return using the risk-free rate:

Jo

4 = 678(9:;) max{0, =P − Q} (5)

c) Repeat the first two steps to obtain a large number of samples of the discounted

value of option returns.

d) Calculate the mean of samples, and get the values of the option price using Monte

Carlo simulation:

_
UV (WXY) ∑`
_aA Z[\{],^Y 9P}
4RS = T
678(9:;) ∑T# 4 = (6)
T

Where Ccd is one of the values calculated by Monte Carlo simulation. This

7
paper carries out Monte Carlo simulation of the predicted implied volatility and

calculates the logarithmic rate of return of the S&P500 index from January 2017 to

May 2018, and then uses it as the expected rate of return and finally predicts closing

price of the S&P500 option.

Ⅱ.4 Estimation of Implied Volatility based on Traditional Model

Considering many time series models, the family of GARCH models is still

of
the most popular, and the most commonly used in a financial modeling. Jiang and

ro
Tian（2005）suggested that the information content of implied volatility is not

necessarily higher than that of

-p
standard time series models such as
re
GARCH/EGARCH. And Pilbeam and Langeland (2005) indicated that a GARCH (1,1)
lP

model is especially shown to be a good parameterization of the process. Moreover, the

na

complex model is not necessarily better than the simple model, like the simple
ur

GARCH model volatility which is the representative of the time series model. What
Jo

we use in estimating equations is rolling estimate method, namely for forecasting

volatility of week (month), respectively, to select 120 daily returns before the week

(month) to establish a GARCH model to forecast individually, rather than building a

model for all week (month) volatility forecast. And the rolling estimation method can

further improve the model's predictive ability, because the prediction of the variance

is dynamic in this approach. We use GARCH (1,1) model for daily data.:

ℎ ' = f + g × i + jℎ (7)

where ℎ ' , ℎ is the predicted variance at time 2 + 1 and 2 respectively, i is

the predicted residual at time t, and g, j are both random parameters.

8
 It is assumed that the first day to be predicted from the 2 + 1 day, the

predicted residual i and the predicted variance ℎ obtained from the 2 day can

predict the conditional variance ℎ ' of the 2 + 1 day, but there is a problem

when predicting the conditional variance of the second day, that is, starting from the

2 day, the prediction residual of the 2 + 1 day is not known. At this time, the

predicted rate of return on the 2 + 1 day is used to replace the true rate of return on

the 2 + 1 day, thereby obtaining the residual i ' on the 2 + 1 day, and

of
ro
placing it into the conditional variance model.

In the same way, the predicted variance ℎ -p ' of the 2 + 2 day is obtained.
re
By analogy, the variance of the rate of return for the day can be predicted, and this
lP

prediction method is the dynamic forecasting method. It is conceivable that since the

predicted value of the mean equation return is substituted for its true value, each
na

prediction error is added into the subsequent prediction, and the longer is the
ur

prediction period, the worse is the prediction result.

Jo

Ⅲ. Prediction results and Option Trading Strategy

Ⅲ.1 Data and statistics

This study is based on the S&P 500 index options accessed through the

Bloomberg terminal, and selecting the daily trading data from January 8th, 2013 to

May 29th, 2018.

In order to ensure the accuracy and efficiency, the steps taken to filter the data

are as follows:

a) The transaction volume is not 0;

9
 b) The price of the option cannot be lower than $0.05;

c) Select at-the-money option, 0.45 < p602i < 0.55, and the delta ratio is

the percentage change in the option premium for each dollar change in the

underlying, and in-the-money option, p602i < 0.45 , and out-the-money

option, p602i > 0.55. And descriptive statistics of the data are shown in

Table

(Insert Table 1 about here)

of
ro
The volatilities include realized volatility, implied volatility, and VIX

-p
(Volatility Index). This paper selects the VIX that best substitute for implied volatility
re
as the training set which will be used in ABC-BP neural network model (Figure 1).
lP

(Insert Figure 1 about here)

The VIX represents market's expectations for volatility over the coming 30
na

days. It is composed of eight groups of options for the call option and put option that
ur

are closest to the at-the-money option of the S&P 500 index option in the near month
Jo

and the next month, whose implied volatility is worked out respectively, and the VIX

is obtained using weighted average method. It provides market participants with an

indicator that is more reflective of the overall market trend. The correlation among

three kinds of volatilities can be seen in Table 2.

