Economic Modelling 53 (2016) 216–223

Contents lists available at ScienceDirect

Economic Modelling

journal homepage: www.elsevier.com/locate/ecmod

Forecasting VaR and ES using dynamic conditional score models and

skew Student distribution
Chun-Ting Gao ⁎, Xiao-Hua Zhou
School of Economics and Business Administration, Chongqing University, China

a r t i c l e i n f o a b s t r a c t

Article history: Dynamic conditional score (DCS) model is a new type of observation-driven model based on the score function.
Accepted 6 December 2015 Based on time-varying scale, Harvey (2013) proposed univariate DCS models with skew Student distribution. We
Available online 30 December 2015 redescribe these models based on log-variance and extend them to models based on variance. In the modeling
process, calculating score of the distribution is a key part. When selecting scale or variance to calculate the
Keywords:
score, two subclasses of DCS models with skew Student distribution are derived. We study them separately in
DCS models
VaR
detail, and the models with leverage have been presented too. For estimating one-day-ahead VaR and ES, we
ES combine these models with AR(1) into new ones. By backtesting their performance for four international stock
Skew Student distribution indices, we find that all these AR–DCS models perform well on forcasting VaR and ES at the confidence levels
Market risk 95%, 99% and 99.5%.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction error distribution (GED) are the wildly used ones. Empirical research
findings also found that the distribution of asset returns is asymmetri-
Value at risk (VaR) may be the most well-known measure of market cal, i.e., it is skewed. So it is natural to bring skewness into a certain
risk. It had been widely adopted by financial institutions and studied by distribution. There are several methods in describing the skewness
researchers after it was first introduced by J.P. Morgan in 1994. By defi- of probability distribution. Such as the skew Student distribution
nition, VaR is the maximum likely loss over some target period, at a of Hansen (1994), the skew generalized Student distribution of
specified probability level. For its easy understanding and sound evalu- Theodossiou (1998). Fernández and Steel (1998) produced a skewed
ating method, the risk measure models based on VaR had developed method which is suitable to many distributions such as Normal
rapidly in literature. However, VaR is not a perfect risk measure. distributon, Student distribution and GED. Another key factor in VaR
Artzner et al. (1999) proved that VaR is not a coherent risk measure and ES calculation is volatility. But the volatility of asset returns is not di-
for it's against the axiom of subadditivity sometimes. Another common- rectly observable in practice. So when estimating VaR and ES, we must
ly used risk measure is expected shortfall (ES). Unlike VaR, ES is a coher- calculate volatility first. There are many volatility models available in
ent risk measure, and ES refers to the expected loss given that the loss the literature. GARCH models which were proposed by Bollerslev
exceeds VaR. Therefore, ES is also called the conditional value at risk (1987) are simple to estimate than many other volatility models and
(CVaR) in some literatures. When calculating VaR and ES, the greatest widely used in VaR and ES calculation. A simple review of the empirical
challenge is to estimate the probability distribution of asset returns. literature proved that the models combining GARCH-type models with
Assuming a probability distribution and effectively measuring volatility skew Student distributon are in the best VaR models. Giot and Laurent
are essential steps in most cases. (2003) studied APARCH model based on the skew Student distribution.
Normal distribution is one of the most widely used distributions in Their empirical findings on three stock market indexes and three US
risk measuring and asset pricing. However, when using the test of stocks showed that their model perform well on forecasting VaR in all
Jarque and Bera (1987) to examine the distribution of asset returns, it cases. Kuester et al. (2006) had done a comparison on several VaR
generally rejects the hypothesis of normality. Much of the literature models and found that only ARCH models yield acceptable forecasts.
had proved that fat-tailed distributions may be more appropriate The skewed-t GARCH is one of the best performance models. Following
choice. Among these distributions, Student distribution and generalized Giot and Laurent (2003), Diamandis et al. (2011) reconsidered the
APARCH model based on the skew Student distribution for daily returns
of 21 stock indexes. Their findings reconfirmed the advantage of skew
Student distribution GARCH-type model on VaR estimation. Zhu and
⁎ Corresponding author at: No. 174 Shazhengjie, Shapingba, Chongqing 400044, China.
Tel.: +86 15922887848.
Galbraith (2011) had used GARCH-class models with a transformation
E-mail addresses: [email protected] (C.-T. Gao), [email protected] of this skew Student distribution to forecast ES of several company
(X.-H. Zhou). stocks and SP500.

http://dx.doi.org/10.1016/j.econmod.2015.12.004
0264-9993/© 2015 Elsevier B.V. All rights reserved.
 C.-T. Gao, X.-H. Zhou / Economic Modelling 53 (2016) 216–223 217

Harvey (2013) and Creal et al. (2013) found that the scores of some where θ denotes a vector of unknown parameters. ω, α and β are the un-
probability distribution function are martingale difference (MD). They known coefficients. st is a specified function of previous data. Eq. (2)
applied this property to GARCH-type models and obtained new type shows the dependence relationship of ft. When ft denotes variance or
ones namely DCS models (Creal et al. called them GAS models). log-variance, Eq. (2) will be a more extensible framework than
Although DCS models have a similar form like GARCH models, but the GARCH-type models. The critical part of GAS models is the choice of st.
research using them on estimating VaR and ES is still rare. Lucas and When st is simply derived from lagged observations, it will be the
Zhang (2014) introduced the GAS structure into EWMA and obtained general GARCH models. In the GAS models, st is calculated by score ∇t
a new range of EWMA models. By examining the VaR forecasting perfor- and matrix function St.
mance, they found that there is the best one in the new type models.
Bernardi and Catania (2015) used the model confidence set (MCS) to st ¼ St ∇t ð3Þ
compare the VaR forecasting performance of Gaussian GAS model,
Student GAS model, ARCH-type models and etc. Their empirical results ∂lnpðyt jf t ; θÞ
∇t ¼ ð4Þ
show that GAS models are equivalent with GARCH models at confidence ∂f t
level 95% in a statistical sense.
In this paper, our main objective is using the DCS models with skew Creal et al. (2012) proposed a choice of St as follows
Student distribution to calculate one-day-ahead VaR and ES. Harvey

