VaR Estimation with Power EWMA Model ―
Conservativeness, Accuracy and Efficiency
EMF code: 450, 570
Correspondent and attending author, Associate Professor, Department of Business
Administration, Soochow University
Address：156, Sec.1, Kwei-Yang St., Taipei 100, Taiwan
Tel: 886-2-2311-1531 ext. 3602 Fax: 886-2-2382-2326
Manager, Information Department, WK Technology Fund
Associate Professor, Department of Economics, National Taiwan University
Financial asset returns are well-known to be non-normal and leptokurtic with the tails
fatter than normal distribution. The Standard EWMA estimator with the normality
assumption (used in JP Morgan's RiskMetrics®
model) will be inefficient and lead to
understate the true value of risk if the asset returns are fat-tailed distributed. On the
basis of the power exponential distribution (also known as the generalized error
distribution) the family EWMA estimators, nesting Power EWMA, Standard EWMA
and Robust EWMA, are proposed by Guermat & Harris (2002). Using these newly
developed estimators, we first forecast the VaR of daily returns for TAIEX, FTSE 100,
and DJIA. Subsequently, the back-testing is performed to evaluate the VaR models.
Performance assessment is based on a range of measures that address the
conservativeness, accuracy and efficiency of each model. The results demonstrate that
the members of the family of EWMA estimators based on power exponential
distribution rather than normal distribution offer a superior coverage for the extreme
risk over the RiskMetrics®
estimator, and show that Power EWMA performs excellent
accuracy in VaR estimation.
Key Words: VaR, Power EWMA, Standard EWMA, Robust EWMA,
Power Exponential Distribution, Fat-tailed
Value-at-Risk (VaR) has emerged as the widely used tool for risk management of
financial institutions. The successful implementation of VaR depends heavily on the
accurate estimation of the conditional distribution of portfolio returns. Owing to the
simple and rapid computations, the exponentially weighted moving average of past
squared returns, or EWMA estimator, become the most common approach to forecast
the conditional volatility of asset returns (JP Morgan, 1994; Dowd, 1998; Jorion,
2000). It has proved to be very effective at forecasting the volatility of returns over
short horizons, and often outperforms the forecasts of more sophisticated models such
as generalized autoregressive conditional heteroscedasticity (GARCH) when the
underlying asset returns is normally distributed. (See Boudoukh, Richardson and
Whitelaw, 1997; Alexander and Leigh, 1997)
However, a number of empirical studies show that asset returns are not normally
distributed. In particular, the conditional distribution of short horizon asset returns is
stylized found to be leptokurtic, with tails that are significantly fatter than those of the
normal distribution (See Mandelbort, 1963；Fama, 1965； Baillie and DeGennaro,
1990; Jansen and de Vries, 1991; Bollerslev, Chou and Kroner, 1992; Koedijk and
Kool, 1994; Loretan and Phillips 1994; Kearns and Pagan 1997). The Standard
EWMA estimator with the normality assumption (used in JP Morgan's RiskMetrics®
model) will be inefficient and lead to understate the true value of risk if the asset
returns are fat-tail distributed.
To remedy this problem, two main directions are proceeded to characterize the tail
behavior. The first is to set up the unconditional distribution as the mixture normal
distribution such as normal-Poisson (Jorion, 1988), normal-lognormal (Hsieh, 1989)
and Bernoulli-normal (Vlaar and Palm, 1993), still preserving the assumption of
homoskedasticity. That is, the volatility of asset returns is time-independent. The
second is to employ the non-normal distribution, for instance, Student-t distribution
(Bollerslev, 1987 ; Baillie and Bollerslev, 1989; Kaiser, 1996; Beine, Laurent and
Lecourt 2000), Laplace and double exponential distribution (Linden 2001) and
exponential power distribution (Varma, 1999; Guermat & Harris, 2002) to substitute
for normal distribution.
Lots of literatures have been addressed the weaknesses and strengths of various
VaR models. However, no single consistent measure of VaR model performance has
been developed. Of concern to supervisors, is whether the required minimum
regulatory-capital calculated by the internal model of the bank can offer an
appropriate coverage for its losses. Alternatively, it is the conservativeness of the
model. We identify the relative conservative models as those that systematically
produce higher estimate of risk relative to other models. With respective to accuracy
the risk manager should be concerned with whether the ex-post performance is
compatible with the theoretical desired level. The regulatory capital-adequacy
framework provides the incentive to develop efficient models, that is, models offer the
enough coverage for the risk to meet the supervisors’ requirement with the minimum
capital that must be held.