(Insert Table 2 about here)

Ⅲ.2 Out-of-sample testing

We calculated the correlation between historical volatility, implied volatility

and VIX, and found that the correlation coefficient of implied volatility and VIX was

10
higher than that of implied volatility and historical volatility, so the subsequent

volatility fitting was fitted and tested using VIX as the standard. And figure 2

suggests that the model have predictability which can be shown in the comparisons

between predicted and sample true values based on ABC-BP neural network model.

(Insert Figure 2 about here)

Comparing the BP neural network model in figure 3 with the findings in

Figure 2, we find that ABC-BP neural network model is closer to the VIX, which

of
ro
reveals that the volatility prediction under this method is more optimal. In addition,

-p
we compare the volatility predicted by the above two models using the following

indicators, mean square error (3= ), mean absolute error (3r ), mean bias error
re
(3s ). The mean square error（(3= )） of the two models decreases as the number
lP

of iterations increases, and it is shown in Figure 4 more intuitively:

na

(Insert Figure 3 about here)

ur

In Figure 3, we find that MSE decreases with the increase of iterations, and the
Jo

MSE of ABC-BP neural network model is clearly smaller than that of BP neural

network model under an optimum fit. So, the validity of model prediction is

illustrated.

(Insert Figure 4 about here)

To further illustrate the advantages of the ABC-BP neural network model, we

use three predictive indicators of MSE, MAE and MBE in Table 3. The calculation of

these three indicators using formulas (8-10), are listed in the following. Starting with

the mean square error (3= ), such as:

11
 (tE 9uE )B
3= = ∑$# × 100% (8)
$

In this formula, = 1,2, ⋯ %, % indicates the number of predicted data in the

experiment, and denote the input and output data(the predicted value and

the real value) respectively.

The formula for the mean absolute error is as follows:

3r = ∑$# | − & | (9)

$

The mean bias error is expressed in the form of MBE, and its formula is given below:

of
3s = $ ∑$# ( − & )

ro
(10)

Table 3 exhibits the results. In this table, we show mean square error (3= ),
-p
mean absolute error (3r ), mean bias error (3s ) of the BP and ABC-BP neural
re
lP

network model and the calculated results of the BS model from January 2017 to May

2018. According to Table 3, the MSE, MAE and MBE of ABC-BP neural network
na

model is the smallest among the three models, hence the overall performance of the
ur

ABC-BP neural network model proposed in this paper is superior to that of the BP
Jo

neural network model and the BS model.

(Insert Table 3 about here)

Finally, we bring the volatility data into Monte Carlo simulation to calculate

the predicted option price and Greeks (including delta and vega). The predicted results

are shown in Table 4.

(Insert Table 4 about here)

Ⅲ.3 Application of predicted option price in three option strategies

In this section, we apply the findings of the predictability of our model to

12
major options trading strategies: straddle, butterfly, and calendar spread to verify the

validity of model prediction. In addition, we also apply our model to volatility

momentum and mean reversion. The momentum refers to the short-term volatility

behavior while the mean-reversion to long-term behavior. We choose three-day period

for the volatility momentum and use the benchmark of 10% and 90% of volatility

movement for the mean-reversion.

We calculate the trading return as min(S, X) − C − C , where = is the

of
forecasted price of underlying asset using Monte Carlo simulation, { is the strike

ro
price, 4 is the price of the call option using predicted price based on Monte Carlo
-p
simulation, and 4 is put option which also uses predicted price based on Monte
re
lP

Carlo simulation. We also compute Sharpe ratio for these strategies. Note that the

transaction cost is set at $0.14 per contract

na

Table 5 describes the bottom straddle strategy back testing of from January
ur

2017 to May 2018, which is based on the prediction of implied volatility. And the
Jo

results of one and a half years' exploration show that the strategy has a good effect in

2018 when the market volatility changes greatly, and a poor performance in 2017

when the volatility is relatively stable.

(Insert Table 5 about here)

In Figure 5, we find the overall volatility mean value in 2017 including the

first and second half of the year is not as high as that in 2018’s first half of the year.

Therefore, we can infer that the average return of options traded in 2017 is less than

that in 2018. It is reasonable because the larger the volatility of the market, the more it

13
will benefit from the higher volatility. And we will continue to explore.

