You are on page 1of 16

Energy Economics 30 (2008) 3156–3171

Contents lists available at ScienceDirect

Energy Economics
j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / e n e c o

Estimating ‘Value at Risk’ of crude oil price and its spillover


effect using the GED-GARCH approach
Ying Fan a, Yue-Jun Zhang a,b, Hsien-Tang Tsai c, Yi-Ming Wei a,⁎
a
Center for Energy and Environmental Policy Research, Institute of Policy and Management,
Chinese Academy of Sciences, Beijing, 100080, China
b
Graduate University of the Chinese Academy of Sciences, Beijing, 100080, China
c
College of Management, National Sun Yat-sen University, Kaohsiung 80424, Taiwan

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

Article history: Estimation has been carried out using GARCH-type models, based on
Received 2 January 2007 the Generalized Error Distribution (GED), for both the extreme
Received in revised form 2 April 2008 downside and upside Value-at-Risks (VaR) of returns in the WTI and
Accepted 3 April 2008
Brent crude oil spot markets. Furthermore, according to a new concept
Available online 11 April 2008
of Granger causality in risk, a kernel-based test is proposed to detect
extreme risk spillover effect between the two oil markets. Results of
Keywords:
an empirical study indicate that the GED-GARCH-based VaR approach
International crude oil markets
appears more effective than the well-recognized HSAF (i.e. historical
GED-GARCH models
Value-at-Risk (VaR) simulation with ARMA forecasts). Moreover, this approach is also
Granger causality in risk more realistic and comprehensive than the standard normal
Risk spillover effect distribution-based VaR model that is commonly used. Results reveal
that there is significant two-way risk spillover effect between WTI
and Brent markets. Supplementary study indicates that at the 99%
confidence level, when negative market news arises that brings about
a slump in oil price return, historical information on risk in the WTI
market helps to forecast the Brent market. Conversely, it is not the case
when positive news occurs and returns rise. Historical information on
risk in the two markets can facilitate forecasts of future extreme
market risks for each other. These results are valuable for anyone who
needs evaluation and forecasts of the risk situation in international
crude oil markets.
© 2008 Elsevier B.V. All rights reserved.

⁎ Corresponding author. Center for Energy and Environmental Policy Research (CEEP), Institute of Policy and Management (IPM),
Chinese Academy of Sciences (CAS), P.O. Box 8712, Beijing 100080, China. Tel./fax: +86 10 62650861.
E-mail address: ymwei@deas.harvard.edu (Y.-M. Wei).

0140-9883/$ – see front matter © 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.eneco.2008.04.002
Y. Fan et al. / Energy Economics 30 (2008) 3156–3171 3157

1. Introduction

Since the mid-1980s, international crude oil price has been primarily influenced by the supply and
demand condition in oil markets but not as rigidly controlled as that during the first and second oil crises.
However, the market mechanism made the international oil price frequently volatile with extensive
amplitude. It can be found that there are significant extreme market risks that have led oil market
participants and governments to endure heavy potential losses. As a result, there has been much interest in
the oil market risk, and its measurement.
To date, numerous literatures regarding oil market risk have discussed numerous topics from a
qualitative perspective. For instance, what kinds of risk there are, what impact they have and what we
should respond to them, are typical exploratory questions (Huang et al., 2005; Li et al., 2005). Clearly, it is
important for us to know the characteristics of oil market risks. Yet, quantitative literatures are extremely
scant.
To measure market risk, application of the Value-at-Risk (VaR) methodology offers comprehensive and
recapitulative advantages. It depicts market risk by means of the probability distribution of a random
variable and evaluates the risk with a single real number. As a result, VaR has become an essential tool
within financial markets (Fan, 2000; Ma et al., 2001; Tan and Chan, 2003; Fan et al., 2004; Hartz et al.,
2006). VaR is also applicable in oil markets to measure risk, and several authors have accomplished
commendable research using VaR. Cabedo and Moya (2003) presented three kinds of VaR methods (i.e. the
historical simulation standard approach, the historical simulation with ARMA forecasts (HSAF) approach
and the variance–covariance method based on autoregressive conditional heteroskedasticity (ARCH)
models forecasts) to calculate the risk of Brent spot oil price with the assumption of standard normal
distribution for oil price returns. Their results indicated that HSAF provides the most flexible and efficient
risk quantification. Giot and Laurent (2003) assessed the performance of the RiskMetrics, skewed Student
APARCH and skewed Student ARCH models when used to calculate the market risk of several commodities,
such as Brent and WTI crude oil. They identified that although the skewed Student APARCH model
performed best in all cases, the skewed Student ARCH model delivered excellent results and was relatively
easy to use. Feng et al. (2004) proposed a semi-parametric VaR method to calculate the risk for losses in
WTI crude oil market. Their results suggested that the new method could largely improve risk
measurement precision when the confidence level was 99%. Sadeghi and Shavvalpour (2006) introduced
the HSAF and variance–covariance method based on the GARCH model to estimate the VaR of OPEC oil
price. They also found that the HSAF could be more efficient. Based on the HSAF and EWMA, Fan and Jiao
(2006) proposed an improved historical simulation approach. They used an exponential decreased
frequency approach with the ARMA forecast (EDFAAF) to estimate the VaR of Brent spot oil price. By
comparing it with the HSAF approach, they offered evidence to show that EDFAAF had more effective
forecasting power for oil risk management. Sadorsky (2006) presented several different univariate and
multivariate statistical models to estimate forecasts of daily volatility in petroleum futures price returns.
Results showed that non-parametric models outperformed parametric models in terms of the number of
exceedances in the backtest.
When the VaR method is used to measure extreme market risk, the risk interaction and spillover effect
among different markets is apparent. It is not uncommon to judge whether historic information about
extreme risk in one market helps to forecast the current or future risk in another market. If this is so, then
we say the former has exerted risk spillover effect on the latter. To test this hypothesis, Hong (2001)
introduced the concept of Granger causality in risk. Based on the new concept, Hong et al. (2003) proposed
a class of kernel-based tests to specifically examine and analyze the spillover of extreme downside market
risk among share A, B and H in the Chinese stock market, and that between Chinese stock market and
oversea equity markets. In fact, the idea provided by Hong et al. (2003) does not just apply to stock markets
but also is helpful in detecting the risk spillover effect among oil markets. Consequently, it is the concept
that is the focus of this paper. Lin and Tamvakis (2001) first investigated the spillover effect among oil
markets, but they were concerned with price information transmission mechanism but not the risk
information (as in this paper).
From the existing literatures which have the VaR method to measure risks in oil markets, it should be
noted that most are solely focused on the market risk when oil price returns go down (i.e. downside risk),
which is similar to research on financial markets. However, crude oil, as a special commodity, has its own
3158 Y. Fan et al. / Energy Economics 30 (2008) 3156–3171