In this paper, therefore, we employ the family of nested power EWMA estimators
based on exponential power distribution that are robust to fat-tailedness and
leptokurtosis in the conditional distribution of returns to forecast the VaR. The data
consists of daily aggregate equity portfolios returns for the TAIEX, FTSE 100, and
DJIA. Next, we focus on three aspects- conservativeness, accuracy and efficiency of
model and propose a range of statistics based on these criteria to compare the
performance of the family models.
The paper is organized as follows. The following section introduces the
methodology, including the data description, the illustration of the nested power
EWMA family, the tail-index estimation and the measures of model evaluation while
Section 3 presents the results of the empirical evaluation. Some concluding remarks
are offered in Section 4.
The empirical evaluation uses aggregate daily equity returns for the US, UK and
Taiwan. The raw data used are daily price observations for the DJIA, FTSE100 and
TAIEX indices obtained from Dow Jones Indexes' Web Site, Datastream and Taiwan
Economic Journal (TEJ) databank respectively, for the period 01/01/84 to 31/12/02.
Continuously compounded returns were then calculated as the first difference of the
natural logarithm of each series, 1lnln −−= ttt IIr , where It is the price index value
for date t. Moreover, to investigate the diversifiable effects among different equity
returns, we compose the equally weighted portfolio of TAIEX, FTSE 100 and DJIA.
2.2 Tail Index Estimation
Financial asset returns are well-known to exhibit fatter tail than normal distribution.
To account for tail behavior, we directly estimate the tail index. The tail index is a
measure of the degree of tail fatness of underlying distribution and estimates with
extreme value theory (EVT) which addresses the characteristic of tail behavior of
distribution. The most famous and often applied estimator for tail index, due to its
easy implementation and asymptotic unbiasedness is proposed by Hill(1975) as
( ) ( )mn
= ∑ξ ， 2m ≥ (1)
where m is the pre-specified number of tail observations to be included. The
selection of m is crucial to obtain unbiased estimator of the tail index. n is the sample
size. ix is the ith increasing order statistic ( ni ....2,1= .), while the tail index α is equal
to ξ1 . Equation (1) means that Hill estimator measures the average of the ratios of
the each observed value relative to the threshold value in the predetermined tail area.
The larger the average, the smaller the tail index and the greater magnitude of the
However, there is considerable empirical evidence shows that it is biased in
relatively small samples and limited to the cases in which a larger sample is available.
To improve the Hill estimator, a recently developed alternative approach proposed by
Huisman et al. (2001) is especially useful for small samples. Their regression-based
approach is based on an approximation of the asymptotic expected value of m
where c is a constant depending on parameters of the distribution and the sample size.
If m becomes small, the bias goes down and the expectation goes to the true value
= . The variance of the estimator increases with small m.
mVar ≈ (3)
The idea of Huisman et al. (2001) is to use equation (2) in a regression analysis and
regress the values )(mξ (computed with conventional Hill estimator) against m as
mmm εββξ ++= , m = 1…κ (4)
0β is an estimator of
= . The authors propose to choose threshold
value κ to equal half the sample (
). Although the parameters in equation (4) can be
estimated by ordinary least squares (OLS). Equation (3) indicates that the variance of
Hill estimates is not constant for different m. The error term )(mε is heteroscedastic.
Accordingly, they propose a weighted least squares (WLS) approach to correct the
form of heteroscedasticity and improve the efficiency of the estimator. We apply the
modified Hill estimator to obtain the tail index estimates by use of both OLS and
2.3 The Power EWMA Variance Estimator
We introduce a general power EWMA estimator proposed by Guermat and Harris
(2002), nesting EWMA models that are more robust to the leptokurtosis of returns,
and which would therefore be expected to be more efficient when the conditional
distribution of returns is fat tailed. The power EWMA estimator is based on the
maximum likelihood estimator of the variance of the power exponential distribution
(also known as the generalized error distribution, or Box-Tiao distribution). The
probability density function of the power exponential distribution is given by
and ( )•Γ is the gamma function. The power exponential distribution has variance
equal to 2
σ , zero skewness and a kurtosis coefficient that depends on the value of
power parameterδ . When δ = 2, the power exponential distribution reduces to the
normal distribution. Whenδ > 2, the power exponential distribution is thin tailed and
platykurtic, and whenδ < 2, the power exponential distribution is fat-tailed and
leptokurtic. When δ = 1, the power exponential distribution reduces to the Laplace
distribution. The power exponential distribution with different power parameterδ is
shown in Figure 1. From Figure 1 we can find that the larger the power parameterδ ,
the more fat-tailed and leptokurtic the distribution is.