(Insert Figure 5 about here)

Furthermore, we use a filter to improve the profitability of the strategy. In

table 6, we discuss the following three aspects. First, we classify the data including

in-the-value, at-the-value and out-the-value options. In the data classification, we

consider the prediction accuracy and the time to maturity. The following data are

classified: one is ABC-BP model, the other is GARCH model. Fahlenbrach and

of
ro
Sandås(2010) found evidence that the order flow in volatility strategies with vega

-p
exposure contains information about future realized volatility, but there is no evidence
re
that the order flow in directional strategies with delta exposure contains information
lP

about future returns. So according to the momentum and reversal effects mentioned in

preceding text, the average return rate, sharp ratio, vega value and the significance of
na

the return results are calculated. The calculation results are analyzed in table 6.
ur

For the two models, we consider three conditions in panel A, 0-2%, 2-5%, and
Jo

5-10%. We find that the rate of return of out-the-money options is higher than that of

at-the-money options and in-the-money on the whole, whatever the model is and the

performance of at-the-money options are worst. According to these three conditions,

we find that the strategic return increases and prediction accuracy improves,

particularly in out-the-money category. In combination with the momentum effect and

reversal effect of volatility, we find that they have the same results and the results are

in line with the bottom straddle strategy’s reality. In addition, vega value and the

strategy return change reversely. The larger the return, the smaller the vega value.

14
Moreover, we find that the return of ABC-BP neural network model is generally

higher than that of GARCH model.

(Insert Table 6 about here)

To examine the impact of time-to-maturity, we separate far month’s volatility

from near month’s volatility, and recalculate the return in panel B of Table 6. We

select the months of 1, 2, 3 and 4 respectively. We find that the return decreases with

time-to-maturity, especially out -the-money option. The results also indicate that the

of
ro
return decreases because the value of volatility prediction error is getting bigger,

-p
especially for GARCH model, but not for ABC-BP neural network model. And the
re
performance of out-the-money options are still the best. These findings suggest that
lP

ABC-BP neural network model performs better than GARCH model.

na

(Insert Table 7 about here)

ur

Table 7 presents the test results of butterfly strategies. As we all know, we can
Jo

buy one call option, sell two call (typically at-the-money) options and buy another call

option to build a butterfly strategy. And their exercise prices are { , { , and {|. The

premiums of these three options are 4 , 4 , and 4| which are predicted based on

Monte Carlo simulation. So, the return of butterfly strategies can be divided into four

intervals. We mainly calculate the trading return as min(S, X ({| )) − C − C , where

= is the forecasted price of underlying asset using Monte Carlo simulation.

We find that the closer to the maturity, the higher the return under both

reversal and momentum effect. In this table, Strategy 1 is shorting a call butterfly and

Strategy 2 is shorting a put butterfly. We find that the call butterfly’s performance

15
surpasses that of put butterfly, indicating ABC-BP neural network model works better

for call options.

We examine the time-to-maturity impact by dividing the options into three

maturity categories: 0-182 days, 183-364 days and 365-546 days. The results show

that the performance decreases as getting closer to the maturity, suggesting that the

model works better for short-term butterfly.

In order to consider the impact of volatility changes, we adopt shorting

of
ro
calendar spread strategy. We construct an options spread by simultaneously entering a

-p
long and a short option positions with the same strike price but different maturity. Our

quantitative signal model is { = 2} − } − }| ,where } , } , and }| are the

re
lP

implied volatility for short-term, mid-term, and long-term maturities, respectively.

When { < 0, the calendar spread portfolio is constructed by a long call option for
na

front month and a short call option for back month.

ur

For credit calendar spread strategy in panel A of table 8, we divide the time to
Jo

maturity into 0-30 days, 30-60 days, and 60-90 days of maturity, respectively. We find

the returns gradually increase in maturity. This finding suggests that the ABC-BP

neural network model works well in this strategy.

In Table 8, Panel B, we find the return range of option strategy is also related to

the change of volatility. As volatility rises, the option return increases, and vice versa.

When we classify our sample into four parts according to the change of volatility,

30-15% drop, 0-15% drop, 0-15% rise, and 15-30% rise, respectively. Again, we find

ABC-BP neural network model is more effective. With these impacts from maturity

16
and volatility, we suggest that momentum effect is stronger than reversal effect.

(Insert Table 8 about here)

Ⅴ. Conclusion

In this paper, we examine the artificial bee colony improved model (ABC-BP

neural network model) in IV predictability and apply the model to three popular

option trading strategies. We document two major findings.

of
First, the inherent convergence speed of BP neural network model is slow.

ro
This model is easy to fall into local optimum and easy to be overfit. It also prolongs

-p
the training time while has lower accuracy of implied volatility. Our experimental
re
results show that ABC-BP neural network model performs better than the BP neural
lP

network model in terms of speed and predictability, and also performs better than the
na

traditional GARCH model. Secondly, we document that ABC-BP neural network

ur

model is applicable to option trading strategies such as straddle, butterfly, and

Jo

calendar spreads. The performance of the model is better than traditional GARCH

models. And we believe that this conclusion is helpful for option traders to select

trading strategies and specific trading products.