−1
(2013) studied the DCS models with skew Student distribution by fo- St ¼ −Et−1 ∇t ∇0t ð5Þ
cusing on the time-varying log-scale of this distribution. We rebuild
these models on time-varying log-variance and extend them by model- Both Creal et al. (2012) and Harvey (2013) studied the Student
ing on time-varying variance. When studying DCS models with skew distribution. We first do some comparative study on the difference
Student distribution, Harvey (2013) points out that there are two between the GAS models and DCS models of this distribution. Let the
methods to construct DCS models (resulting from two ut). By combining observations be generated by
them with an AR(1) conditional mean equation, we construct 4 AR–DCS
models respectively, such as variance or log-variance, including or not yt ¼ φt ε t ; t ¼ 1; …; T: ð6Þ
including leverage. Selecting four stock indices as sample, we test VaR
and ES out-sample forecasting performance of these AR–DCS models where εt follows Student distribution and is serially independent, φt is
at the confidence levels 95%, 99% and 99.5%. Our results show that: all the time-varying scale of this distribution. So yt follows a distribution
these AR–DCS models perform very well on forecasting VaR and ES with the following probability density function
not only at the confidence level 95% but also at more extreme
confidence levels 99%, 99.5% even without combining EVT.  −ðνþ1Þ=2
Γ ððν þ 1Þ=2Þ y2
The rest of this paper is organized as follows: Section 2 begins with f ðyt jφt ; ν Þ ¼ pffiffiffiffiffiffi 1 þ t 2 ; ν N2 ð7Þ
Γ ðν=2Þφt πν νφt
summarizing the main idea of DCS models with the Student distribu-
tion, and then extends them with skewness. In this part, we obtain where Γ(•)is the gamma function, v is the number of degrees of freedom.
two subsets of DCS models including variance form, log-variance form With the relation between φt and the standard deviation σt of yt, we can
and with or without leverage form. Section 3 provides the calculating also let the observations be written as
methods of VaR and ES. Section 4 analyzes the backtesting performance
of one-day-ahead VaR and ES forcasting results. At last, we conclude the yt ¼ σ t ηt ; t ¼ 1; …; T: ð8Þ
paper with brief remarks in Section 5.
where ηt = ((ν − 2)/ν)1/2εt still has a Student distribution, but
2. Methods standardized so as to have unit variance. And the probability density
function in Eq. (7) can be changed to
A new general framework for modeling time variation in parametric
models was proposed by Harvey (2013) and Creal et al. (2013) sepa-  −ðνþ1Þ=2
Γ ððν þ 1Þ=2Þ y2t
rately. The former called it as dynamic conditional score (DCS) model, f ðyt jσ t ; νÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 þ ð9Þ
Γ ðν=2Þσ t π ðν−2Þ ðν−2Þσ 2t
and the later named it as generalized autoregressive score (GAS)
model. This new general framework not only can be applied to take
With the upper Eq. (9), we can get the value of ∇t and st by calcula-
the place of GARCH-type models and ACD-type models, but also can
tion. For ft = σt2 and ft = ln σt2, we may obtain two basic forms of
be used in modeling the time-varying correlation with copulas. In this
GAS(1,1) models with the specification of st. For ft = σt2,
paper, we will focus on its application to replace GARCH-type models.
For the great contribution of Harvey on univariate models, we mainly
∂ ln f ðyt ; σ t ; νÞ ðν þ 1Þy2t 1
study DCS models. ∇t ¼ ¼  − 2 ð10Þ
∂σ t2 2σ t ðν−2Þσ 2t þ y2t
2 2σ t
2.1. Univariate DCS models  
νþ3 2 ðν þ 1Þy2t
st ¼ ∇t St ¼ σt −1 ð11Þ
Although the main idea of Creal et al. (2013) and Harvey (2013) is ν ðν−2Þσ t þ yt
2 2

similar, their deducing process is different from each other. The starting
points of both their modeling come from calculating the score of some For ft = ln σt2, following the upper way, we can get the representa-
conditional probability distribution. tion of st step by step. That is
Creal et al. (2012) proposed the deducing process of univariate GAS  
models in detail. Assuming that the asset returns yt follows conditional νþ3 ðν þ 1Þy2t
st ¼ −1 ð12Þ
probability distribution p(yt| ft;θ), the basic form of GAS (1,1) models is ν ðν−2Þσ t þ yt
2 2

specified as:
Harvey (2013) also did similar work when the distribution is
yt
pðyt jf t ; θÞ; ð1Þ Student. Their derivation process is straightforward. He got two
main models and named them Beta-t-GARCH and Beta-t-EGARCH
f tþ1 ¼ ω þ αst þ β f t ; ð2Þ respectively.
218 C.-T. Gao, X.-H. Zhou / Economic Modelling 53 (2016) 216–223

The Beta-t-GARCH model is composed of the following two respectively, whereas γ = 1 indicates that it is the same as without
equations: skewness. Eq. (19) also can be expressed as a more simple form


σ 2tþ1 ¼ ω þ ασ 2t ut þ βσ 2t ð13Þ 2
f ðyt jγ Þ ¼ f yt =γ signðyt Þ ð20Þ
γ þ γ −1
ðν þ 1Þy2t
ut ¼ −1 ð14Þ where sign(•) is a sign function. When f(•) is Student distribution in Eq.
ðν−2Þσ 2t þ y2t
(7), we will obtain the skew Student distribution
where ut is in proportion to the score of log-density of Student distribu-  
νþ1
tion. ut is a martingale difference (MD). ω N 0, α N 0 and 1 N β N 0. 2Γ  −νþ1
2 y2t 2
Actually ut = 2σt2∇t holds in the model Beta-t-GARCH. Harvey (2013) f ðyt j0; φt ; ν; γÞ ¼   1þ ð21Þ
pffiffiffiffiffiffi ν νφ2 γ 2signðyt Þ
proposed that (ut + 1)/(ν + 1) follows the distribution of Beta(1/2, ν/2). ðγ þ γ−1 Þφt νπΓ t

The Beta-t-EGARCH model is not directly built on log-variance but 2

time-varying log-scale. It is very easy to transform it to the form based
on log-variance: Here, we introduce another parameter δt which Lambert and
pffiffiffiffiffiffi
Laurent (2001) named dispersion parameter. If letting φt ¼ δt ν−2
ν ,

lnσ 2tþ1 ¼ ω þ αut þ β ln σ 2t ð15Þ there will be another equivalent formation with Eq. (21). Giot and
Laurent (2003) proposed that this skew Student distribution has a
mean m and variance s2 when δt = 1.
Where the ut is equal to the one in Eq. (14). By a simple comparison of
the models from Creal et al. (2012) and Harvey (2013), we can find that rffiffiffiffiffiffiffiffiffi
 