Conditional on the power parameterδ , the maximum likelihood estimator of the
standard deviation of the power exponential distribution is given by
Equation (7) is the unconditional variance estimator independent of past information.
Guermat and Harris (2002) transformed it into the conditional variance estimator and
replaced the unweighted average in (7) by an exponentially weighted average to yield
the power EWMA estimator
t rkg −
+ ∑−= λλσ
1 )()1( (9)
By recursive substitution, the Power EWMA estimator can be rewritten as
t rkg )()1(1 λλσσ −+=+ (10)
and so the Power EWMA estimator can be seen as an infinite weighted average of
past squared returns, incorporating information from all past shocks to the power
parameter k of returns, but with exponentially declining weights. Alternatively, by
Figure 1 Power Exponential Distribution with
Different Power Parameter δ
using the fact that 2
1 +++ += tttr εσ where 2
1+tε is a zero mean random shock that is
orthogonal to the time t information set, the power EWMA estimator can also be
interpreted as an infinite order autoregressive model for the kth powered return. When
k = 2, the power EWMA estimator coincides with the standard EWMA estimator
1 )1( ttt rλλσσ −+=+ (11)1
The standard EWMA estimator is a special case of the generalized autoregressive
conditional heteroscedasticity, or GARCH model (Engle, 1982; Bollerslev, 1986).
The GARCH(1,1) model for the conditional variance of returns is given by
1 ttt rβσαασ ++=+ (12)
where 0α , 1α and 1β are parameters to be estimated. When 0α =0 and 1β = 1− 1α ,
the GARCH model reduces to the standard EWMA estimator, and is alternatively
known as Integrated GARCH or IGARCH. For k < 2, the power EWMA estimator is
less sensitive to extreme observations and thus we may expect it to be more efficient
when the conditional distribution of returns is leptokurtic. The role of the function g
(k) is to preserve the integrated nature of the volatility, in keeping with the standard
When k = 1, the power exponential reduces to the Laplace distribution, and the
power EWMA estimator reduces to
The Laplace distribution is commonly used in the context of robust estimation, and so
the EWMA estimator given by (13) might therefore be thought of as a ‘robust’
when 2k = , 1
when 1k = , 22
EWMA estimator. The power EWMA estimator therefore nests the standard EWMA
estimator, the robust EWMA estimator, and a continuum of estimators that lie
between the two, as well as estimators that are even more sensitive to outlying
observations than the standard EWMA estimator, and those that are even less
sensitive to them than the robust EWMA estimator. The power EWMA estimator
described above is a special case of the NGARCH model of Higgins and Bera (1992),
t r1101 βσαασ ++=+ (14)
When 00 =α and ( ) ( )kg11 1 αβ −= , the NGARCH model reduces to the power
EWMA estimator. As with the standard GARCH model, the parameters of the
NGARCH model can be estimated by maximum likelihood. In contrast, members of
the family of power EWMA estimators have only a single parameter – the decay
factor – and consequently their implementation is very much more straightforward
than that of the more sophisticated NGARCH model. The relationship among these
nested models can be summarized as Figure 2. The parameters of the models for the
four equity returns are firstly estimated by maximum likelihood based on power
exponential distribution, using BHHH algorithm (Berndt et al, 1974) with a
convergence criterion of 0.00001 applied to the function value. Next, we apply the
likelihood ratio statistics3
to test the restrictions on the NGARCH model that are
implied by the power EWMA estimator, and introduce the standard EWMA based on
normal distribution (well known as the RiskMetrics) as the benchmark to compare the
out-of-sample performance of various power EWMA estimators.
Under regularity, the large sample of likelihood ratio test statistic λχ ln22
−= is chi-squared, with
degrees of freedom equal to the number of restrictions imposed, where UR L/L=λ ， UL denotes
the likelihood function evaluated without regard to constraints and RL is the constrained likelihood
function estimate. Since both likelihoods are positive, and a restricted optimum is never superior to an
unrestricted one. Such that0＜ λ ＜1. If λ is too small, then the doubt is cast on the restriction,
otherwise when λ is close to 1, the null hypothesis is accepted.（see Greene (2000).