Acknowledgements

This study was supported by the National Social Science Fund of China

(19BJY242). The authors would like to thank Carl Chen (the Editor) and two

anonymous Referees for their insightful comments.

17
 of
Reference

ro
Andreou, P. C., Charalambous, C. & Martzoukos, S. H. (2006). Robust artificial
neural networks for pricing of European options. Computational Economics, 27(2-3),
329-351.
-p
re
lP

Bollen, N. P. B., & Whaley, R. E. (2004). Does net buying pressure affect the shape of
implied volatility functions? The Journal of Finance, 59(2), 711-753.
na

Buraschi, A., & Jackwerth, J. (2001). The price of a smile: hedging and spanning in
ur

option markets. Review of Financial Studies, 14(2), 495-527.

Jo

Callen, J. L., Kwan, C. C. Y., Yip, P. C. Y., & Yuan, Y. (1996). Neural network
forecasting of quarterly accounting earnings. International Journal of Forecasting,
12(4), 475-482.

Chang, C. C., Hsieh, P. F., & Wang, Y. H. (2010). Information content of options
trading volume for future volatility: evidence from the Taiwan options market.
Journal of Banking & Finance, 34(1), 174-183.

Engle, R. F., & Sokalska, M. E. (2012). Forecasting intraday volatility in the us equity
market. multiplicative component GARCH. Journal of Financial Econometrics, 10(1),
54-83.

Fahlenbrach, R. & Sandås P. (2010). Does information drive trading in option

18
strategies? Journal of Banking & Finance, 34(10), 0-2385.

Fang, L., Chen, B., Yu, H., & Qian, Y. (2017). The importance of global economic
policy uncertainty in predicting gold futures market volatility: a GARCH-MIDAS
approach. Journal of Futures Markets, 38(3), 413-422.

Guo, B., Qian, H., & Hai, L. (2018). Are there gains from using information over the
surface of implied volatilities? Journal of Futures Markets, 38(6), 645-672.

Guo, D. (2000). Dynamic volatility trading strategies in the currency option market.

of
Review of Derivatives Research, 4(2), 133-154.

ro
-p
Harvey, C. R., & Whaley, R. E. (1992). Market volatility prediction and the efficiency
of the S&P 100 index option market. Journal of Financial Economics, 31(1), 43-73.
re
lP

Hong, H., Sung, H. C., & Yang, J. (2018). On profitability of volatility trading on
S&P 500 equity index options: the role of trading frictions. International Review of
Economics & Finance, 55(55), 295-307.
na
ur

Huang, Y. Z.& Zheng, Z. L. (2009). Model-free implied volatility and its information
content: Evidence from Hang Seng index options. Systems Engineering-Theory &
Jo

Practice, 29(11), 46-59.

Jiang, G. J. (2002). Testing option pricing models with stochastic volatility, random
jumps and stochastic interest rates. International Review of Finance, 3(3-4), 233-272.

Jiang, G. J. & Tian, Y. S. (2005). The model-free implied volatility and its information
content. Review of Financial Studies, 18(4), 1305-1342.

Jain, B. A., & Nag, B. N. (2010). Artificial neural network models for pricing initial
public offerings. Decision Sciences, 26(3), 283-302.

Kongsilp, W, & Mateus, C. (2017). Volatility risk and stock return predictability on
19
global financial crises. China Finance Review International, 7(1), 33-66.

Laborda, J., Laborda, R., & Olmo, J. (2014). Optimal currency carry trade strategies.
International Review of Economics & Finance, 33(3), 52-66.

Li, H., Ware, A.; Di, Lan.; Yuan, G., Swishchuk, A.& Yuan, S. (2016). The application
of nonlinear fuzzy parameters PDE method in pricing and hedging European options.
Fuzzy Sets & Systems, 331(331), 14-25.