traits that differ from financial products, and so does its market risk. Specifically, its market risk has quite
distinct implications for disparate market participants. For example, when crude oil price returns slump,
crude oil producers or sellers incur losses. In this scenario, profits grow among purchasers. And when the
returns rise, the outcomes will reverse. Therefore, in contrast to the risk measurement in financial markets,
both the downside and upside risks have to be considered in crude oil markets. However, there are still
shortcomings in existing VaR methods (such as the assumption of standard normal distribution). Hence,
exploration should be conducted for appropriate VaR methods, and convincing empirical research.
Furthermore, with globalization of the economy, it is likely that market risks will transfer from one market
to another. Accordingly, we should also investigate the risk spillover effect among various international
crude oil markets besides the risk in a single oil market. This work would be of great significance to identify
future trends of oil price returns, and thus help to support the scientific decision making of those
governmental departments and enterprises concerned.
In consequence, the following three activities extend and develop extant work. First, we estimate both
the downside and upside extreme risks of international crude oil price returns. Second, we introduce
GARCH type models based on generalized error distribution (GED) to develop the VaR model with respect
to the international crude oil price returns. When the VaR estimation was made by means of GARCH type
models previously, most literatures assumed that the residual series follow the standard normal or student
distribution. However, oil price returns usually have a leptokurtic distribution with a fat tail, which is quite
different from the assumption. As a result, the estimated VaR model proves to be inappropriate and
inadequate, which subsequently affects the judgment of the oil market risk situation. In contrast, GED offers
a comprehensive distribution, and turns out to be advantageous when used here. Meanwhile, in order to
evaluate objectively whether the VaR model is adequate or not, a statistical backtest method provided by
Kupiec (1995) is used. Finally, we use the method of Granger causality in risk developed by Hong (2001)
and Hong et al. (2003) to examine the risk spillover effect between WTI and Brent crude oil markets.
There are four major sections in this paper. The models and econometric methodologies used are
presented in Section 2. Section 3 presents the data and empirical results and discussions, while Section 4
provides some concluding remarks.

2. Methodologies of risk measurement and spillover effect

2.1. GARCH-VaR model based on GED

When modeling crude oil price return, often a clustering phenomenon of return volatility is evident. In
order to describe it, Engle (1982) developed the standard ARCH model. When the lag of ARCH models
became too large, Bollerslev (1986) proposed adopting the generalized ARCH, known as the GARCH model.
The GARCH model can be expressed as follows:

Yt ¼ X tVb þ et ;

X
p X
q ð1Þ
ht ¼ a0 þ ai e2ti þ bj htj ;
i¼1 j¼1

where Yt denotes the oil price return in this paper, Xt is a column vector consisting of independent variables,
and β is the coefficient column vector. There are also some other constraints, such as p N 0, q ≥ 0, α0 N 0, αi ≥ 0
P
p P
q
(i = 1, 2,…, p), βj ≥ 0 (j = 1, 2,…, q), and ai þ bj b1 which reflects the duration of return volatility.
i¼1 j¼1
Meanwhile, significant leverage effect often can be seen in the volatility of oil price returns, which tells
that the current volatile degree of returns caused by the previous oil price return increase and decrease is
quite asymmetric. The TGARCH (Zakoian, 1994) model is applied to discuss this topic, and its conditional
variance can be depicted as follows:

X
p X
q
ht ¼ a0 þ ai e2ti þ we2t1 dt1 þ bj htj ð2Þ
i¼1 j¼1
Y. Fan et al. / Energy Economics 30 (2008) 3156–3171 3159

where if ɛt − 1b0, dt − 1= 1; otherwise, dt − 1= 0. Due to the use of dt − 1, the influence of return increase (ɛt − 1N0)
and decrease (ɛt − 1b0) on the conditional variance (ht) can be found distinct. Specifically, when the return
goes up, ψɛt − 12dt − 1= 0, and the influence can be expressed by P
p
ai ; whereas if the return goes down, the
i¼1
P
p
influence should be ai þ w. In brief, as long as ψ ≠ 0 in the model, then it can be said that the influence of
i¼1
information concerned appears asymmetric.
Based on the research on volatility clustering and leverage effect, crucial importance should be attached
to the calculation and supervision of the extreme risk in oil markets, in which the key point is how to
measure the risk quantitatively. Therefore, the concise and effective methodology VaR is introduced here.
According to Hendricks (1996) and Hilton (2003), the idea of VaR indicates the maximum amount of
money that may be lost on a portfolio over a given period of time with a given confidence level, due to
exposure to the market risk. From the perspective of statistics, VaR implies the quantile of the
distribution function. In this paper, the left quantile of the international oil price returns is used to
measure the downside risk, which means the loss of sales income for crude oil producers owing to the
fall of oil returns; whereas the right quantile is adopted to depict the upside risk, which represents the
extra expenses for oil purchasers caused by the sharp rise of oil price. In this way, the method focused on
the downside extreme risk in financial markets is considered. On the other hand, the special
characteristics of crude oil markets are recognized here. This helps to detect extreme risks in oil
markets more accurately.
In reality, there are numerous methods to calculate the VaR. Generally, they can be classified as three
types, i.e., variance–covariance method, historical simulation approach and Monte Carlo simulation
method. Here, the first one is used. During the calculation, it is very important to estimate the parameters of
VaR models. In fact, the most common estimating methods include the GARCH model and RiskMetrics
methodology. The latter usually assumes that the return series have a standard normal distribution, thus it
is hard to describe some characteristics of return volatility (e.g. leverage effect). Relatively, the former has
gained more popularity recently, whereas it is also based on the assumption that the residual asset returns
follow a normal distribution (Barone-Adesi et al., 1999). However, because the oil price return often has a
leptokurtic distribution and fat tail, which are not in accordance with the standard normal distribution, this
results in the fact that the assumption of standard normal distribution tends to underestimate the extreme
risk. For this purpose, the generalized error distribution (GED) (Nelson, 1990) is presented here to estimate
the residual series from GARCH type models. The probability density function of GED can be formulated as
follows:
 
