2014 Seventh International Symposium on Computational Intelligence and Design

The Leverage Effect and Fat-tails in China’s Futures Market

A bayesian analysis of stochastic volatility models

Guozhi An Lanjun Lao

Department of Management Science Department of Finance
School of Management, Fudan University School of Management, Fudan University
Shanghai, China Shanghai, China
e-mail: 13110690008@[Link] e-mail: gan13@[Link]

Abstract—Under the background that China is about to launch are introduced, section III displays the Bayesian approach to
options on futures trading, we investigate quantitatively the estimate the parameters, section IV is devoted to methods of
leverage effect and the degree of fat-tails in China’s futures Markov Chain Monte Carlo (MCMC) and Gibbs sampling,
market in a Bayesian approach by using the stochastic section V presents empirical study results of China’s futures
volatility models. To estimate the parameters of the models, market, and section VI is allocated to the conclusion.
Markov Chain Monte Carlo (MCMC) method and Gibbs
sampling are applied. II. STOCHASTIC VOLATILITY MODELS
Keywords-China’s Futures Market; Bayesian; MCMC; Stochastic volatility (SV) model, proposed by Taylor
leverage effect; fat-tails; stochastic volatility models (1986) [5], is an alternative class of nonlinear financial
models to the ARCH/GARCH specification which can
capture the time-varying properties of the conditional
I. INTRODUCTION volatility. Different with the ARCH/GARCH model, the SV
As part of its push to liberalize the economy, China will model allows the conditional volatility to follow a certain
launch equity index options on the China Financial Futures latent stochastic process. Compared to the ARCH/GARCH
Exchange and several options on futures on the main futures model’s one error process, there are two innovations to
exchanges. This makes it meaningful to investigate the describe the time-varying characteristics in the SV model.
volatility, especially the so called leverage effect and fat fails Nelson (1990) [6] and Duan (1997) [7] show that the
in the China’s futures market. ARCH/GARCH and SV models have strong similarities
Volatility is a so important issue in financial markets that asymptotically. However, with the more flexibilities
most forms of investing are affected by it to some degree. In introduced by the additional uncertainty in the volatility
investment terms, volatility relates to the rate at which the process from the SV model, a SV model often provides a
price of a financial instrument moves up or down. For those better in-sample fit than the ARCH/GARCH model, see
who deal with options and other derivative securities, Kim, Shephard and Chib (1998) [8]. Therefore, in this paper,
forecasting volatility accurately and managing the exposure we choose the Leverage-SV model and SV-T model to
of their investment portfolios to its effects are crucial. It is investigate the leverage effect and fat-tails of the volatility.
nigh on impossible to make any kind of precise forecasts
about how the price of options will move without having a A. The Basic SV Model
clear insight into volatility and the impact it has. The basic SV model was originally introduced by Taylor
Increasing attention has been focused on the analysis of (1986) [5]. There are two stochastic processes specified in a
the so called leverage effect, which corresponds to a negative SV model,
correlation between past returns and future volatility. This
stylized fact is first discussed by Black (1976) [1], who yt = exp(θ t / 2)ε t , ε t ~ i.i.d N (0,1)
observed that the volatility of stocks tends to increase when (1)
the price drops. This effect is particularly important for θt = μ + φ (θ t −1 − μ ) + ηt , ηt ~ i.i.d N (0,τ 2 )
option markets: not only does it imply that at-the-money
volatilities tend to increase after price drops, but also that a
significant skew in the volatility smile should appear [2] [3]. where yt is the return time series, which is defined as the
Another well-known stylized fact is ‘fat-tails’ in the logarithmic closing price differences. The process θ t is the
distribution of returns or long ranged volatility correlations,
latent log-volatility. The latent log-volatility is assumed to
which was proposed by Mandelbrot (1997) [4].
follow a stationary AR(1) process, which requires the
Our aim is to investigate the leverage effect and fat-tails
persistent parameter φ to be bounded by 1 in absolute
in China’s futures market. To analyze these stylized facts,
Leverage-SV and SV-t models based on Bayesian inference value, i.e., | φ |< 1 . τ is the standard deviation of the shock to
are used. In section II, the main stochastic volatility models θ t , ε t and ηt are normally and independently distributed.