1t rβσαασ ++=+
1t rβσαασ ++=+
1t r)1( λλσσ −+=+
1t r)1( λλσσ −+=+
1t r)k(g)1( λλσσ −+=+
1t r)k(g)1( λλσσ −+=+
tt1t r2)1( λλσσ −+=+
tt1t r2)1( λλσσ −+=+
1t rβσαασ ++=+
1t rβσαασ ++=+
When ( ) ( )kg10α 110 αβ −== 、
When k = 1
When k = 2
When 110 10、α αβ −==
2.4 The Estimation of VaR
We compute out-of-sample one-day VaR forecasts for the four portfolios, using each
of the EWMA estimators. A rolling window is used for the estimation of each model.
Each estimated model is then used to forecast the VaR of the portfolios with window
length of 10, 50, 100, 250, 500, 10004
observations respectively. Moreover, VaR is
computed for the 99% and 95% confidence levels. The VaR of each portfolio in each
period t is forecast by the formula
1t1t )(VaR ++ −= σαδ (15)
where 1+tσ is the standard deviation of the portfolio’s return, r 1+t , conditional on the
time t information set, andδ (α ) is theα - quantile of the standardized (i.e. zero mean,
unit variance) empirical power exponential distribution and α is one minus the VaR
confidence level. The standardized empirical distribution is defined as the return
series over the window, scaled by the estimated standard deviation for each of those
days. The standard deviation estimate used to standardize the return is obtained from
According to the guideline of BIS the window lengths must be at least 250 business days. Our results
of VaR estimated show that when the window lengths greater than 100 days, the discrepancy among
estimated VaRs of different window lengths are such small that can be neglected. Consequently,
window length of 250 observations is enough. It is not necessary to extend window length longer.
Figure 2 The Nested Power EWMA Models
the EWMA model (see Hull and White, 1998).
2.5 Model Evaluation
The evaluation of VaR forecasts is not straightforward. As with the evaluation of
volatility forecasting models, a direct comparison between the forecast VaR and the
actual VaR can not be made, since the latter is unobservable. A variety of evaluation
methods have been proposed (see, for instance, Kupiec, 1995; Christofferson, 1998;
Lopez; 1999). Up to now, no single definitive of VaR model performance has been
developed. To evaluate the performance of the family models we propose a range of
statistics that address different aspect of the usefulness of VaR models to risk
manager and supervisory authorities. We focus on three aspects of model:
conservativeness, accuracy and efficiency (Engel and Gizycki, 1999).
Mean Relative Bias (MRB)
Engel and Gizycki (1999) defined the conservativeness of models in terms of the
relative size of VaR for the risk assessment. The larger the VaR value is, the more
conservative the model. To measure relative size of VaR among different models,
they apply the mean relative bias developed by Hendricks (1996). The mean relative
bias statistic captures the degree of the average bias of the VaR of specific model
from the all-model average. Given T time period and N VaR models, the MRB of
model i can be calculated as:
Different user of the VaR model will focus on different types of inaccuracies. It may
be expected that supervisors will pay more attention to the underestimate of losses
while the financial institutions will be concerned more about over-predictions of
losses due to the capital adequacy requirement. In this study, we define the accuracy
as what extent is the rate of failure (or exception) of specific model close to the preset
significant level. The three accuracy measures：binary loss function, LR test of
unconditional coverage (Kupiec ,1995) and the scaling multiple to obtain coverage are
presented as follows.
a. Binary Loss Function (BLF)
The binary loss function is based on whether the actual loss is larger or smaller than
the VaR estimate. Here we are simply concerned with the number of the failure rather
than the magnitude of the exception. If the actual loss is larger than the VaR then it is
termed as an“exception”(or failure) and has the equal value of 1 , all others are 0.
Aggregate the number of failure across the all dates and divide it by the sample size.