k4exp  12 je=kjk
f ðeÞ ¼ ð0VkVlÞ; ð3Þ
k42½ðkþ1Þ=k Cð1=kÞ

h ð2=kÞ
i1=2
Cð1=kÞ
where k ¼ 2 Cð3=k Þ
; Cð●Þ is the gamma function, and k is the parameter of GED and also called its degree
of freedom, which implies how fat the tail is. Specifically, k = 2 indicates the GED turns to be the standard
normal distribution; k N 2 shows its tail is thinner than that of the standard normal distribution; while k b 2
suggests its tail is thicker.
On the basis of the definition of VaR above, we can obtain the following Eqs. (4) and (5) for VaR
calculation, which are used for the upside VaR and downside VaR in oil markets respectively.
qffiffiffiffiffiffiffiffiffi
VaRup
m;t ¼ Am;t  zm;a hm;t ; ðm ¼ 1; 2Þ ð4Þ

qffiffiffiffiffiffiffiffiffi
VaRdown
m;t ¼ Am;t þ zm;a hm;t ; ðm ¼ 1; 2Þ ð5Þ

where μm,t is the conditional expectation return of market m; zm,α b 0 denotes the left α-quantile of the GED
distribution which is used for the residual series of (T) GARCH model in market m; and hm,t is the conditional
variance series in market m, so hm,t N 0.
After the calculation of VaRs, for the sake of reliability, it is necessary to backtest whether the VaR model
used has adequately estimated the real extreme risk or not. To this end, the backtest method provided by
3160 Y. Fan et al. / Energy Economics 30 (2008) 3156–3171

Kupiec (1995) is employed here. Its main idea can be summarized as that if we assume the confidence level
1 − α, sample size T, days of failure N, then the frequency of failure can be regarded as f = N / T. Subsequently,
Kupiec (1995) proposed the most proper likelihood ratio test with the null hypothesis f = α, where the
statistic
h i h i
TN N
LR ¼ 2ln ð1  f Þ f  2ln ð1  aÞTN aN : ð6Þ

Under the null hypothesis, LR ~ χ2(1), and the critical values of its 95% and 99% confidence level are 3.84
and 6.64 respectively. According to the definition of χ2 distribution, if the value of LR is larger than the
corresponding critical value, then the null hypothesis should be rejected, in other words, it can be said that
the VaR model is not adequate.

2.2. Risk spillover effect test

The methodology used is Granger causality in risk provided by Hong (2001) and Hong et al. (2003) to
test the risk spillover effect between WTI and Brent crude oil markets. It requires that the time-varying VaR
should be calculated for each return, and then whether the historical information about risk in one market
increases one's ability to forecast its occurrence in another market should be determined in terms of the
Granger causality concept.
First of all, an indicator function of VaR is defined as follows, taking downside VaR indicator function
based on VaR series as an example:
 
Zm;t ¼ I Ym;t b  VaRm;t ; ðm ¼ 1; 2Þ ð7Þ

where I(●) is the indicator function. When the actual loss exceeds VaR, Zm,t takes value 1; otherwise 0.
To test the one-way downside risk spillover effect from market 2 to market 1, the null hypothesis can be
stated as H0: E(Z1,t|I1,t − 1) = E(Z1,t|It − 1), whereas HA: E(Z1,t|I1,t − 1) ≠ E(Z1,t|It − 1), in which It − 1= {Ym,t − 1, Ym,t − 2,…}
indicates the information set available at time t − 1 and I1,t − 1is the information set for market 1. If H0 holds,
that is, there is no one-way downside Granger causality in risk from market 2 to market 1, then when
extreme risk is found in market 2, it can not be used to forecast the probable risk in market 1 in the future.
Similarly, upside risk spillover effect test can be conducted.
Suppose VaRm,t = VaRm (Im,t − 1, α) is the VaR  series of market  m at the significance level α, which is
obtained from Eqs. (4) and (5). Let Ẑ m;t ¼ I Ym;t b  VaRm;t just as Eq. (7), then the sample cross-
correlation function (CCF) between Z 1,t and Z 2,t can be given as follows:
8
> XT   
>
> T 1 Ẑ 1;t  â1 Ẑ 2;tj  â2 ; 0VjVT 1
>
>
< t¼1þj
Ĉ ð jÞ ¼ ð8Þ
>
> X
T   
>
> 1
>
: T Ẑ 1;tþj  â1 Ẑ 2;t  â2 ; 1T Vjb0
t¼1j

P
T
where âm ¼ T 1 Ẑ m;t , m = 1,2, and T is the sample size of return series {Y1,t, Y2,t}t = 1T. And the sample
t¼1
cross-correlation is
q̂ð jÞ ¼ Ĉ ð jÞ=Ŝ1Ŝ2 ; j ¼ 0; F1;: : :; FðT  1Þ; ð9Þ
2  
where Ŝm ¼ âm 1  âm is the sample variance of Ẑ m;t .
Hong et al. (2003) proposed a class of kernel-based tests for Granger causality in risk. To test one-way
Granger causality in risk from market 2 to market 1, the test statistic is
8 9
< X T 1 =
2
Q1 ðMÞ ¼ T k ð j=MÞq̂ ð jÞ  C1T ðM Þ =f2D1T ðM Þg1=2
2
ð10Þ
: j¼1 ;
Y. Fan et al. / Energy Economics 30 (2008) 3156–3171 3161

where k(●) is a kernel function that assigns weights to various lags whereas M is the lag order, and the
centering factor and scaling factor can be expressed respectively as