ν−1 ν−2
ν
for variance and log-variance, st = (ν + 3)σt2ut/ν and st = (ν + 3)ut/ν m ¼ γ−γ −1 Γ =Γ ð22Þ
2 π 2
hold respectively. When estimating parameters, these two class models
will get the same results except for the parameter α.
s2 ¼ γ2 þ γ −2 −m2 −1 ð23Þ
In the financial market, there are asymmetric effects on volatility
between positive shocks and negative shocks. To address this problem,
Harvey and Sucarrat (2014) had provided the mean and variance
Glosten et al. (1993) incorporated leverage effects in GARCH and
of yt.
obtained a new model. Subsequent researchers called this model as
GJR model. Harvey (2013) adopted a similar way to introduce leverage qffiffiffiffiffiffiffiffiffi
ν
effects into Beta-t-GARCH and Beta-t-EGARCH. Eðyt Þ ¼ μ t ¼ mφt ð24Þ
ν−2
The Beta-t-GARCH model with leverage is
ν 2 2
varðyt Þ ¼ σ 2t ¼ s φt ð25Þ
ν−2
σ 2tþ1 ¼ωþ ασ 2t ut þ κσ 2t ðut þ 1ÞIt ðyt b0Þ þ βσ 2t ð16Þ
Harvey (2013) proposed that when the distribution is skew Student,
where ω N 0 and β + κ/2 b 1. ut has two different forms (page 144 and page 146). We deduce in a
The Beta-t-EGARCH model with leverage is different way and obtain two ut forms which are transformations of
Harvey's. For convenience, we label the two processing modes by
lnσ 2tþ1 ¼ ω þ αut þ κ ðut þ 1ÞIt ðyt b 0Þ þ β ln σ 2t ð17Þ method A and method B.

In Eqs. (16) and (17), ut equals 2σt2∇t too. κσt2(ut + 1)It(yt b 0) and 2.2.1. Method A
κ(ut + 1)It(yt b 0) are the leverage term in each model where It is a To obtain the form with time-varying variance, we first do some
indicator function, and transformation to Eq. (21). Let zt = yt − mσt/s, so E(zt) = 0 and
var(zt) = var(yt) = σt2 hold. Further combining with Eq. (25), the
1; yt b 0 density function (21) can be changed to
It ¼ ð18Þ
0; yt ≥ 0  
νþ1 !−νþ1
2Γ s
ðzt s þ mσ t Þ2
2
2
When κ N 0, negative shocks will increase the volatility at next time. f ðzt j0; σ t ; ν; γ Þ ¼   1þ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ν ðν−2Þσ t γ
2 2sign ðz t sþmσ t Þ
And when κ b 0, negative shocks will decrease the volatility. ðγ þ γ−1 Þσ t ðν−2ÞπΓ
2
ð26Þ
2.2. Univariate DCS models with skewness

Fernández and Steel (1998) proposed a widely used method to Following the derivation process in Appendix A.1, we can obtain the
describe skewness feature of asset returns. Harvey (2013) studied the score against σt2
DCS models with this skewed method for Student distribution and
GED distribution. In this paper, we review the DCS models with skew ∂ ln f ðzt j0; σ t ; ν; γÞ
∇t ¼
Student distribution and do some extensions. For the convenience of ∂σh 2t i
empirical study latter, we first talk about the skewed method. The ðν þ 1Þ ðzt s þ mσ t Þ2 −mσ t ðzt s þ mσ t Þ 1
skewed method of Fernández and Steel (1998) is as below ¼
− 2 ð27Þ
2σ 2t ðν−2Þσ 2t γ 2signðzt sþmσ t Þ þ ðzt s þ mσ t Þ2 2σ t
2

f ðyt jγ Þ ¼ f ðyt =γÞI ½0;þ∞Þ ðyt Þ þ f ðyt γÞIð−∞;0Þ ðyt Þ ð19Þ
γ þ γ −1
Here we get ut by ut = 2σt2∇t, so
where f(•) in the right part is a standardized probability density function
h i
with the characters of unimodal and symmetric about zero. I is an ðν þ 1Þ ðzt s þ mσ t Þ2 −mσ t ðzt s þ mσ t Þ
indicator function, when yt ∈ [0, + ∞), I = 1; or else I = 0. γ is skewness ut ¼ −1 ð28Þ
parameter. γ N 1 and 0 b γ b 1 can generate right and left skewness ðν−2Þσ 2t γ2signðzt sþmσ t Þ þ ðzt s þ mσ t Þ2
 C.-T. Gao, X.-H. Zhou / Economic Modelling 53 (2016) 216–223 219

For forecasting one-day-ahead VaR, it is valuable to introduce a Following the way mentioned before, let zt = yt − mσt/s, we can
conditional mean Eq. AR (1) which was once used in McNeil and Frey change the upper Eq. (35) to
(2000).
ðν þ 1Þðzt s þ mσ t Þ2
  ut ¼ −1 ð36Þ
r t ¼ λr t−1 þ zt ; zt ∼skt 0; σ 2t ; ν; γ ð29Þ ðν−2Þσ 2t γ 2signðzt sþmσ t Þ þ ðzt s þ mσ t Þ2

where rt is log asset return at time t, skt(0, σt2, ν, γ) indicates a skew Comparing ut in Eqs. (28) and (36), we can find that the latter has a
Student distribution with v degrees of freedom, mean zero, variance term ðν−2Þσ−mσ
γ2
ðz sþmσ Þðνþ1Þ
t t t
2signðzt sþmσ t Þ
þðz sþmσ Þ
. Harvey (2013) proved that both Eqs. (28)
t t
2
t