978-1-4799-7005-6/14 $31.00 © 2014 IEEE 120

DOI 10.1109/ISCID.2014.163
B. SV-T Model According to Bayes theorem, the joint posterior
To investigate the fat-tails, we need to change the distribution of the non-observables given the data is
assumption that the return time series are normally proportional to the prior times the likelihood, i.e.
distributed. Hence, we use the student-T distribution to
describe the distribution characteristics. π ( μ , φ ,τ , ω , θ0:n | y0:n ) ∝ π ( μ , φ ,τ , ω , θ 0:n ) L( μ , φ ,τ ,θ 0:n ) (5)

yt = exp(θ t / 2)ε t , ε t ~ i.i.d t (0,1, ω )

(2)
θt = μ + φ (θ t −1 − μ ) + ηt , ηt ~ i.i.d N (0,τ 2 ) § 1 n · § ω + 1 · −n § ω ·
n

L ( μ , φ ,τ ,θ 0:n ) = exp ¨ − ¦θt ¸  n ¨ ¸ ¨ ¸

© 2 t =1 ¹ © 2 ¹ ©2¹
where ω is the degree of freedom. ω
(6)
n
yt exp(−θt ) − 2 1
2 +

(ωπ ) ∏[1 +
−n/ 2
C. Leverage-SV Model ]
t =1 ω
Under the assumption of the basic SV model, the
leverage effect cannot be explained. Therefore, we follow the where L ( μ , φ ,τ ,θ 0:n ) is the likelihood of the joint model,
leveraged SV specification from Harvey and Shephard
(1996) [9] and Yu (2005) [10] and assume the following θ0:n denotes {θ0 ,θ1 ,θ 2 ," ,θ n } .
bivariate structure: Set the parameters of the prior distribution is a
prerequisite for Bayesian statistical analysis. According to
§ εt · ­°§ 0 · § 1 ρτ · °½ Jun Yu and Renate Meyer (2000) [12], we set the prior
¨ ¸ ~ i.i.d Ν ®¨ ¸ , ¨ ¾ distributions of the variables μ , φ , ω , τ and the latent log-
τ 2 ¸¹ ¿°
(3)
η
© t +1 ¹ °¯© 0 ¹ © ρτ volatility θ as

where ρ measures the correlation between ε t and ηt +1 .

φ ~ Be ( 20,1.5) , τ 2 ~ IGa ( 2.5, 0.025 ) , μ ~ N ( 0, 0.01)
(7)
When ρ < 0 , it refers to a so-called leverage effect, a drop ω ~ Χ 2 ( 8 ) I ( 4, 40 ) , θ 0 ~ N ( μ ,τ 2 )
in the return followed by an increase in the volatility.
The principles of the Bayesian inference of the Leverage-
SV model is similar to the SV-T model. Hence, in this paper
III. BAYESIAN INFERENCE OF THE PARAMETERS we just show the initial parameters set of the prior
Zhu (2008) [11] show that SV-T model is more distributions, according to Kim (1998) [8],
appropriate to describe the characteristics of China’s
financial markets than the basic SV model. Consequently, in φ ~ Be ( 20,1.5) , τ 2 ~ IGa ( 2.5, 0.025 ) , μ ~ N ( 0, 0.01)
this paper, we will use the SV-T and Leverage-SV model to (8)
analyze the stylized facts of China’s futures market. A θ0 ~ N ( μ ,τ 2 ), ρ ~ U (−1,1)
complete Bayesian analysis model consists of the joint prior
distribution of all the non-observables and the observables.
In the SV-T model, the unobservable variables are μ , φ , IV. MCMC AND GIBBS SAMPLING
ω , τ and the latent log-volatility θ , the observable variable Up to now, SV model parameter estimation methods
is the return time series y . Bayesian inference is based on usually include: a Quasi-Maximum Likelihood method
(QML), Generalized Method of Moments (GMM),
the non-observables’ posterior distributions. The joint prior
Simulated Maximum Likelihood method (SML), the Monte
distribution of the parameters μ , φ , ω , τ and the latent log- Carlo method for Maximum Likelihood (MCML), Markov
volatility θ can be re-expressed as a product of three chain Monte Carlo method (MCMC). Among which,
components: MCMC method is a relatively accurate and efficient
parameter estimation method which developed in recent
years Chib (2002) [13]. This method introduces Markov
π ( μ , φ ,τ , ω , θ 0:n ) = π ( μ , φ ,τ , ω ) π (θ 0 |μ , φ ,τ , ω )
process into the Monte Carlo simulation, which effectuates
n
the dynamic simulation and overcome the existing static and
∏π (θ |θ
t =1
t t −1 , μ , φ ,τ , ω ) high-dimensional defects in the traditional Monte Carlo
(4) simulation.
=  ( μ ) π (φ ) π (τ ) π (ω ) π (θ 0 | μ , φ ,τ , ω ) The core idea of the MCMC method is: (1) preset the
n prior distribution of the parameters to be estimate. (2)
∏π (θ |θ
t =1
t t −1 , μ , φ ,τ , ω ) according to the Bayes theorem, introduce the corresponding
model, construct a Markov chain with stationary distribution
π ( x) (i.e., the posterior distribution of the estimated
parameters). (3) with the help of ergodic characteristic of