The BLF is obtained as the rate of failure. The closer the BLF value is to the
confidence level of the model, the more accuracy the model.
b. LR Test of Unconditional Coverage (LRuc)
The BLF provides a point estimate of the probability of failure. In other words, the
accuracy of the VaR model requires that the BLF on average be equal to one minus
the prescribed confidential level of VaR model. The model should provide the correct
unconditional coverage of loss. Kupiec (1995) proposed a likelihood ratio test based
on the binomial process which can be applied to determine if the rate of failure is
statistically compatible with the expected level of confidence. Given the sample size
T and the frequency of failure N governed by a binomial probability, the likelihood
ratio statistic of the unconditional coverage hypothesis α=p:H0 can be stated as
[ ] 2
Under the null correct hypothesis of correct unconditional coverage, the LRuc has a
chi-squared distribution with one degree of freedom.
c. Multiple to Obtain Coverage (MOC)
To highlight the magnitude of the deviation that the losses from VaR estimate, we
compare the multiple to obtain coverage proposed by Engel and Gizycki (1999). The
multiple equivalent, Xi , of risk measure for model i is calculate so that
αii TF = ，
where the Fi is equivalent to the total number of failures. Ti is the sample size and α
is the significant level of the model. 1, +∆ tiP denotes the realized profit or loss on t +1
Mean Relative Scaled Bias (MRSB)
Efficiency is important since VaR measures are used by supervisor and the internal
management of financial institutions to influence investors’ incentives. A more
efficient VaR model provides more precise resource allocation signals to the financial
institutions. Hence we address the aspect of efficiency on the capacity for a model to
provide adequate risk coverage with minimum capital. The Mean Relative Scaled
Bias (MRSB) (Engel and Gizycki (1999)) is aimed to evaluate which model, once
suitably obtain the desired risk coverage level, produces the smallest VaR measure.
There are two steps in calculating the MRSB measure. First, the scaling should be
calculated by multiplying the VaR for each model by the multiple needed to obtain
the 95% or 99% coverage as described in MOC measure. Subsequently, we compare
the scaled VaR measures with the average relative size to all-model average. The
MRSB measure is given as following.
3. Empirical Results
3.1 Descriptive Statistics
The preliminary statistics for all four equity returns are summarized in Table1. As is
commonly found daily financial asset returns are not normally distributed. In all cases
the Jarque-Bera test for normality are highly significant with excess kurtosis5
negatively skewed. Moreover, DJIA exhibits the most leptokurtic distributed with the
kurtosis of 66.52. The variance of TAIEX is the largest of all whereas that of
portfolio is the smallest due to the diversification effects.
Table1 Descriptive Statistics
Excess kurtosis means the distribution has kurtosis greater than that of normal distribution whose
kurtosis equal to 3.
TAIEX FTSE 100 DJIA Portfolio
Mean 0.00034 0.00030 0.00041 0.00031
Standard deviation 0.01841 0.01060 0.01116 0.00836
Maximum 6.58% 7.60% 9.67% 5.04%
Minimum -7.05% -13.03% -25.63% -11.32%
Sample size 5289 4768 4757 5620
Note: P-values are reported in parentheses for the skewness, kurtosis and the
3.2 Estimation of Tail Index
The most important information in terms of characterizing the limiting extreme
distribution of the tail is the tail index. The index of right-tail, left-tail and both
for the four equity return distributions are estimated by OLS and WLS.
The modified Hill estimate results are shown in Table 2. From Table 2 we can
find that the estimates of normal distribution, as a benchmark, are all around 8.5.
All tail index estimates, varying between 3.24 and 5.88, are greater than 8.5
indicates that all the equity return distributions exhibit fatter tails than the normal
distribution, as commonly found in literatures. Further, the DJIA has the highest
degree of fat-tailedness, also consistent with the findings of preliminary statistics.
If these estimates differ significantly over both tails, it is inappropriate to use the
estimate obtained from the combined information in both. However, our results
show that the left-tail estimates display little fatter tail than the right-tail do for
the majority of equity return.
Table2 Tail- Index Estimates
The Jarque-Bera statistic has a chi-squared distribution with two degrees of freedom under the null
hypothesis of normality.
All observations are taken in excess of their sample mean. The left tail is examined by using the
absolute value of all negative returns, the right tail is examined by using the absolute value of all
positive returns and to both tails simultaneously we use the absolute values of all returns.
DJIA Portfolio Normal
Sample size 2,644 2,383 2,378 2,808 24,999
Sample size 1,310 1,155 1,171 1,380 12,527
Sample size 1,333 1,228 1,206 1,428 12,472
Note: P-values are reported in parentheses for the estimate of the tail
3.2 Estimation of Power EWMA
The NGARCH model, and the power EWMA estimator that it nests, are based on the
maximum likelihood estimators of the variance of the power exponential distribution.