σt2 and skewness parameter γ. and (36) are MD. So, E½ðν−2Þσ−mσ ðz sþmσ Þðνþ1Þ
 ¼ 0.
t t t
2signðzt sþmσ t Þ 2
γ þðz sþmσ Þ
2
t t t
When combining the upper Eqs.(28) and (29) with Eqs. (13), (15),
Here we also consider the conditional mean Eq. (29). When making
(16) and (17) separately, we will obtain a range of DCS models includ-
up the upper Eqs. (29) and (36) with Eqs. (13), (15), (16) and (17)
ing conditional mean equation.
separately, we will obtain another range of DCS models including
When integrating with Eq. (13), we will obtain the following model
conditional mean equation.
and named it AR-DCS-A:
When integrating with Eq. (13), we will obtain the following model
  and named it AR-DCS-B:
r t ¼ λr t−1 þ zt ; zt ∼skt 0; σ 2t ; ν; γ
2 2 2  
σ t ¼ ω þ ασ t−1 h ut−1 þ βσ t−1 i r t ¼ λrt−1 þ zt ; zt ∼skt 0; σ 2t ; ν; γ
2 ð30Þ 2 2
σ t ¼ ω þ ασ t−1 ut−1 þ βσ t−1 2
ðν þ 1Þ ðzt−1 s þ mσ t−1 Þ −mσ t−1 ðzt−1 s þ mσ t−1 Þ
ut−1 ¼ −1 ð37Þ
ðν þ 1Þðzt−1 s þ mσ t−1 Þ2
ðν−2Þσ 2t−1 γ 2signððzt−1 sþmσ t−1 Þ þ ðzt−1 s þ mσ t−1 Þ2 ut−1 ¼ −1
ðν−2Þσ t−1 γ 2signðzt−1 sþmσ t−1 Þ þ ðzt−1 s þ mσ t−1 Þ2
2

When integrating with Eq. (15), we will obtain the following model
When integrating with Eq. (15), we will obtain the following model
and named it AR-EDCS-A:
and named it AR-EDCS-B:
   
r t ¼ λr t−1 þ zt ; zt ∼skt 0; σ 2t ; ν; γ r t ¼ λrt−1 þ zt ; zt ∼skt 0; σ 2t ; ν; γ
ln σ 2t ¼ ω þ αuht−1 þ β ln σ 2t−1 i
2
lnσ t ¼ ω þ αut−1 þ β ln σ t−1 2
ð31Þ ð38Þ
ðν þ 1Þ ðzt−1 s þ mσ t−1 Þ2 −mσ t−1 ðzt−1 s þ mσ t−1 Þ ðν þ 1Þðzt−1 s þ mσ t−1 Þ2
ut−1 ¼ −1 ut−1 ¼ −1
ðν−2Þσ 2t−1 γ2signðzt−1 sþmσ t−1 Þ þ ðzt−1 s þ mσ t−1 Þ2 ðν−2Þσ t−1 γ 2signðzt−1 sþmσ t−1 Þ þ ðzt−1 s þ mσ t−1 Þ2
2

When integrating with Eq. (16), we will obtain the following model
When integrating with Eq. (16), we will obtain the following model
and named it AR-DCSL-B:
and named it AR-DCSL-A:
 
  r t ¼ λrt−1 þ zt ; zt ∼skt 0; σ 2t ; ν; γ
r t ¼ λr t−1 þ zt ; zt ∼skt 0; σ 2t ; ν; γ σ 2t ¼ ω þ ασ 2t−1 ut−1 þ κσ 2t−1 ðut−1 þ 1ÞIt−1 ðzt−1 b 0Þ þ βσ 2t−1
σ 2t ¼ ω þ ασ 2t−1 2 2 ð39Þ
h ut−1 þ κσ t−1 ðut−1 þ 1ÞIt−1 ðzt−1 b 0Þ þ βσi t−1 ðν þ 1Þðzt−1 s þ mσ t−1 Þ2
2 ð32Þ ut−1 ¼ −1
ðν þ 1Þ ðzt−1 s þ mσ t−1 Þ −mσ t−1 ðzt−1 s þ mσ t−1 Þ ðν−2Þσ t−1 γ 2signðzt−1 sþmσ t−1 Þ þ ðzt−1 s þ mσ t−1 Þ2
2
ut−1 ¼ −1
ðν−2Þσ 2t−1 γ2signðzt−1 sþmσ t−1 Þ þ ðzt−1 s þ mσ t−1 Þ2
When integrating with Eq. (17), we will obtain the following model
When integrating with Eq. (17), we will obtain the following model and named it AR-EDCSL-B:
and named it AR-EDCSL-A:  
r t ¼ λrt−1 þ zt ; zt ∼skt 0; σ 2t ; ν; γ
  lnσ 2t ¼ ω þ αut−1 þ κ ðut−1 þ 1ÞI t−1 ðzt−1 b 0Þ þ β ln σ 2t−1
r t ¼ λr t−1 þ zt ; zt ∼skt 0; σ 2t ; ν; γ ð40Þ
ðν þ 1Þðzt−1 s þ mσ t−1 Þ2
ln σ t ¼ ω þ αuht−1 þ κ ðut−1 þ 1ÞIt−1 ðzt−1 b 0Þ þ β ln σ 2t−1 i
2 ut−1 ¼ −1
ð33Þ ðν−2Þσ 2t−1 γ 2signðzt−1 sþmσ t−1 Þ þ ðzt−1 s þ mσ t−1 Þ2
ðν þ 1Þ ðzt−1 s þ mσ t−1 Þ2 −mσ t−1 ðzt−1 s þ mσ t−1 Þ
ut−1 ¼ −1
ðν−2Þσ 2t−1 γ2signðzt−1 sþmσ t−1 Þ þ ðzt−1 s þ mσ t−1 Þ2
2.3. Maximum likelihood estimation of DCS models

2.2.2. Method B Applying maximum likelihood theory to estimate these AR-DCS

The difference between method B and method A is choosing which models, we can get the parameters. The log likelihood of standardized
parameter to calculate score against. In this method, we not calculate skew Student distribution can be calculated from its probability density
the score against variance but the square of scale. function. When the sample size is T, the log likelihood is
Following the derivation process in Appendix A.2, we first get the !  2
score against φ2t 2Γ ðνþ1Þ 1X T
s
lnLðθÞ ¼ T ln pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 þ ln
ðν−2Þπðγ þ γ ÞΓ ð2Þ −1 ν 2 t¼1 σ 2t
! ð41Þ
∂ ln f ðyt j0; φt ; ν; γ Þ ðν þ 1Þy2t 1 ν þ 1XT
ðzt s þ mσ t Þ2
∇t ¼ ¼  − 2 ð34Þ − ln 1 þ
∂φ2t 2φ2t νφ2t γ 2signðyt Þ þ y2t 2φt 2 t¼1 ðν−2Þσ 2t γ 2signðzt sþmσ t Þ