X
T 1
C1T ðM Þ ¼ ð1  j=T Þk2 ð j=M Þ;
j¼1

X
T 2
D1T ðM Þ ¼ ð1  j=T Þf1  ð j þ 1Þ=T gk4 ð j=M Þ:
j¼1

Hong et al. (2003) also constructed tests for downside risk spillover effect between two markets,
including instantaneous downside risk spillovers. The test statistic is
8 9
< XT 1 =
2
Q2 ðMÞ ¼ T k ð j=M Þ q̂ ð jÞ  C2T ðM Þ =f2D2T ðMÞg1=2
2
ð11Þ
: ;
j¼1T

where the centering factor and scaling factor are specified respectively as

X
T 1
C2T ðM Þ ¼ ð1  j=T Þk2 ð j=M Þ;
j¼1T

X
T 2
D2T ðM Þ ¼ ð1  j=T Þf1  ð j þ 1Þ=T gk4 ð j=MÞ:
j¼2T

When H0 holds, then Q1(M) and Q 2(M) follow an asymptotic standard normal distribution, i.e., N(0,1).
Therefore, if Q1(M) and Q 2(M) are greater than the right-tailed critical value of N(0,1) distribution at a
specified significance level, then the null hypothesis is rejected.
In this paper, we extend the idea in Hong et al. (2003) and introduce a new concept according to the
specific situation in oil markets, that is, the Granger causality in upside risk.
In practice, we may test the two-way Granger causality in risk first. If its p-value is significant, we may
further check the one-way Granger causality in risk to discover the specific risk spillover direction.

3. Empirical analysis and discussion

3.1. Data

We use daily spot WTI and Brent crude oil prices from May 20th in 1987 to August 1st in 2006, which
are quoted in US dollars per barrel and come from the Energy Information Administration of America.

Fig. 1. Daily spot WTI and Brent crude oil prices (1987.5.20–2006.8.1).
3162 Y. Fan et al. / Energy Economics 30 (2008) 3156–3171

Fig. 2. Daily spot WTI and Brent crude oil price returns (1987.5.20–2006.8.1).

However, there are some missing prices in the original series. Considering the non-linear characteristics
of the oil prices (see Fig. 1), we interpolate the missing prices with the geometric average values of its
two fore-and-aft neighboring daily prices. As a result, 4944 samples are obtained. We divide the whole
time-period into two parts. The first time-period, which is called “in-sample period” from May 20th, 1987
to July 29th, 2005, is used to estimate VaR models. The second time-period, which is called “out-of-
sample period” from August 1st, 2005 to August 1st, 2006, is used to construct the out-of-sample
forecasts.

3.2. Summary statistics for WTI and Brent crude oil price returns

Suppose oil price series of WTI and Brent are P1,t and P2,t at date t respectively, then their logarithmicreturns
could be formulated as Y1,t =ln(P1,t /P1,t − 1) and Y2,t =ln(P2,t /P2,t − 1). In this way, 4943 daily return samplescould be
obtained and their trends could be seen from Fig. 2. It should be noted that both reveal volatility clustering.
First, we check the basic statistical characteristics of WTI and Brent oil price returns during the in-
sample period, and results are shown in Table 1. Generally, both the means and variances of the two returns
are in the neighborhood respectively, which can also be confirmed from Fig. 2. Meanwhile, compared with
the standard normal distribution with skewness 0 and kurtosis 3, the skewnesses of both returns here are
negative (i.e. left skewed) while kurtoses are far larger than 3, hence we may deduce that each return has a
leptokurtic distribution with a fat left tail. In other words, returns here do not have the standard normal
distribution. Verification is from the results of Jarque Bera (JB) test. In order to check the autocorrelation of
the returns, a Ljung Box (LB) test confirms each return, with results indicating that both have significant
autocorrelation. In addition, both returns are stationary series by means of ADF unit root test.

Table 1
Summary statistics for WTI and Brent returns during the in-sample period

Return Mean Variance Skewness Kurtosis JB test LB-Q(10) ADF test


WTI 0.00024 0.0246 −1.2322 24.0772 87869.30 (0.0000) 45.998 (0.000) −31.0462 (0.0000)
Brent 0.00025 0.0234 −0.9099 20.9386 63436.07 (0.0000) 25.617 (0.004) −32.3834 (0.0000)

Note: p-values for corresponding null hypotheses are reported in parentheses.