Table 3 reports the parameter estimates of each model for each of the three return
series. The first column of each panel in Table 3 gives the results for the unrestricted
NGARCH model. The sum of the parameters 1α and 2α is 0.99168 for the TAIEX,
0.97781 for the FTSE 100, 0.99096 for the DJIA, 0.98757 for the portfolio. They are
all very close but not equal to 1.The second column reports the results for the power
EWMA estimator which imposes the restrictions that 0α =0 and 2α = 1- 1α .The
estimated power parameter of the conditional distribution,δ , in the power EWMA
model is 1.52346 for the TAIEX ,1.57267 for the FTSE 100, 1.24451for the DJIA and
1.50543 for the portfolio. The estimated power parameter of the conditional
distribution of the power EWMA model in each case is very close to the estimated
power parameter, k, of the conditional variance model, which is consistent with the
results of the previous studies (Nelson and Foster ,1994; Guermat and Harris,
2002).The power parameter, k, varying between 1.10735 for the DJIA and 1.34356 for
the FTSE 100, for all four series is closer to unity than to two suggest that the robust
EWMA estimator might be expected to perform better than the standard EWMA
estimator. The estimated decay factors of the power EWMA model are 0.95535 for
the DJIA, 0.94212 for the portfolio, 0.93792 for the FTSE 100 and 0.92986 for the
TAIEX. These are all very close to the value of 0.94 that is suggested by JP Morgan.
The estimated power parameters of the conditional distribution,δ , varying between
1.22393 and 1.60166, are all smaller than 2 which confirms again our previous
findings that all the equity returns exhibit fat tailed and leptokurtic distribution.
Table 3 Estimates of EWMAs
NGARCH Power EWMA
  
[ 11 α− ] [ 11 α− ] [ 11 α− ]
LOGL 18216.77374 18194.93186 18190.44856 18174.73734
  
[ 11 α− ] [ 11 α− ] [ 11 α− ]
LOGL 18859.79677 18834.25421 18824.40090 18811.12709
  
[ 11 α− ] [ 11 α− ] [ 11 α− ]
LOGL 18856.36568 18846.36007 18828.91324 18843.71799
  
[ 11 α− ] [ 11 α− ] [ 11 α− ]
LOGL 23680.06706 23665.59646 23653.39508 23650.26249
Note: 1. The restricted parameter values imposed on NGARCH are
reported in square brackets [ ].
2. P-values are reported in parentheses ( ) for the parameter
3. LOGL is the maximum value of the log likelihood function.
3.3 Restrictions Test on the Nested Model
The power EWMA estimator is nested by the NGARCH model, and therefore
imposes certain restrictions on the NGARCH model. In this section, we test whether
those restrictions are supported by the data. Table 4 reports likelihood ratio tests of
the various restrictions. Table 4 shows that, owing to the precision with which the
NGARCH parameters are estimated, the null hypothesis that the sum of the
NGARCH parameters is unity can be rejected for all four series at significant level of
1%. These results suggest that while the true data generating process is not quite
integrated, the sum of the estimated parameters is very close to unity and so, over
short horizons, their dynamic properties should be reasonably well described by an
integrated NGARCH, or power EWMA process.
The third and fourth rows report results for the standard EWMA estimator and the
robust EWMA estimator. These models impose the restrictions on the power EWMA
estimator that k = 2 and k = 1, respectively. On the basis of likelihood ratio tests
reported in Table 4, both models can be rejected for the TAIEX, FTSE 100 and
portfolio at significant level of 1% with the except of DJIA, the robust EWMA
estimator can be rejected only at significant level of 5%. A more inspection of the
results, we can find that, due to the different degree of the fat-tailedness, the extent of
rejection of Standard EWMA for the DJIA is stronger than the others, whereas the
rejections of Robust EWMA for the TAIEX, FTSE 100 and portfolio are stronger than
the DJIA. In sum, all the restrictions imposed on the EWMA models can not be
supported by the sample asset returns. The unconstrained model, NGARCH, is the
most suitable one for our data.
Table 4 L R Test of Restrictions on the Nested Model
H0 TAIEX FTSE 100 DJIA Portfolio
Note: 1.This table reports the likelihood ratio test statistics to test the respective restrictions. The LR statistics are
chi-square distributed with degrees of freedom equal to the number of restrictions imposed.
2.*denotes the significant level at 5% and ** denotes the significant level at 1%.