Here we get ut not by ut = 2σt2∇t but by ut = 2φ2t ∇t, so All these AR-DCS models with skew Student distribution have the
same log likelihood, so we can estimate them by optimizing Eq. (41).
The asymptotic properties of maximum likelihood estimators for
ðν þ 1Þy2t another equivalent skew Student distribution had been investigated
ut ¼ 2φ2t ∇t ¼ −1 ð35Þ
νφ2t γ2signðyt Þ
þ y2t by Zhu and Galbraith (2010).
220 C.-T. Gao, X.-H. Zhou / Economic Modelling 53 (2016) 216–223

Table 1 are S&P 500 Index (SP500: 12/14/1993–6/9/2015), Germany DAX30

Descriptive statistics of daily return for stock indices. Index (DAX: 2/17/1992–6/9/2015), Hang Seng Index (HSI: 1/26/1998–
Index Mean Std. dev. Skewness Excess kurtosis J–B Q2(20) obs 6/9/2015) and Shanghai Composite Index (SCI: 6/6/1997–6/9/2015).
SP500 0.0720 0.7808 −0.6868* 9.4241* 3779.18⁎ 146.6⁎ 5409
Daily returns of these stock indices are calculated by the following
DAX 0.0368 0.9290 −0.0899* 1.3481* 77.07⁎ 60.6⁎ 5900 equation:
HSI 0.0181 2.0063 0.3142* 3.4668* 517.24⁎ 52.5⁎ 4284
SCI 0.0496 1.4952 −0.0991* 5.1001* 1085.44⁎ 214.3⁎ 4360
r t ¼ 100  ð ln P t − lnP t−1 Þ ð44Þ
Notes: * denotes significantly at the 1% level.

where rt and Pt denote daily return and closing price at time t respec-
3. Prediction methods of VaR and ES tively. The number of observations (obs) and the descriptive statistics
of the first 1000 obs for each stock index are shown in Table 1.
From the conditional mean equation of the above mentioned The J–B normality test statistics present that the distributions of
AR(1)–DCS(1,1) models, we may obtain one-day-ahead conditional these return series reject the hypothesis of normality. Significantly, all
mean μ^ tþ1 ¼ λrt . From the conditional variance equation of them, the values of excess kurtosis are bigger than 0, which indicate that
2
^ tþ1 . With μ^ tþ1
we may obtain one-day-ahead conditional variance σ the distributions of these return series have thicker tails in contrast to
2
^ tþ1, VaR and ES at the time t + 1 can be estimated in the following normal distribution. The values of skewness indicate that SP500, DAX
and σ
and SCI have significant negative skewness except HSI.
way.
4.2. Estimation results of AR–DCS models
3.1. The computation of VaR
We use eight AR–DCS models to estimate the four indices one by
Assuming the distribution of asset returns is standardized skew
one. For simplicity, we only present the estimated values of SP500.
Student distribution skt, we can calculate the one-day-ahead VaR by
The estimation results for the first 1000 obs in SP500 return series are
the equation below
listed in Table 2.
From Table 2, we can find that γ in all AR–DCS models is less than 1,
VaRtþ1 ðpÞ ¼ μ^ tþ1 þ σ
^ tþ1 qskt p;ν;γ ð42Þ
which indicates negative skewness. The values of Q2(20) statistic show
that all the skew Student AR(1)–DCS(1,1) models effectively depict the
where qsktp,ν,γ is the left-tailed quantile at p of skt with v degrees of
heteroscedasticity in the return series. The values of κ are greater than 0,
freedom and γ degrees of skewness. μ^ tþ1 and σ ^ tþ1 are the estimated
which shows that the negative shock will increase volatility.
values of conditional mean and volatility at time t + 1.
4.3. The backtesting results of VaR and ES
3.2. The computation of ES
In this section, we examine the one-day-ahead VaR and ES results for
By estimating these AR(1)–DCS(1,1) models, we can obtain the con-
the index SP500, DAX, HSI and SCI at the confidence levels 95%, 99% and
ditional mean μ^ tþ1 and volatility σ
^ tþ1 at next period. Meanwhile, the
99.5% (i.e., p is 5%, 1% and 0.5%). Using a rolling window size 1000, we
parameters v and γ will be figured out. By definition, we can calculate obtain 4409, 4900, 3284 and 3360 one-day-ahead forecasts for each
the one-day-ahead ES with the following equation: model and at confidence level, respectively for the four indices.
Z VaRtþ1
x f ðxÞdx 4.3.1. Accuracy evaluation for VaR
EStþ1 ðpÞ ¼ EðXjXbVaRtþ1 Þ ¼ −∞
ð43Þ If a realized loss is greater than the VaR at day t, it means that a
p failure happens. Following these rules, we can define hit series as

Where f(x) is the density function of nonstandardized skt with v degrees

of freedom and γ degrees of skewness, and x have mean μ^ tþ1 and 1 yt b VaRt
hit t ¼ ð45Þ
^ tþ1 .
volatility σ 0 yt ≥ VaRt

4. Empirical findings
When we set the length of test sample is n and the number of viola-
4.1. Data and descriptive statistics ^ ¼ n1 =n. From
tions in hit series is n1, we can work out the failure rate π
the definition of VaR, it is necessary to check the relation between the
In this paper, we choose four international stock market indices as realized failure rate and the expected failure rate. The unconditional
sample and collect the data from Wind Financial database. These indices coverage test proposed by Kupiec (1995) is the mostly used method.

Table 2
The in sample estimation results of SP500.