Y. Fan et al. / Energy Economics 30 (2008) 3156–3171 3163

3.3. Estimation of GARCH type models for WTI and Brent oil price returns

In order to filter out the autocorrelation of the WTI return, the ARMA model is used here. According to
the censored orders of autocorrelation and partial autocorrelation function graphs, as well as the principle
that AIC value must be relatively minimum, an ARMA(1,1) model is singled out through numerous trials.
Examining the result of the LB test for the residual series of the ARMA(1,1) model, it can be seen that all of
the autocorrelation coefficients fall within the given random interval and all of their absolute values have
no evident differences from 0. This indicates that the residual series is random. In other words, the original
WTI return series can be well fitted by the ARMA(1,1) model.
Whereas the WTI return series has significant volatility clustering, therefore, the ARCH LM test is carried
out for the residual series of the ARMA(1,1) model above. The result shows that there is high-order ARCH
effect; so a GARCH model needs to be adopted. Because of the fat tail of the return, the GED distribution is
introduced to depict the residual of the GARCH model. According to the requirements that the AIC value
should be relatively small, and model coefficients must be significant and positive, the GARCH(1,1) model is
the best when comparison are made among GARCH(1,1), GARCH(1,2), GARCH(2,1) and GARCH(2,2) models.
Next, TGARCH(1,1) and GARCH-M(1,1) models are developed so as to further investigate the volatility
characteristics of the WTI return. Empirical results imply that there are significant TGARCH and GARCH-M
effects. In other words, the volatility of the return appears asymmetric, and the return is markedly
influenced by its expected risk. According to the minimum AIC value and the need to describe the
asymmetric volatility, a TGARCH(1,1) model becomes the final choice for the WTI return, whose estimation
results can be seen from Table 2 below.
Similarly, an MA(1)-GARCH(1,1) model is singled out for the volatility of Brent return, whose parameter
estimation results are stated in Table 2.
Meanwhile, as can be seen from Table 2, GED degree parameters of the two models are 1.2608 and
1.3246 respectively, both less than 2, which confirms the fact that the tails of the international oil returns
are thicker than that of the standard normal distribution.
From the variance equation of WTI return, it can be found that when the return goes down, the impact
2
of ɛ1,t − 1 on h1,t is α1 + ψ, that is, 0.0572; while the return goes up, the impact should only be α1, that is,
0.0836, which is about 1.5 times larger than the former. The coefficient of h1,t − 1 is 0.9205, which means
92.05% of current variance shock can still be seen in the following period. Therefore, the decay of the
volatility shock is quite slow, and the volatility clustering is very evident. Investigation of the residual
series in the mean equation indicates that its autocorrelation function fluctuates within the given random
interval and the significance probabilities of Q and Q2 statistics are both larger than 10%. Therefore, we
can deduce that there are no longer autocorrelation and volatility clustering in the residual series.
Additionally, the result from ARCH-LM test also shows that the residual series no longer has volatility
clustering. All of this indicates that the TGARCH(1,1) model has fitted the WTI return very well.
On the other hand, according to the volatility model of Brent return, we find that there are no significant
asymmetric leverage effect and GARCH-M effect either. Moreover, In the variance equation of the GARCH

Table 2
Estimation results of TGARCH and GARCH models for oil price returns

Parameter WTI-TGARCH(1,1) Brent-GARCH(1,1)


Mean equation
AR(1) − 0.6667 (0.0441) –
MA(1) 0.6548 (0.0512) 0.0463 (0.0013)

Variance equation
α0 6.38E−06 (0.0000) 5.31E−06 (0.0000)
α1 0.0836 (0.0000) 0.0801 (0.0000)
β1 0.9205 (0.0000) 0.9127 (0.0000)
ψ − 0.0264 (0.0201) –
AIC value − 4.8986 − 4.9935
Log likelihood 11474.52 11697.19
GED degree 1.2608 (0.0000) 1.3246 (0.0000)

Note: p-values are reported in parentheses.


3164 Y. Fan et al. / Energy Economics 30 (2008) 3156–3171

Fig. 3. Conditional variances of WTI and Brent returns.

(1,1) model of Brent return, the coefficient of h2,t − 1 is 0.9127, which means 91.27% of the current
variance shock can be found in the next period. Hence, similar to the WTI return, the Brent return also has a
slow speed of decay for its volatility shock. Investigating the residual series of the mean equation, we find
that its autocorrelation function also falls into the given random interval, and both of its Q and Q2 statistics
have large significance probabilities (both are larger than 20%). In this way, we can say that the residual
series has no longer autocorrelation and volatility clustering, and the result from ARCH-LM test confirms
that it has no volatility clustering. Therefore, we can judge the goodness-of-fit of the GARCH(1,1) for the
Brent return is very nice.
As for the volatility extent comparison, Fig. 3 depicts the trends of the conditional variance series for
WTI and Brent returns. These represent their volatility extent respectively. Judging from the figure, we can
find that, the volatility levels of the two returns are close, even if the volatility of the WTI return is a little
larger on occasions; while the volatility of the Brent return is sharper during the Gulf War. Both of the
returns have an evident trait, namely that the largest variance during those periods with high volatility hits
more than twenty times greater than that on the average volatility level. This kind of large-scale vibration of
volatility not only demonstrates the extreme risk in the international crude oil market, but also significantly
helps to forecast the market volatility and the risk concerned.

3.4. VaR model estimation and test for WTI and Brent oil price returns

In accordance with prior discussion, it is necessary to measure both the upside and downside risk for the
oil price returns to support the scientific decisions of crude oil producers and purchasers concerned.
Therefore, we adopt the variance–covariance method to measure the upside and downside VaRs by means
of the TGARCH(1,1) and GARCH(1,1) models based on GED above for WTI and Brent returns respectively.
(1) Determination of GED quantiles
Based on the probability density function of GED, we calculate its 95% and 99% quantiles under the given
parameters (see Table 3). The results in the table show that its 95% quantile is close to that of the standard

Table 3
GED parameters and quantiles for WTI and Brent returns

Return Parameter 95% quantile 99% quantile


WTI 1.260823 1.642 2.578
Brent 1.324630 1.644 2.548
Y. Fan et al. / Energy Economics 30 (2008) 3156–3171 3165

Table 4
Summary of VaRs for the WTI return

Conf. level Risk type Mean Standard deviation Max Min Failure time Rate of Failure LR statistic
95% Upside 0.0375 0.0152 0.1796 0.0070 194 0.0414 7.6590
Downside −0.0375 0.0152 −0.0070 − 0.1796 221 0.0472 0.7857
99% Upside 0.0589 0.0239 0.2789 0.0149 41 0.0088 0.7628
Downside −0.0589 0.0239 −0.0149 − 0.2789 60 0.0128 3.4416