Li, J., Cheng, J. H., Shi, J. Y., & Huang, F. (2012). Brief Introduction of Back

of
Propagation (BP) Neural Network Algorithm and Its Improvement. Advances in
Computer Science and Information Engineering. Springer Berlin Heidelberg.

ro
-p
Madeleine, B., & Bin, L. J. (2008). Foraging in honeybees—when does it pay to
dance? Behavioral Ecology,19(2), 255-261.
re
lP

Markose, S. M., Peng, Y., & Alentorn, A. (2012). Forecasting extreme volatility of
ftse-100 with model free VFTSE, Carr-Wu and generalized extreme value (GEV)
na

option implied volatility indices. Economics Discussion Papers, 34(4), 871-881.

ur

Muzzioli, S. (2010). Option-based forecasts of volatility: an empirical study in the

Jo

dax-index options market. The European Journal of Finance, 16(6), 561-586.

Sermpinis, G, Laws, J, & Dunis, C. L. (2013). Modelling and trading the realised
volatility of the FTSE100 futures with higher order neural networks. The European
Journal of Finance, 19(3), 165-179.

Ser-Huang, P. & Clive, G. (2005). Practical issues in forecasting volatility. Financial

Analysts Journal, 61(1), 45-56.

SOPHIE X., N. I., Pan, J., & Poteshman, A. M. (2008). Volatility information trading
in the option market. The Journal of Finance, 63(3), 1059-1091.

20
Tissaoui, K. (2019). Forecasting implied volatility risk indexes: International evidence
using Hammerstein-ARX approach. International Review of Financial Analysis,
64(64), 232-249.

Taylor, S. J., Yadav, P. K., & Zhang, Y. (2010). The information content of implied
volatilities and model-free volatility expectations: evidence from options written on
individual stocks. Journal of Banking & Finance, 34(4), 871-881.

Xu, L., & Yin, X. (2017). Exchange traded funds and stock market volatility.
International Review of Finance, 17(4), 525-560.

of
Yang, H., Kutan, A. M.; Ryu, D. (2019). Volatility information trading in the index

ro
options market: An intraday analysis. International Review of Economics and Finance,
64(64), 412-426.
-p
re
Zhang, Y., Yao, T., He, L.& Ripple, R. (2018). Volatility forecasting of crude oil
market: Can the regime switching GARCH model beat the single-regime GARCH
lP

models? International Review of Economics and Finance, 59,302-317.

na

Zhang, H. M.& Watada, J. (2019). An analysis of the arbitrage efficiency of the

Chinese SSE 50ETF options market, International Review of Economics and Finance,
ur

59,474-489.
Jo

21
 of
Table 1 Descriptive Analysis

ro
Maximum Minimum Mean Median Std.Dev
PRICE 208.5900 0.050000 88.92787 91.50000 46.02796
RATE
STRIKE
232.3000
2875.000
-p
7.600000
1450.000
62.06735
2109.478
32.10000
2075.000
62.48160
344.6443
re
TIME 1.936986 0.000000 0.716983 0.720548 0.429604
lP

UNDERLYING PRICE 2872.820 1457.150 2104.769 2071.260 334.6420

VIX 0.321800 0.090100 0.145826 0.137800 0.036855
IV 0.393900 0.048700 0.134003 0.133100 0.026361
na
ur
Jo

22
 of
ro
-p
Figure 1 Volatility Comparison
re
lP
na
ur
Jo

23
 Table 2 Correlation among three volatility
IV Realized Volatility VIX
IV 1 0.5028 0.7220
Realized Volatility 0.5028 1 0.6988
VIX 0.7220 0.6988 1

of
ro
-p
re
lP
na
ur
Jo

24
 of
ro
-p
Figure 2 Comparisons between predicted and sample true values based on ABC-BP
re
neural network model
lP
na
ur
Jo

25
 of
ro
-p
Figure 3 Comparisons based on BP and ABC-BP neural network model optimal and
re
true value
lP
na
ur
Jo

26
 of
ro
-p
re
Figure 4 Comparison of MSE based on BP neural network model and ABC-BP neural
network model
lP

Note: Figure 4 shows that MSE decreases with the increase of iterations, and the MSE
na

of ABC-BP neural network model is obviously smaller than that of BP neural network
model under the optimal fitting condition.
ur
Jo

27
 Table 3 Error comparison under three models

MSE MAE MBE

BS 0.001 0.024 0.001

BP 0.000** 0.010** -0.000**

ABC-BP 0.000*** 0.010*** 0.000***

of
ro
-p
re
lP
na
ur
Jo

28
 Table 4 Descriptive Analysis of option prices in Monte Carlo simulation

Maximum Minimum Mean Median Std.Dev

Underlying price 2872.820 2257.830 2530.086 2648.980 234.174

Call 1611.265 108.0463 1088.706 1334.268 395.1518

Put 1121.573 -110.7903 633.1015 593.3589 292.2569

delta 0.970963 0.536306 0.757982 0.812611 0.126825

of
vega 1234.165 171.0937 864.9403 876.3806 199.771

ro
-p
re
lP
na
ur
Jo

29
 of
Figure 5 2017-2018.5 volatility trend

ro
-p
re
lP
na
ur
Jo

30
 Table 5 Annual performance of options volatility trading strategies

Y Rate of return Sharp Ratio

The first half year of 2017 -4.422% -1.086
The second half year of 2017 -2.082 -0.731
The first half year of 2018 5.022%*** 0.259