λ ω α κ β v γ LL Q2(20)

AR–DCS-A 0.0276 0.0024 0.0534 – 0.9982 4.9879 0.9194 −1049.3130 27.1726

AR–EDCS-A 0.0270 −0.0002 0.0483 – 0.9972 4.9362 0.9188 −1049.6760 26.6528
AR-DCSL-A 0.0532 0.0168 0.0249 0.1861 0.9006 5.9635 0.9006 −1040.1620 9.4248
AR-EDCSL-A 0.0558 −0.0890 0.0254 0.1606 0.9611 5.8439 0.8986 −1040.7870 8.5394
AR–DCS-B 0.0268 0.0041 0.0652 – 0.9968 5.0456 0.9021 −1048.1140 22.5129
AR–EDCS-B 0.0272 −0.0013 0.0618 – 0.9935 5.007 0.9031 −1048.5500 21.8571
AR-DCSL-B 0.0536 0.0170 0.0285 0.1648 0.8989 6.0366 0.9049 −1040.2440 9.5799
AR–EDCSL-B 0.0565 −0.0916 0.0287 0.1437 0.9606 5.9368 0.9014 −1040.7660 8.7276

Notes: LL is maximum likelihood estimation value of each model, Q2(20) is the Ljung–Box Q-statistic of order 20 computed on the squared standardized residuals.
 C.-T. Gao, X.-H. Zhou / Economic Modelling 53 (2016) 216–223 221

Table 3
The failure rates and unconditional coverage test results.

MODEL SP500 DAX HSI SCI

FR LRuc FR LRuc FR LRuc FR LRuc

95% confidence level

AR–DCS-A 0.0540 0.2309 0.0539 0.2185 0.0493 0.8599 0.0563 0.1029
AR–EDCS-A 0.0547 0.1615 0.0555 0.0818 0.0490 0.7972 0.0560 0.1200
AR–DCSL-A 0.0515 0.6523 0.0529 0.3630 0.0493 0.8599 0.0521 0.5820
AR–EDCSL-A 0.0531 0.3537 0.0535 0.2703 0.0505 0.8856 0.0518 0.6367
AR–DCS-B 0.0497 0.9201 0.0508 0.7937 0.0463 0.3228 0.0512 0.7524
AR–EDCS-B 0.0499 0.9752 0.0518 0.5575 0.0481 0.6175 0.0524 0.5296
AR–DCSL-B 0.0501 0.9697 0.0522 0.4740 0.0487 0.7356 0.0518 0.6367
AR–EDCSL-B 0.0524 0.4693 0.0520 0.5148 0.0502 0.9490 0.0515 0.6936

99% confidence level

AR–DCS-A 0.0093 0.6360 0.0094 0.6634 0.0107 0.7078 0.0089 0.5249
AR–EDCS-A 0.0098 0.8684 0.0096 0.7725 0.0116 0.3773 0.0095 0.7798
AR–DCSL-A 0.0086 0.3451 0.0086 0.3029 0.0094 0.7446 0.0077 0.1700
AR–EDCSL-A 0.0082 0.2060 0.0088 0.3790 0.0097 0.8824 0.0095 0.7798
AR–DCS-B 0.0082 0.2060 0.0082 0.1820 0.0091 0.6132 0.0086 0.4141
AR–EDCS-B 0.0084 0.2697 0.0082 0.1820 0.0100 0.9776 0.0080 0.2361
AR–DCSL-B 0.0091 0.5294 0.0086 0.3029 0.0094 0.7446 0.0074 0.1183
AR–EDCSL-B 0.0084 0.2697 0.0090 0.4651 0.0094 0.7446 0.0092 0.6478

99.5% confidence level

AR–DCS-A 0.0057 0.5368 0.0045 0.6063 0.0046 0.7214 0.0051 0.9611
AR–EDCS-A 0.0059 0.4116 0.0053 0.7636 0.0046 0.7214 0.0054 0.7717
AR–DCSL-A 0.0045 0.6573 0.0053 0.7636 0.0037 0.2506 0.0054 0.7717
AR–EDCSL-A 0.0045 0.6573 0.0051 0.9196 0.0040 0.3798 0.0057 0.5982
AR–DCS-B 0.0043 0.5054 0.0042 0.4132 0.0037 0.2506 0.0051 0.9611
AR–EDCS-B 0.0048 0.8220 0.0043 0.4674 0.0043 0.5390 0.0051 0.9611
AR–DCSL-B 0.0050 0.9923 0.0049 0.9191 0.0037 0.2506 0.0051 0.9611
AR–EDCSL-B 0.0043 0.5054 0.0053 0.7636 0.0037 0.2506 0.0057 0.5982

The likelihood ratio statistic LRuc in this test behaves as chi-square Where p is the expected failure rate. n0 indicates that the number of
distribution with freedom 1. hit = 0. The failure rates (FR) and p-values of unconditional coverage
test are listed in Table 3.


n1  Only doing unconditional coverage test will ignore the violation
LRuc ¼ −2 ln pn1 ð1−pÞn0 þ 2 ln π ^ Þn0 ∼χ 2 ð1Þ
^ ð1−π ð46Þ cluster which may cause extreme losses. To address this problem,

Table 4
The independent test and conditional coverage test results.

MODEL SP500 DAX HSI SCI

LRind LRcc LRind LRcc LRind LRcc LRind LRcc

95% confidence level

AR–DCS-A 0.3837 0.3339 0.8297 0.4582 0.9982 0.9845 0.6630 0.2405
AR–EDCS-A 0.3357 0.2362 0.8083 0.2136 0.7330 0.9128 0.8636 0.2942
AR–DCSL-A 0.1207 0.2709 0.8421 0.6482 0.7061 0.9170 0.4423 0.6398
AR–EDCSL-A 0.1590 0.2412 0.3788 0.3698 0.8854 0.9795 0.4633 0.6835
AR–DCS-B 0.7761 0.9555 0.9191 0.9614 0.6746 0.5617 0.6778 0.8728
AR–EDCS-B 0.7523 0.9510 0.8109 0.8182 0.8158 0.8591 0.6647 0.7471
AR–DCSL-B 0.1689 0.3879 0.6864 0.7133 0.7603 0.9016 0.4633 0.6835
AR–EDCSL-B 0.1870 0.3223 0.4988 0.6435 0.9143 0.9922 0.4849 0.7250

99% confidence level

AR–DCS-A 0.3803 0.6084 0.4554 0.6885 0.3851 0.6393 0.2720 0.4470
AR–EDCS-A 0.3573 0.6458 0.4761 0.7440 0.3455 0.4340 0.3134 0.5785
AR–DCSL-A 0.4163 0.4602 0.3763 0.3977 0.4420 0.7057 0.1977 0.1702
AR–EDCSL-A 0.4413 0.3341 0.3955 0.4733 0.4274 0.7219 0.3134 0.5785
AR–DCS-B 0.4413 0.3341 0.3390 0.2598 0.4570 0.6674 0.2524 0.3721
AR–EDCS-B 0.4287 0.3976 0.3390 0.2598 0.4130 0.7150 0.2152 0.2299
AR–DCSL-B 0.3920 0.5689 0.3763 0.3977 0.4420 0.7057 0.1810 0.1207
AR–EDCSL-B 0.4287 0.3976 0.4151 0.5494 0.4420 0.7057 0.2924 0.5176