Table 5
Summary of VaRs for the Brent return

Conf. level Risk type Mean Standard deviation Max Min Failure time Rate of Failure LR statistic
95% Upside 0.0358 0.0154 0.1932 0.0107 194 0.0414 7.6770
Downside −0.0358 0.0154 − 0.0107 −0.1932 229 0.0489 0.1201
99% Upside 0.0555 0.0239 0.3016 0.0167 47 0.0100 0.0006
Downside −0.0555 0.0239 − 0.0167 −0.3016 51 0.0109 0.3645

normal distribution, 1.645, while the 99% quantile is evidently larger than that of the standard normal
distribution, 2.326, which also implies that international crude oil price returns have marked fat tails.
(2) Estimation of VaRs based on GED-GARCH models
Based on Eqs. (4) and (5), the upside and downside VaRs for WTI and Brent returns can be obtained at
the confidence levels 95% and 99%, and their summary statistics are shown in Tables 4 and 5.
Several findings can be drawn from the results in Tables 4 and 5:
• With the exception of the upside risk at the confidence level of 95%, all of other LR statistics are less than
their corresponding critical values. Therefore, according to the backtest method provided by Kupiec
(1995), it can be argued that the TGARCH(1,1) and GARCH(1,1) models have adequately estimated the
VaRs for the two returns. It also can be confirmed from the return trends and VaRs (see Fig. 4 with the
example of Brent return and its VaR at the confidence level of 99%).
• At the 99% confidence level, the estimation of upside VaRs for WTI and Brent returns is more precise than
that of the downside VaRs, which can be gotten from their LR statistic values comparison. This results

Fig. 4. Brent return and its VaRs at the confidence level of 99%.
3166 Y. Fan et al. / Energy Economics 30 (2008) 3156–3171

Fig. 5. Upside and downside VaRs of WTI and Brent returns at the 99% confidence level.

from the fact that the return distribution has a long left tail and GED has not totally caught this kind of fat-
tail characteristic. While at the 95% confidence level, the estimation of downside VaRs is more precise.
• If one considers the means of VaR series, it can be found that at identical confidence levels, no matter if
the returns go up or down, VaR means of the WTI return are larger than those of the Brent return. Thus,
more risk reserves should be prepared for the former. Examining the VaR trends in Fig. 5, it can seen that,
the two returns have similar VaR trends, although during some sample periods, the risk of the WTI return
may exceed that of the Brent return.

(3) Comparison of VaR models


When GARCH models are used for VaR calculation on oil price returns or stock price returns, it will be
assumed that their residual series follow the standard normal distribution for the sake of calculation
convenience, thus directly taking the quantile of the standard normal distribution as zm,α in Eqs. (4) and
(5). However, in reality, international oil price returns and their residual series of related models often do
not follow the standard normal distribution. As a result, the VaR model under the usual assumption will not
be adequate. In order to explain this point, we develop a VaR model based on the standard normal
distribution quantile at the confidence level of 99% and compare the results (see Table 6) with those in
Tables 4 and 5.
Comparison of results reveal that, on the one hand, VaR means based on the standard normal
distribution quantile are closer to 0 than those in Table 4 and 5. This trait can also be confirmed from the
comparison of failure times. On the other hand, because all of the failure times in Table 6 exceed the given
time (about 47) with the confidence level of 99%. Therefore, we may say the VaR model here has
underestimated the real market risk.

Table 6
Summary of VaRs based on the standard normal distribution quantile at the 99% confidence level

Return Risk type Mean Failure time Rate of Failure LR statistic


WTI Upside 0.0531 62 0.0132 4.5119
Downside − 0.0531 79 0.0169 18.5196
Brent Upside 0.0507 69 0.0147 9.2526
Downside − 0.0507 71 0.0152 10.8804
Y. Fan et al. / Energy Economics 30 (2008) 3156–3171 3167

According to the backtesting method of Kupiec (1995), the critical value of χ2 test at the 99% confidence
level is 6.64, hence, basically any directions and any returns do not see an adequate and reliable VaR model
under the assumption of standard normal distribution quantiles. Even if the upside VaR model for WTI
return does seem acceptable from the perspective of its LR value, whereas its counterpart in Table 4 appears
more precise due to a smaller LR value. Consequently, we may argue that overall it is more acceptable to use
VaR models based on GED than those based on the standard normal distribution, i.e., N(0,1).
Nevertheless, at the 95% confidence level, VaR models based on GED and standard normal distribution
have approximately the same LR values. This arises from their approximate quantiles, which are close to
1.645 (see Table 3).
Additionally, we also find that if we assume the residual series from GARCH and TGARCH models have
standard normal distributions and adopt their quantiles when VaR models are developed, then all of the
subsequent VaR models are inadequate for the risk at any directions and in any markets. Interestingly, prior
research has adopted this approach.
(4) Forecasting performance analysis
The aforementioned discussion shows VaR models based on the GED-GARCH and GED-TGARCH models
are able to effectively estimate and test the return risk over the in-sample period. Furthermore, in order to
investigate the forecasting power of those VaR models from a holistic perspective, we forecast the VaRs
over the out-of-sample period at the 95% confidence level, and compare them with corresponding real
return series. The results imply that for both the WTI and Brent returns, the rate of returns which fall
between the downside VaR and upside VaR to all the returns over the out-of-sample period is some 95.65%,
very close to 95%, and the LR value of VaR models is 0.3407, less than the critical value 3.84 at the 95%
confidence level. Thus, the results can be accepted (see Figs. 6 and 7). In other words, when VaR models are
used for the returns over the in-sample period to forecast the VaRs over the out-of-sample period, its
forecasting power is anticipated to be reliable.
Additionally, for the sake of comparison, we also introduce the popular HSAF method to develop
VaR models and test out-of-sample forecasting. However, the results indicate that its forecasting
performance is not very good and cannot be accepted according to the backtesting method provided by
Kupiec (1995). Specifically, whether for the WTI or Brent returns, the rate of returns between the
downside VaR and upside VaR to all the returns in the forecasting period is 91.87%, a little less than
95%. The corresponding LR value is 4.43, larger than the critical value concerned. Hence, the null
hypothesis is rejected. Which means the HSAF method is inappropriate to forecast the VaRs over the
out-of-sample period here. Therefore, we may say the GED based VaR method outperforms the HSAF
method.