3. P-values are reported in parentheses.
3.4 Model Evaluation
For each model, the average value of the performance criteria across sample assets
for the 95th
percentile and the average of the two are summarized in Table 5.
The mean relative bias tends to fall between 7.5% and -7% for average, indicating that
there is little difference in the magnitude of risk measure across the models. The most
conservative model is the robust EWMA which produces the largest average VaR
estimate, while the RiskMetrics®
is the least conservative model which produces the
lowest average VaR estimate.
With the exceptions of the RiskMetrics®
in 99 th
VaR estimate and the Robust
EWMA in 95th
VaR estimate, the VaR measures produce the rates of failure (i.e. BLF)
close to the benchmarks of 0.01 and 0.05 respectively. The RiskMetrics®
normal distribution appears to be much more accurate when forecasting the 95th
than the 99th
VaR, which suggests that it is less sensitive to the outlying observations
than other models based exponential power distribution. The most conservative model
Robust EWMA thereby produce the least BLF either in 99th
VaR estimate or in
VaR estimate. The average BLF is 2.36%, for the Robust EWMA, 2.72% for the
NGARCH and 2.85% for the Power EWMA, all lower than the benchmarks of 3%,
indicating that these models understate risk while others 3.20% for the Standard
EWMA and 3.46% for the RiskMetrics®
, higher than the benchmark of 3%, indicating
that these models overstate risk. The Power EWMA provides the BLF which most
close to the benchmark, suggesting that Power EWMA is the most accurate model.
The results of likelihood ratio test show that the numbers of the NGARCH, Power
EWMA, Standard EWMA, Robust EWMA, RiskMetrics models that reject the null
hypothesis, for the 99th
VaR estimates are 2, 0, 1, 2, 4 and for the 95th
are 2, 1, 0, 4, respectively. The Power EWMA, rejecting the null hypothesis only once
and having the least average LR statistics, achieve the best accurate results. Next to
the Power EWMA, the sequence are Standard EWMA, NGARCH, Robust EWMA
For the 99th
VaR estimate the multiples needed to obtain coverage are all larger
than one except for the Robust EWMA model. Put it another way, the Robust EWMA
model overestimates risk while the other models under-predict risk. Worthy to note
that for the 95 th
VaR estimate the two models, the Standard EWMA and the
, with power parameter of 2 both require the multiples very close to one.
This indicates that they are both accurate models at 95% confidence level. For
average the Power EWMA need the multiple most close to one consistent with the
results of previous accuracy measures, achieving the best accuracy, The followings
are Standard EWMA, NGARCH, Robust EWMA, RiskMetrics®
Comparing across all models, we find that the NGARCH exhibit greatly different
performance of efficiency. It provides the most efficient at 99th
measure whereas the
poorest efficient at 95th
measure. The RiskMetrics®
dominates the other models for
Table 5 The Performance Measures of Models
Robust EWMA RiskMetrics®
MRB 0.02537 0.01204 -0.02654 0.08799 -0.09886
BLF 0.98% 1.04% 1.25% 0.80% 1.72%
MOC 1.00286 1.02196 1.06662 0.95036 1.14936
LRuc 2.26346(2) 0.75996(0) 2.74555(1) 3.16126(2) 17.16379(4)
MRSB -0.00650 -0.00005 0.00448 0.00008 0.00199
MRB 0.01729 -0.00128 -0.03687 0.06299 -0.04213
BLF 4.46% 4.66% 5.15% 3.92% 5.19%
MOC 0.96373 0.97798 1.01458 0.91985 1.01823
LRuc 6.40036(2) 1.64764(1) 1.01450(0) 11.83509(4) 0.89814(0)
MRSB 0.00214 -0.00065 -0.00021 0.00072 -0.00201
MRB 0.02133 0.00538 -0.03170 0.07549 -0.07049
BLF 2.72% 2.85% 3.20% 2.36% 3.46%
MOC 0.98330 0.99997 1.04060 0.93511 1.08380
LRuc 4.33191(4) 1.20380(1) 1.88002(1) 7.49817(6) 9.03097(5)
MRSB -0.00218 -0.00035 0.00213 0.00040 -0.00001
Note: The numbers of the model that reject the null hypothesis are reported in
standard EWMA based on normal distribution, is widely used to
forecast the variance of the conditional distribution of asset returns. It is appropriate
when the asset returns are drown from a normal distribution. However, there is
considerable evidence that the distribution of most financial returns is not well
approximated by normal distribution, even conditionally. The conditional distribution
of assets returns is typically found to be leptokurtic, and have fatter tail than normal
distribution. To improve the inefficiency of EWMA estimators, we introduce the
power exponential distribution to construct a serial of EWMA family estimators such
as Power EWMA, Standard EWMA, and Robust EWMA proposed by Guerma &
Harris (2002). The daily returns of TAIEX, FTSE 100 and DJIA are used to forecast
the VaR. Considering the different aspects of the usefulness of VaR models to risk
manager and supervisory authorities, we focus on three aspects- conservativeness,
accuracy and efficiency of model and propose a range of statistics based on these
criteria to evaluate the performance of the family models. The concluding remarks are
provided from the empirical findings as follows.