99.5% confidence level

AR–DCS-A 0.5933 0.7165 0.0884 0.2052 0.7106 0.8760 0.0746 0.2038
AR–EDCS-A 0.5786 0.6118 0.1298 0.3033 0.7106 0.8760 0.0852 0.2179
AR–DCSL-A 0.6694 0.8273 0.1298 0.3033 0.7667 0.4947 0.0852 0.2179
AR–EDCSL-A 0.6694 0.8273 0.1186 0.2944 0.7478 0.6458 0.0965 0.2188
AR–DCS-B 0.6850 0.7378 0.0710 0.1258 0.7667 0.4947 0.0746 0.2038
AR–EDCS-B 0.6539 0.8818 0.0795 0.1649 0.7291 0.7798 0.0746 0.2038
AR–DCSL-B 0.6385 0.8955 0.1080 0.2735 0.7667 0.4947 0.0746 0.2038
AR–EDCSL-B 0.6850 0.7378 0.1298 0.3033 0.7667 0.4947 0.0965 0.2188
222 C.-T. Gao, X.-H. Zhou / Economic Modelling 53 (2016) 216–223

many researchers advise to do independent test. Christoffersen (1998) 5. Conclusion

proposed the independent test statistic LRind to test whether there is
temporal dependence in the violation series. He also proposed a condi- In this paper, we construct DCS models with skew Student dis-
tional coverage test statistic LRcc which is a joint test statistic of LRuc and tribution based on two methods. These DCS models include
LRind. LRind and the relation of LRcc with the other two are given as AR(1) conditional mean equation and have basic form and exponential
follows form, with leverage form and without leverage form. We examine their

n01 þ n11  performance on calculating VaR and ES by daily returns of four stock in-
LRind ¼ −2 ln π^ ð1−π ^ Þn00 þ n10 dices. Via backtesting, we find that all these AR–DCS models perform

n01 n00 n11 
^ 01 ð1−π
þ 2 ln π ^ 01 Þ π ^ 11 ð1−π ^ 11 Þn10 ∼χ 2 ð1Þ ð47Þ well on forecasting VaR and ES. So, these models are good choices for
measuring market risk and they provide more alternatives for
LRcc ¼ LRuc þ LRind ∼χ 2 ð2Þ ð48Þ managers.

where nij indicates the number of j following i and π ^ i j ¼ ni j =n. Acknowledgments

The p-values of LRind and LRcc are listed in the Table 4.
From the test results of Tables 3 and 4, we can observe that all the The authors are very grateful to anonymous reviewers and the
VaR series can pass the unconditional coverage test, independent test managing editor for their constructive comments and suggestions
and conditional coverage test. That is, all the skew Student distribution which help us to make the paper more perfect.
AR–DCS models based method A and method B perform well on
forcasting VaR at the confidence levels 95%, 99% and 99.5%. Appendix A. Derivation process of Eqs. (27) and (34)

4.3.2. Accuracy evaluation for ES A.1. For Eq. (27), the derivation process is as follows:
We use Eq. (43) and the estimated parameters to calculate the one-
day-ahead ES. To assess the performance, we use the evaluation method ∂ ln f ðzt j0; σ t ; ν; γ Þ
proposed by McNeil and Frey (2000). Their test is based on exceedance ∂σ 2t  
0 1
νþ1 !
residuals which are defined as: B
2Γ
2
s
C νþ1 ðzt s þ mσ t Þ2
∂ ln B@  C
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ν A− 2 ∂ ln 1 þ ðν−2Þσ 2 γ2signðzt sþmσ t Þ
^ tþ1 ð49Þ
fRtþ1 : t∈T; if r tþ1 bVaRtþ1 ðpÞg; where Rtþ1 ¼ ½r tþ1 −EStþ1 ðpÞ=σ ðγ þ γ−1 Þσ t ðν−2Þπ Γ
2
t

¼
∂σ t! 2

McNeil and Frey (2000) pointed out that these exceedance residuals νþ1 ðzt s þ mσ t Þ2 1
− ∂ ln 1 þ − ∂ lnσ 2t
will follow an iid with mean zero when ES is correctly estimated. They 2 ðν−2Þσ 2t γ 2signðzt sþmσ t Þ 2
¼
did not assume a distribution to {Rt + 1 : t ∈ T}, but use a bootstrap ∂σ 2t !
ðzt s þ mσ t Þ2
test to test the hypothesis of mean zero. The p-values which are tested !∂
2 2signðzt sþmσ t Þ ðν−2Þσ t γ 2 2signðz sþmσ Þ
νþ1 ðν−2Þσ t γ t t
1
by this method have been shown in Table 5. In this paper, we set the ¼− − 2
2 ðν−2Þσ 2t γ2signðzt sþmσ t Þ þ ðzt s þ mσ t Þ2 ∂σ 2t 2σ t
bootstrap samples as 10,000. !
ðzt s þ mσ t Þ2
From the results in Table 5, we find that all the p-values are greater !∂
νþ1 σt 2 σ 2t 1
than 0.05 which means that all the AR–DCS models perform well on ¼− − 2
2 ðν−2Þσ 2t γ2signðzt sþmσ t Þ þ ðzt s þ mσ t! Þ2 ∂σ 2t 2σ t
forecasting ES at the confidence levels 95%, 99% and 99.5%. νþ1 ðzt s þ mσ t Þ2 −mσ t ðzt s þ mσ t Þ 1
¼ − 2
2σ 2t ðν−2Þσ 2t γ 2signðzt sþmσ t Þ þ ðzt s þ mσ t Þ2 2σ t

Table 5
The backtesting results of ES forecasting.
A.2. For Eq. (34), the derivation process is as follows:
MODEL SP500 DAX HSI SCI