Fig. 6. Real WTI return and its forecasted VaRs over the out-of-sample period at the 95% confidence level.
3168 Y. Fan et al. / Energy Economics 30 (2008) 3156–3171

Fig. 7. Real Brent return and its forecasted VaRs over the out-of-sample period at the 95% confidence level.

3.5. Risk spillover effect test for WTI and Brent returns

After the downside and upside VaR series are obtained for WTI and Brent returns, we develop two
statistics according to the concept of Granger causality in risk (Hong, 2001; Hong et al., 2003), and calculate
the statistics and their corresponding p-values. Subsequently, the one-way and two-way risk spillover
effects between WTI and Brent returns are tested. The results are shown in Table 7, in which M, the largest
effective lag truncation order, is equal to 10, 20, 30 respectively.
From the results in Table 7, it can be seen that, at the confidence level of 95% or 99%, whether the upside
or downside risk, both WTI and Brent returns have significant two-way Granger causality in risk. This
indicates that there is evident risk spillover effect between the two oil markets. It is not difficult to
understand this result. Economic globalization has made oil market a stage in a global and interactive
manner. Consequently, as two largest benchmark oil markets in the world, the evident and timely
interaction between WTI and Brent is not uncommon.
Moreover, to determine the direction of risk spillover, we calculate the statistic for the one-way Granger
causality in risk. The result reveals that there exists significant risk spillover effect from WTI to Brent
returns whether the confidence level is 95% or 99% and whether the return is upside or downside.
However, when we examine the risk spillover effect from Brent to WTI returns, complex considerations
need addressing. Specifically, at the 95% confidence level, there only exists risk spillover effect for the
downside returns, but not for the upside return. While at the 99% confidence level, the reverse is evident.
That is, there only exists risk spillover effect for the upside returns, but not for the upside return. The reason
for this difference, in our opinions, is that the former result might be caused from the fact that the VaR
model for the upside return at the confidence level of 95% is not adequate (see Tables 4 and 5). Whereas, all
the VaR models at the 99% confidence level are adequate. Hence, the latter result may be more reliable. Put
another way, we may say that there is no downside risk spillover effect from Brent to WTI returns at the
confidence level of 99%. This means when negative news occurs (which gives rise to the slump in oil price
returns), the risk information of the WTI return can help forecast the risk of the Brent return. But it will not
be the same case in the opposite direction. When there is positive news that can bring about the rise of oil
price returns, historical risk information in both WTI and Brent returns can be conducive to forecast their
future extreme return risks for one another.
To sum up, in terms of market risk, WTI oil price plays the dominant role in the international oil market
and has greater influence. This is caused by many factors. On the one hand, US dollar is the main invoicing
currency in the international oil market, and the dominance of US dollar enables American WTI oil market
to guide the developing direction of international oil markets. On the other hand, Brent crude oil is traded
Y. Fan et al. / Energy Economics 30 (2008) 3156–3171
Table 7
Test of risk spillover for the WTI and Brent returns

Conf. level Null hypothesis Upside risk spillover Downside risk spillover

M = 10 M = 20 M = 30 M = 10 M = 20 M = 30
95% WTI⇎Brent 100.5395 (0.0000) 70.8746 (0.0000) 57.9768 (0.0000) 119.1944 (0.0000) 84.3711 (0.0000) 68.7780 (0.0000)
WTI⇏Brent 4.0696 (0.0000) 2.0190 (0.0217) 1.4338 (0.0758) 32.2097 (0.0000) 22.1607 (0.0000) 17.9013 (0.0000)
Brent⇏WTI − 0.3940 (0.6532) 0.3099 (0.3783) 0.8464 (0.1987) 2.4689 (0.0068) 1.5226 (0.0639) 1.0506 (0.1467)
99% WTI⇎Brent 62.6007 (0.0000) 44.0779 (0.0000) 36.0828 (0.0000) 165.4863 (0.0000) 116.8463 (0.0000) 95.1344 (0.0000)
WTI⇏Brent 3.5124 (0.0000) 2.2833 (0.0112) 2.4514 (0.0071) 6.7315 (0.0000) 3.9886 (0.0000) 2.7193 (0.0033)
Brent⇏WTI 9.5487 (0.0000) 6.4280 (0.0000) 4.8993 (0.0000) 0.0446 (0.4822) −0.0388 (0.5155) − 0.2353 (0.5930)

Note: ⇎ represents there is no two-way Granger causality in risk, and ⇏ represents there is no one-way Granger causality in risk from the former to the latter. The corresponding p-values are
reported in parentheses.

3169
3170 Y. Fan et al. / Energy Economics 30 (2008) 3156–3171

primarily in London while WTI crude oil is traded primarily in NYMEX in America, as a result, by means of
the world-largest oil consumption of America as well as the strength and influence of NYMEX, the role of
WTI becomes outstanding in an increasing manner.

4. Conclusions and remarks

Resulting from the prior discussion above, we may safely draw the following conclusions:
(1) Significant volatility clustering can be found in both WTI and Brent returns. However, their volatile
levels are in the same cluster for most sample periods; even though the WTI return appears more
volatile in some periods. During the Gulf War, volatility of the Brent return turns out to be sharper
than that of the WTI return.
(2) The speed of decay of volatility shock in both WTI and Brent returns is very slow. Furthermore,
empirical study shows that there exists asymmetric leverage effect in the WTI return. Specifically,
the upside return has one and a half times larger impact on the future return than that of the
downside return with the same amplitude. We also find the volatility of the Brent return does not
have a leverage effect.
(3) We find that at the 99% confidence level, whether for the upside risk or downside risk, the VaR model
based on GED method, which is not uncommonly used, proves to be more robust and accurate than
that based on the standard normal distribution. However, at the confidence level of 95%, both are the
same and are, statistically speaking, adequate.
(4) When we examine the Granger causality in risk, the results indicate that at the confidence level of
95% or 99%, whether for the upside risk or downside risk, both WTI and Brent returns have
significant two-way Granger causality in risk. This implies that there exists tremendous risk spillover
effect. Further study reveals that at the confidence level 99%, when negative news makes the oil price
returns decline, risk information of the WTI market can help to forecast the risk of the Brent market.
But the process cannot be mirrored. When positive news brings about upward returns, risk
information in both of the markets can promote future risk forecasting for each other.
Overall, it should be noted that whether for the in-sample test or out-of-sample forecasting, the VaR
method for crude oil market risk proposed here appears more prudent. Moreover, we find that when the
VaR is calculated, whether it adopts the VaR model in this paper or the well-recognized HSAF method
largely depends on the sample data periods. For instance, the HSAF method does not have an advantage
over others to forecast the return in out-of-sample data here. Therefore, future work is expected to
investigate how to improve the adapting ability of VaR measurement and forecasting models. Only by doing
so, can the extreme risk in the crude oil markets be detected and forecasted in a more reliable and veracious
manner.