121
Markov chain theory, repeated sampling on the chain,
eventually get the estimated parameter values.
This paper will select Gibbs sampling to realize the
MCMC simulation, which can be regarded as a special case
of Metropolis-Hastings algorithm. When the sample is large
enough, it can produce accurate posterior mean. Because of
the good convergence and easiness to direct sampling and
simulation, Gibbs sampling is widely used in recent years
[13], [14] and [15]. This paper will use MCMC-Gibbs
method for Bayesian inference of the SV models, mainly
using the statistical software R and Openbugs.
V. EMPIRICAL STUDY

A. Data Description
Because of the expiration limits of futures contracts, we
will choose the continuous indexes of the futures to analyze Figure 1. Bayesian estimation results of parameter ρ
volatility of China's futures market. Mainly choose copper,
PTA, soybean meal (SM) and stock index futures (IF). The
sample data are selected from 2009 to present (February 11,
2014), in which the stock index futures data are from the As Table.2 shows, the overall level of the persistent
launch date on April 19, 2010. All the data are from CSMAR parameter φ is higher, especially in copper and PTA, of
database. Returns are calculated by the following formula, which the level of persistence reached to 0.9542 and 0.9405,
where Pt is the closing price of day t. indicating that these two futures have strong volatility
persistence. However, the persistent level of stock index
futures is only 0.5538, lower than copper and PTA.
Rt = log(
Pt
) (9) Disturbance parameter τ indicates the standard deviation of
Pt −1 the shock to θ , it is higher in the stock index futures and
soybean meal futures. Parameter ρ represents the leverage
The demeaned return series yt is defined as, yt = Rt − R1:t effect of volatility, the more negative the value indicates that
the leverage effect is more obvious. From Table.2, we can
B. Emprical Analysis find that stock index futures and copper futures have the
The descriptive statistics of the four kinds of futures are most obvious leverage effect. The leverage effect of the
shown in Table.1, the fat tail phenomenon is very common. stock index futures reached to 0.2369, the 95% confidence
Such as SM’s kurtosis reach to 10.6398, skewness reach to - interval is [-0.5473, 0.02434], which means stock index
1.42256. futures has a strong leverage effect.

TABLE I. SUMMARY STATISTICS OF THE DAILY RETURNS TABLE II. BAYESIAN ESTIMATE RESULTS OF LEVERAGE-SV MODEL

Copper IF SM PTA Futures Paras mean sd MCerror 95% Interval

Mean 0.0003 -0.0002 0.0001 0.0001 ߤ -0.9100 0.2896 0.0174 [-1.3560,-0.1769]

Max 0.0263 0.0243 0.0203 0.0418 Copper

߶ 0.9542 0.0287 0.0019 [0.8728,0.9829]
ߩ -0.2372 0.1141 0.0073 [-0.4364,-0.0067]
Min -0.0285 -0.0321 -0.0537 -0.0283
߬ 0.1881 0.0335 0.0023 [0.1316,0.2515]
Std 0.0071 0.0061 0.0060 0.0062 ߤ -0.2200 0.6019 0.0422 [-0.6869,1.8010]

Median 0.0002 -0.0003 0.0003 0.0001 Stock ߶ 0.5538 0.2920 0.0202 [0.0927,0.9495]
Index ߩ -0.2369 0.1633 0.0106 [-0.5473,0.0243]
Skew -0.0807 0.0582 -1.4226 0.1965
߬ 0.5929 0.2859 0.0200 [0.1511,1.0110]
Kurt 2.2840 2.1950 10.6398 3.9530 ߤ -0.6383 0.0827 0.0031 [-0.7959,-0.4688]
Soybean ߶ 0.7372 0.0897 0.0059 [0.5365,0.8694]
Meal ߩ -0.1047 0.0825 0.0045 [-0.2610,0.0622]
Leverage-SV model is used to analysis the leverage ߬ 0.5606 0.0993 0.0067 [0.3963,0.7595]
effect of four futures, Table.2 shows the mean, stand
deviation, MC simulation error and 95% interval of the
ߤ -0.5797 0.1641 0.0045 [-0.8962,-0.2490]

models’ parameters. The Bayesian estimation results of the PT A

߶ 0.9405 0.0222 0.0014 [0.8852,0.9729]

four futures’ posterior density functions of the leverage ߩ 0.0612 0.0895 0.0052 [-0.1104,0.2320]
effect parameter ρ are shown in Figure.1. ߬ 0.2863 0.0499 0.0034 [0.2160,0.4152]