95% confidence level ∂ ln f ðyt j0; φt ; ν; γÞ

AR–DCS-A 0.5485 0.7565 0.6732 0.7354
∂φ2t  
AR–EDCS-A 0.5041 0.7445 0.5582 0.6967 0 1
νþ1
AR–DCSL-A 0.8646 0.9699 0.9501 0.6567 2Γ  
B 2 C νþ1 y2t
∂ ln B C−
AR–EDCSL-A 0.8926 0.9746 0.9454 0.5868
@   ∂ ln 1 þ
AR–DCS-B 0.7766 0.9763 0.7813 0.6987 pffiffiffiffiffiffi ν A 2 νφ2t γ 2signðyt Þ
AR–EDCS-B 0.7010 0.9716 0.8450 0.7057 ðγ þ γ −1 Þφt νπ Γ
2
AR–DCSL-B 0.7980 0.9721 0.9394 0.6559 ¼
∂φ2t
AR–EDCSL-B 0.8664 0.9616 0.9594 0.6241  
νþ1 2
yt 1
99% confidence level − ∂ ln 1 þ − ∂ ln φ2t
2 νφ2t γ2signðyt Þ 2
AR–DCS-A 0.1820 0.2512 0.8528 0.2202 ¼
∂φ2t
AR–EDCS-A 0.1659 0.2504 0.9038 0.2745  
AR–DCSL-A 0.5390 0.2333 0.8862 0.1077 y2t
  ∂
AR–EDCSL-A 0.4884 0.3126 0.8948 0.3542 νþ1 νφ2t γ2signðyt Þ νφ2t γ2signðyt Þ 1
AR–DCS-B 0.2590 0.3558 0.7992 0.3230 ¼− − 2
2 νφ2t γ 2signðyt Þ þ y2t ∂φ2 2φt
AR–EDCS-B 0.2573 0.2786 0.9317 0.1689  2 t
AR–DCSL-B 0.5971 0.2343 0.9109 0.0677 y
  ∂ t2
AR–EDCSL-B 0.5092 0.3496 0.8938 0.3296 νþ1 φ2t φt 1
¼− − 2
99.5% confidence level 2 νφ2t γ 2signðyt Þ þ y2t ∂φ2t 2φt
AR–DCS-A 0.4201 0.1334 0.7628 0.3215 ðν þ 1Þy2t 1
¼  2 − 2
AR–EDCS-A 0.3970 0.1692 0.7618 0.3713 2φt νφt γ 2signðyt Þ þ y2t
2 2φt
AR–DCSL-A 0.7700 0.5513 0.8114 0.4934
AR–EDCSL-A 0.7268 0.3525 0.7759 0.5832
AR–DCS-B 0.5806 0.3689 0.7896 0.4724
AR–EDCS-B 0.6793 0.3217 0.9201 0.4161 References
AR–DCSL-B 0.8303 0.4708 0.8333 0.3972
Artzner, P., Delbaen, F., Eber, J., Heath, D., 1999. Coherent measures of risk. Math. Financ. 9
AR–EDCSL-B 0.5920 0.5150 0.7491 0.6022
(3), 203–228.
 C.-T. Gao, X.-H. Zhou / Economic Modelling 53 (2016) 216–223 223

Bernardi, M., Catania, L., 2015. Comparison of value-at-risk models: the MCS package. Jarque, C.M., Bera, A.K., 1987. A test for normality of observations and regression residuals.
http://arxiv.org/abs/1502.04472. Int. Stat. Rev. 55, 163–172.
Bollerslev, T., 1987. A conditionally heteroskedastic time series model for speculative Kuester, K., Mittnik, S., Paolella, M.S., 2006. Value-at-risk prediction: a comparison of
prices and rates of return. Rev. Econ. Stat. 69, 542–547. alternative strategies. J. Financ. Econ. 4, 53–89.
Christoffersen, P., 1998. Evaluating interval forecasting. Int. Econ. Rev. 39, 841–862. Kupiec, P., 1995. Techniques for verifying the accuracy of risk measurement models.
Creal, D., Koopman, S.J., Lucas, A., 2012. Univariate generalized autoregressive score J. Deriv. 2, 174–184.
volatility models. http://faculty.chicagobooth.edu/drew.creal/. Lambert, P., Laurent, S., 2001. Modeling skewness dynamics in series of financial data
Creal, D., Koopman, S.J., Lucas, A., 2013. Generalized autoregressive score models with using skewed location-scale distributions. Stat Discussion Paper-0119. Institut de
applications. J. Appl. Econ. 28 (5), 777–795. Statistique, Louvain-la-Neuve.
Diamandis, P.F., Drakos, A.A., Kouretas, G.P., Zarangas, L., 2011. Value-at-risk for long and Lucas, A., Zhang, X., 2014. Score driven exponentially weighted moving average and
short trading positions: evidence from developed and emerging equity markets. Int. value-at-risk forecasting. Duisenberg school of finance-Tinbergen Institute Discussion
Rev. Financ. Anal. 20 (3), 165–176. Paper, TI 14-092/IV/DSF77.
Fernández, C., Steel, M., 1998. On Bayesian modelling of fat tails and skewness. J. Am. Stat. McNeil, A., Frey, R., 2000. Estimation of tail-related risk measures for heteroscedastic fi-
Assoc. 93, 359–371. nancial time series: an extreme value approach. J. Empir. Financ. 7 (3–4), 271–300.
Giot, P., Laurent, S., 2003. Value at risk for long and short trading positions. J. Appl. Econ. Theodossiou, P., 1998. Financial data and the skewed generalized t distribution. Manag.
18, 641–664. Sci. 44, 1650–1661.
Glosten, L.R., Jagannathan, R., Runkle, D., 1993. On the relation between the expected Zhu, D., Galbraith, J.W., 2010. A generalized asymmetric student-t distribution with
value and the volatility of the nominal excess return on stocks. J. Financ. 48, application to financial econometrics. J. Econ. 157, 297–305.
1779–1801. Zhu, D., Galbraith, J.W., 2011. Modeling and forecasting expected shortfall with the gen-
Hansen, B., 1994. Autoregressive conditional density estimation. Int. Econ. Rev. 35, eralized asymmetric Student-t and asymmetric exponential power distributions.
705–730. J. Empir. Financ. 18 (4), 765–778.
Harvey, A.C., 2013. Dynamic Modes for Volatility and Heavy Tails. Cambridge University
Press.
Harvey, A.C., Sucarrat, G., 2014. EGARCH models with fat tails, skewness and leverage.
Comput. Stat. Data Anal. 26, 320–338.