Acknowledgements

The authors gratefully acknowledge the financial support from the National Natural Science Foundation
of China under grant Nos.70425001, 70733005 and 70573104, and the National Key Projects from the
Ministry of Science and Technology of China (grants 2006-BAB08B01). We also would like to thank
Professor R.S.J. Tol and the anonymous referees for their helpful suggestions and corrections on the earlier
draft of our paper and upon which we have improved the content.

References

Barone-Adesi, G., Giannopoulos, K., Vosper, L., 1999. VaR without correlations for portfolios of derivative securities. The Journal of
Futures Markets 19, 583–602.
Bollerslev, T., 1986. Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics 31, 307–327.
Cabedo, J.D., Moya, I., 2003. Estimating oil price ‘Value at Risk’ using the historical simulation approach. Energy Economics 25,
239–253.
Engle, R.F., 1982. Autoregressive conditional heteroskedasticity with estimates of the variance of United Kingdom inflation.
Econometrica 50, 987–1007.
Fan, Y., 2000. VaR methodology and its application in stock market risk analysis. Chinese Journal of Management Science, 8, 26–30
(in Chinese).
Y. Fan et al. / Energy Economics 30 (2008) 3156–3171 3171

Fan, Y., Jiao, J.L., 2006. An improved historical simulation approach for estimating ‘value at risk’ of crude oil price. International Journal
of Global Energy Issues 25, 83–93.
Fan, Y., Wei, Y.M., Xu, W.X., 2004. Application of VaR methodology to risk management in the stock market in China. Computers &
Industrial Engineering 46, 383–388.
Feng, C.S., Wu, J.C., Jiang, F., 2004. To compute the oil market value at risk by applying semi-parametric approach. Journal of Hubei
University (Natural Science Edition) 26, 213–217 (in Chinese).
Giot, P., Laurent, S., 2003. Market risk in commodity markets: a VaR approach. Energy Economics 25, 435–457.
Hartz, C., Mittnik, S., Paolella, M., 2006. Accurate value-at-risk forecasting based on the normal-GARCH model. Computational
Statistics & Data Analysis 51, 2295–2312.
Hendricks, D., 1996. Evaluation of value at risk modeling using historical data. Economic Policy Review. Federal Reserve Bank of New
York.
Hilton, G.A., 2003. Value-at-risk, Theory and Practice. New York.
Hong, Y.M., 2001. Granger causality in risk and detection of risk transmission between financial markets. Working paper. Department
of Economics and Department of Statistical Science, Cornell University.
Hong, Y.M., Cheng, S.W., Liu, Y.H., Wang, S.Y., 2003. Extreme risk spillover between Chinese stock markets and international stock
markets. Working paper. Department of Economics and Department of Statistical Science, Cornell University.
Huang, Y.C., Li, C., Ma, W.F., 2005. The dilemma and outlet of China's oil price risk management. World Economy Study 10, 22–26.
Kupiec, P.H., 1995. Techniques for verifying the accuracy of risk measurement models. Journal of Derivatives 3, 73–84.
Li, P.M., Jia, M., Zhang, G.Y., 2005. China's strategies for the high oil price. Macroeconomics 12, 8–14 (in Chinese).
Lin, S.X., Tamvakis, M.N., 2001. Spillover effects in energy futures markets. Energy Economics 23, 43–56.
Ma, C.Q., Li, H.Q., Xu, S.Y., Yang, X.G., Li, H., 2001. Total-parametric methods of VaR and its applications in risk management of financial
market. Systems Engineering-theory & Practice 4, 74–79 (in Chinese).
Nelson, D.B., 1990. ARCH models as diffusion approximations. Journal of Econometrics 45, 7–38.
Sadeghi, M., Shavvalpour, S., 2006. Energy risk management and value at risk modeling. Energy Policy 34, 3367–3373.
Sadorsky, P., 2006. Modeling and forecasting petroleum futures volatility. Energy Economics 28, 467–488.
Tan, K.H., Chan, I.L., 2003. Stress testing using VaR approach — a case for Asian currencies. Journal of International Financial Markets
13, 39–55.
Zakoian, J.M., 1994. Threshold heteroskedastic models. Journal of Economic Dynamics and Control 18, 931–995.

Dr. Ying Fan is a professor at the Center for Energy and Environmental Policy Research, Institute of Policy and Management, Chinese
Academy of Sciences, China. Her research field is energy policy and system engineering. In 2004, she was a visiting scholar at Cornell
University, USA.

Mr. Yue-Jun Zhang is a Ph.D. candidate in Management Science at the Center for Energy and Environmental Policy Research,
Institute of Policy and Management, Chinese Academy of Sciences, China.

Dr. Hsien-Tang Tsai is a Professor of College of Management, National Sun Yat-sen University,Taiwan.

Dr. Yi-Ming Wei is a professor at the Center for Energy and Environmental Policy Research, Institute of Policy and Management of
the Chinese Academy of Sciences; and he was a visiting scholar at Harvard University, USA.

You might also like