Note: *, **, *** indicate significance at the 10%, 5%, 1%.

of
ro
-p
re
lP
na
ur
Jo

31
 Table 6 The impact of forecast accuracy, maturity and execution price matching on

return using straddle strategy

Panel A: The influence of the accuracy of volatility forecast on the return

ITM ATM OTM

Annum
Volatility
Annum Annum
Filter Effect
Return (%) Sharp vega Sharp vega Sharp vega
Forecast
Return (%) Return (%)

of
1
-0.116 -0.711 419.543 -0.852 -5.371 936.970 4.750*** 0.370 255.361

ro
0-2% ABC-BP

2 -0.128 -1.037 390.801 -0.816 -5.886 922.648 0.083** 0.561 267.825

2-5%
ABC-BP
1
-0.070*** -1.992 92.227 -p -0.8481 -8.201 592.329 -0.034 -0.389 286.661
re
2 -0.081*** -0.459 384.305 -0.813 -6.247 914.902 -0.061 -0.589 264.606

GARCH 2 -0.209 -2.770 489.562 -0.893 -17.46 887.755 0.028 0.476 189.461
lP

1
-0.217* -4.002 431.809 -0.8555 -10.63 973.604 0.015** 0.055 130.080
ABC-BP

2 -0.143 -1.557 82.743 -0.820 -3.969 884.684 0.082** 0.764 57.773

na

5-10%

1 -0.442 -0.686 161.418 -0.860 -12.45 969.302 -0.0839 -0.994 210.752

GARCH
2
ur

-0.135 -1.169 403.818 -0.937 -9.751 194.078 -0.054 -0.618 260.417

Panel B: The effect of the time to maturity on the return

Jo

ABC-BP 1 -0.337 -7.793 85.290 -0.933 -21.32 223.099 0.082* 1.411 14.047
0-30d
2 -0.201 -3.649 77.904 -0.920 -16.47 208.995 0.097** 1.369 51.015
GARCH 1 -0.045 -0.549 76.622 -0.925 -10.56 207.341 0.026 0.120 44.684

2 -0.039 -0.366 53.833 -0.930 -5.820 199.676 0.002 -0.126 21.955

ABC-BP 1 -0.215 -2.774 250.907 -0.922 -12.42 382.640 0.017** -0.024 66.676
30-60d
2 0.041* 0.510 116.406 -0.919 -15.99 346.170 0.051** 0.634 79.361
GARCH 1 -0.025 -0.356 166.943 -0.925 -8.852 363.670 0.001 -0.152 97.343

2 -0.067 -0.525 187.829 -0.921 -5.942 373.426 -0.053 -0.497 83.439

ABC-BP 1 -0.321* -5.141 316.449 -0.926 -14.09 480.924 0.066** 0.874 112.489
60-90d
2 0.034** 0.377 98.374 -0.932 -16.36 459.499 0.073** 1.029 120.475

GARCH 1 -0.109 -0.900 264.813 -0.922 -7.107 476.368 -0.050 -0.677 165.845

2 -0.171 -1.096 308.106 -0.915 -5.905 476.339 -0.132 -1.102 180.182

32
 ABC-BP 1 -0.136 -1.146 279.842 -0.917 -6.999 550.288 -0.034 -0.440 147.632

2 -0.172 -1.207 316.049 -0.913 -5.570 557.628 -0.017 -0.276 104.863

90-120d
GARCH 1 -0.052 -1.107 209.567 -0.927 -20.78 581.079 -0.164 -4.006 241.430

2 -0.093 -1.773 241.373 -0.942 -15.35 550.838 -0.052 -1.107 209.567

In this table: In the “effect” column, “1” represents momentum effect, and “2” represents inversion

effect. In Panel A with filters, options are only traded when the predicted price deviation is larger than

the filter value. The filter value means the difference between the predicted option price and the actual

option price. When the value of the filter increases, the number of trades decreases, and the agent is

of
allowed to invest in the risk-free asset on no trading days. This article selects three parts, 0-2%, 2-5%,

ro
5-10% in panel A. And panel B is divided into 1, 2, 3 and 4 months to the maturity. "-" means the data
-p
is missing. Missing data of the whole row has been deleted. And *, **, *** indicate significance at the
re
10%, 5%, 1%.
lP
na
ur
Jo