122
 VI. CONCLUSION
In this paper, we investigate the leverage effect and fat-
tails in the China’s four main futures markets. Based on the
Bayesian inference and MCMC simulation, we use SV-T
and Leverage-SV model to detect these two stylized facts.
We got many useful conclusions by analyzing the volatility
of different kinds of China’s futures markets. The main
conclusions are that the four studied futures markets all have
some degree of leverage effect and fat-tails features, in
which the copper and the stock index futures have the
sharpest leverage effect. The degree of fat-tails of stock
index futures is less than the other three commodity futures.
The analysis results in this paper can provide some supports
to the research on China’s financial market in the future.
Figure 2. Bayesian estimation results of parameter ω Hopefully, the discussion of the leverage effect and fat-tails
in China’s futures market can provide useful information for
the formulation of investment strategies and portfolio
management.
Then the fat-tails are analyzed by using the SV-T model.
The relevant Bayesian estimate results of the four futures’ REFERENCES
parameters are shown in Table.3. The volatility persistence [1] F. Black, Studies of stock market volatility changes. Proceedings of
levels of the four futures are still very high. The parameter the American Statistical Association, Business and Economic
Statistics Section, 1976, pp. 177–181.
ω indicates the degree of fat-tails, values of four futures in
[2] D. Backus, S. Foresi, K. Li, et al. Accounting for biases in Black-
the table are significantly larger than 0, which means that the Scholes. CRIF Working Paper series, 1997: 30.
four futures all have some degree of fat-tails. The parameter [3] F. Jean-Peirre, G. Papanicolaou and K [Link]. Derivatives in
ω of Copper is 15.42, while the stock index future’s is just financial markets with stochastic volatility. Cambridge University
3.543. The fat-tails of these three commodity futures are Press, 2000.
significantly greater than the degree of stock index futures. [4] B B. Mandelbrot, The variation of certain speculative prices. Springer
That also reflects that the stock index futures market has New York, 1997.
more efficiency than the commodity futures markets. [5] S.J. Taylor, Modelling financial time series, Chichester:John Wiley,
1986
TABLE III. BAYESIAN ESTIMATE RESULTS OF SV-T MODEL [6] D. Nelson, “Stationarity and persistence in the GARCH(1,1) model”.
Econometric Theory,1990,6, pp. 318–34.
Futures Paras mean sd MCerror 95% Interval [7] J. Duan, “Augmented GARCH(p,q) Process and its Diffusion Limit”,
ߤ -0.1734 2.6430 0.0019 [-1.701,10.23] Journal of Econometrics, 1997, 79, pp. 97-127
[8] S. Kim, N. Shephard and S. Chib. “Stochastic volatility: likelihood
Copper ߱ 15.4200 9.8740 0.0068 [2.885,40.37]
inference and comparison with ARCH models”. The Review of
߶ 0.9520 0.0769 0.0053 [0.6573,0.9934] Economic Studies, 1998, 65(3),pp. 361-393.
߬ 0.1362 0.0367 0.0025 [0.0714,0.2085] [9] A. Harvey,E. Ruiz, and N. Shephard, “Estimation of an Asymmetric
ߤ -7.1080 8.1570 0.0058 [-19.38,-0.2246] Stochastic Volatility Model for Asset Returns,” Journal of Business
and Economic Statistics, 1996, 14,pp. 429-434.
Stock ߱ 3.5430 4.7120 0.0033 [0.0985,15.99]
[10] J. Yu, On leverage in a stochastic volatility model. Journal of
Index ߶ 0.8746 0.0966 0.0065 [0.6296,0.9791] Econometrics , 2005, 127, pp. 165–178.
߬ 0.1097 0.0347 0.0024 [0.0647,0.1948] [11] R. Meyer, J. Yu. “BUGS for a Bayesian analysis of stochastic
ߤ 0.8366 3.3040 0.0023 [-1.097,11.51] volatility models”. The Econometrics Journal, 2000, 3(2), pp. 198-
215.
Soybean ߱ 14.5400 13.2600 0.0092 [4.594,44.94]
[12] H.M. Zhu, “Bayesian Heavey-tailed Stochastic Volatility Model in
Meal ߶ 0.9545 0.0606 0.0042 [0.7499,0.9939] Financial Analysis Based on MCMC Simulaiton”, Journal of System
߬ 0.1168 0.0261 0.0018 [0.0752,0.1787] Simulation, 2008, 20(9) pp. 2479-2482.(in Chinese)
ߤ 0.2349 2.8070 0.0020 [-1.206,10.97] [13] S. Chib, Greenberg, Understanding the Metropolis–Hastings
algorithm. The American Statistician ,1995, 49,pp. 327–335
PT A ߱ 13.1100 11.2400 0.0078 [4.791,40.81]
[14] S. Chib, F. Nardari,N. Shephard, Markov chain Monte Carlo methods
߶ 0.9586 0.0552 0.0038 [0.7935,0.9941] for stochastic volatility models. Journal of Econometrics ,2002,
߬ 0.1459 0.0312 0.0021 [0.0974,0.2178] 108,pp. 281–316.
[15] C.K. Carter,R. Kohn, On Gibbs sampling for state space models.
Biometrika 1994, 81,pp. 541–553

123