33
 Table 7 The impact of the distance between the price of the option in the middle and the execution

price on return using short butterfly strategy

Volatility Strategy 1 Strategy 2

Filter Effect
Annum Sharp Annum Sharp
Forecast

1 0.421* 3.662 0.153** -117.776

ABC-BP

2 0.480** 2.508 -0.102* -98.871

0-182d
0.289 4.174 -0.815 -12.870
GARCH 1
-0.502 -9.134 -0.401 -7.355
2

of
0.388* 3.431* 0.666* 12.775
1
ABC-BP

ro
0.419 6.068** 0.427 5.985
2
183-364d

GARCH
1

2
0.158
0.514
-p
1.833
2.198
0.115
0.688
1.272
2.973
re
1 0.291 1.672 0.003* -0.099
ABC-BP
lP

2 0.377* 1.848 - -
365-546d
0.646 4.311 -0.217 -2.792
na

GARCH
1
0.239 3.193 0.997 16.520
2
ur

Note: In the “effect” column, “1” represents momentum effect, and “2” represents inversion effect.
Jo

Strategy 1 is to sell a call option with lower execution price and higher execution price, and then buy

two call options with intermediate execution price; strategy 2 is to sell a put option with lower

execution price and higher execution price, and then buy two call options with intermediate execution

price. We divide the data in "-" indicates that the data is missing. Missing data of the whole row has

been deleted. Where *, **, *** indicate significance at the 10%, 5%, 1%.

34
 Table 8 The influence of time and volatility on the return of calendar spread trading strategy

Strategy 1 Strategy 2

Volatility Annum Annum

Filter Effect
Forecast Return Sharp vega Return Sharp vega

(%) (%)

1 0.279 2.950 48.650

0.000 0.418 56.816
ABC-BP

of
2 -0.300 -2.188 70.488 0.130 0.833 47.233
0-30d
0.000 -0.199 23.693 - - -

ro
1
GARCH
2 0.002 -0.173 45.764 0.259 6.142 31.128

ABC-BP
1 0.002*
-p
-0.162 127.907 0.180* 1.660 109.436
re
30-60d 2 -0.297 -2.122 159.083 0.099 0.557 116.754
lP

GARCH 2 0.001 -0.213 101.164 0.677 16.366 67.346

1 0.003* -0.136 183.577 0.161 1.182 167.494

ABC-BP
na

2 -0.216 -1.480 274.586 0.100 0.542 177.857

60-90d
0.002 -0.185 96.235 - - -
1
ur

GARCH
0.003 -0.170 164.372 0.747 18.247 113.425
2
Jo

The influence of the change range of volatility on the strategic return

ABC-BP 1 0.003 -0.137 137.320 0.299** 4.392 104.779

2 -0.275 -3.074 186.310 0.824*** 8.645 135.281

-30~-15%
1 0.950 18.916 22.513 0.098 19.638 28.716
GARCH
2 0.932 16.437 17.765 0.095 16.903 26.487

ABC-BP 1 0.002 -0.152 91.272 0.390* 6.285 82.303

2 - - - 0.214** 1.675 34.766

-15~-0%
1 0.152 2.024 7.948 0.239 3.746 8.923
GARCH
2 0.094 0.960 22.513 0.090 17.948 24.1232

ABC-BP 1 0.002 -0.164 110.092 0.247*** 3.355 121.915

0~+15% 2 - - - 0.189*** 1.427 75.364

GARCH 1 0.037 -0.463 0.100 0.000 -1.000 14.265

35
 2 0.992 6.702 71.770 -0.097 -8.597 -1.025

ABC-BP 1 0.002 -0.167 142.171 0.252*** 4.298 144.056

+15-30% 2 -0.243 -3.788 249.808 0.357 4.742 151.759

GARCH 1 0.992 6.702 71.770 -0.098 -8.599 -1.041

Note: In the “effect” column, “1” represents momentum effect, and “2” represents inversion effect.

Strategy 1 is the calendar spread strategy of reverse call option and strategy 2 is the calendar

spread strategy of reverse put option. So the buy calendar spread is equal to the long implied

of
volatility, while the sell is the opposite. In table 8, the filter of panel a is the distance between near

ro
and far months, mainly divided into 0-30 days, 30-60 days and 60-90 days. "-" indicates that the

-p
data is missing. Missing data of the whole row has been deleted. Where *, **, *** indicate
re
significance at the 10%, 5%, 1%.
lP
na
ur
Jo

36
 of
ro
-p
re
Figure 6 Comparison of volatility based on different four models (Robustness Test)
Note: The BP, ABC-BP neural network model here depicts the best and worst average
lP

level.
na
ur
Jo

37
 Table 9 Comparison of volatility errors in GARCH and ABC-BP neural network
model
MSE MAE MBE

GARCH 0.264 0.510 0.022

BS 0.001** 0.024* 0.001**
ABC-BP 0.000*** 0.010*** 0.000***
Note: Where *, **, *** indicate significance at the 10%, 5%, 1%.

of
ro
-p
re
lP
na
ur
Jo

38
Appendix A

The computational process of ABC-BP neural network model

1) Create a BP neural network model.

2) Initialize the parameters of the ABC algorithm, comprising the bee colony's size

(%~ ), the number of employed bees (%U ), the number of onlookers (%. ), the number

of solutions (%/ ), the limit (0 1 2), the maximum number of cycles (34%), and the

of
initial solution of the D-dimension { ( = 1, … , %/ ), and %~ , %U , %. and %/

ro
satisfy the following relation:
-p
%~ = 2%/ = %U + %. , %U + %.
re
(1)
lP

The D-dimensional vector { ( = 1, … , %/ ) represents the connection weights

na

and thresholds of the network created in (1), and the dimension D of each solution
ur

satisfies the following equation:

Jo

= %) * ×% U) +% U) +% U) × %.* * + %.* * (2)

%) * , % U) and %.* * are the number of neurons in the input layer,

the hidden layer, and the output layer respectively. The value of the initial solution is a

randomly generated number between (-1, 1).

3) Calculate the fitness of each solution according to equation (1).

1 3= =0
<({ ) = •
3= >0
(3)
R^€E

In this equation, = 1, … , %/ . 3= represents the BP neural network

39
 th
model mean squared error of the solution. When the fitness reaches 1, this would

be the most ideal situation clearly.

4) The employed bees search for a new solution based on the current solution

memoried before.

} = { + ?i(p(−1,1)({ − { ) (4)

is the number of the solution, ∈ {1, … … , }, ∈ {1, … … , %/ } are randomly

of
generated, and ≠ . The employed bees adopt the greedy mechanism. If the fitness

ro
of the new solution is larger than the fitness of the old solution, write down the new
-p
solution, otherwise the number of update failures of the old solution is increased by
re
one.5) Calculate the possible values (5 ) of each solution.
lP

ƒ(„E )
5 = †‡ (5)
∑…aA ƒ(„… )
na

$
<({ ) is the fitness of the solution. And ∑)#‡ <({) ) is the sum of the fitness of
ur
Jo

solutions. The onlooker searches for a new solution (Formula 4) from the

neighborhood of the existing solution based on these possible values.

6) If the failures number of the solution { updates exceeds the preset limit

value(0 1 2), it means that the solution can no longer be optimized, and it must be

discarded, and replaced by a new solution generated by the following formula:

{ = {T ) + ?i(p(0,1)({TˆV − {T ) ) (6)

{T ) is the smallest solution, and {TˆV is the largest one. Save the optimal

solution.

40
7) If the number of iterations is greater than the maximum number of cycles (34%),

the training ends. Otherwise, return to step (4).

8) The obtained optimal solution is transformed into the connection weights and

thresholds of the BP neural network model, and the neural network is simulated and

tested with data.

of
ro
-p
re
lP
na
ur
Jo

41
 Author statement

Manuscript title: Implied Volatility Forecast and Option Trading Strategy

I agree to be accountable for all aspects of the work in ensuring that questions
related to the accuracy or integrity of any part of the work are appropriately
investigated and resolved.

All persons who have made substantial contributions to the work are reported in the
manuscript, and those who provided editing and writing assistance but are not in the
authors list, are named in the Acknowledgments section of the manuscript and have

of
given their written permission to be named. If the manuscript does not include
Acknowledgments, it is because the authors have not received substantial

ro
contributions from non-authors.

-p
re
Lili Zhang
Associate Professor of Finance
lP

School of Economics and Management

Beijing Jiaotong University, China
na

Tel: +8613522593360
E-mail: [email protected]
ur
Jo