From the results of descriptive statistics, estimated tail-index and the estimated
power parameter of power exponential distribution, we obtain the consistent findings
that all equity returns have significant fat-tailed and leptokurtic distribution. Among
of them, the DJIA exhibits the highest degree of the fat-tailedness with leptokurtosis.
Next are FTSE 100, portfolio and the least is TAIEX. The estimated decay factors of
the power EWMA model are 0.95535 for the DJIA, 0.94212 for the portfolio, 0.93792
for the FTSE 100 and 0.92986 for the TAIEX. These are all very close to the value of
0.94 that is suggested by JP Morgan.
Comparing across all models, we find that the most conservative model is the
robust EWMA which produces the largest average VaR estimate, while the
is the least conservative model which produces the lowest average VaR
estimate. The Power EWMA providing the BLF and MOC which are both most close
to the benchmark, rejecting the null hypothesis of LR test only once and having the
least average LR statistics, achieve the best accurate results. A closer inspection of the
results reveal that for the 95 th
VaR estimate, the Standard EWMA and the
, with power parameter of 2 both require the multiples very close to one..
They are more accurate at 95% confidence level (less extreme risk) than at 99%
confidence level (more extreme risk). It is worthy to note that the NGARCH exhibit
greatly different performance of efficiency, which provides the most efficient at
measure whereas the poorest efficient at 95th
measure. Another finding is that the
dominates the other models for 95th
The back-testing results demonstrate that the members of the family of EWMA
estimators with the power exponential distribution have improved the inefficiency of
- the traditional Standard EWMA estimator based on normal
distribution, and offer an appropriate coverage of the extreme risk. Due to the
flexibility of power parameter of the model, the Power EWMA performs a superior
accuracy in VaR estimation over the other EWMA estimators.
Artzner, P., F. Delbaen, J-M Eber & D. Health (1997), “Thinking Coherently,” Risk,
10, pp. 68-71.
Baillie, R., and T. Bollerslev (1989), “The Message in Daily Exchange Rates: A
Conditional-Variance Tale,” Journal of Business and Economic Statistics, 7,
Baillie, R. & R. de Gennaro (1990), “Stock Returns & Volatility,” Journal of
Financial & Quantitative Analysis, 25, pp. 203-14.
Berndt, E. K., B. H. Hall, R. E. Hall & J. A. Hausman (1974), “Estimation and
Inference in Nonlinear Structural Models,” Annals of Economic and Social
Measurement, 4, pp. 653-65.
Bollerslev, T. (1986), “Generalised Autoregressive Conditional Heteroskedasticity,”
Journal of Econometrics, 31, pp. 307-27.
Bollerslev, T. (1986), “Generalized Autoregressive Condtional Heteroskedasticity,”
Journal of Econometrics, 31, pp.307–327
Bollerslev, T., R. Chou & K. Kroner (1992), “ARCH Modelling in Finance,” Journal
of Econometrics, 52, pp. 5-59.
Boudoukh, J., M. Richardson & R. F. Whitelaw (1997), “Investigation of a Class of
Volatility Estimators,” Journal of Derivatives, 4, pp. 63-72.
Cassidy, C. & M. Gizycki (1997), “Measuring Traded Market Risk: Value-at-Risk and
Backtesting Techniques,” Reserve Bank of Australia Research Discussion Paper,
Christoffersen, P. (1998), “Evaluating Interval Forecasts,” International Economic
Review, 39, pp. 841-62.
Dowd, K. (1998), Beyond Value-at-Risk: The New Science of Risk Management, New
Engle, R. (1982), “Autoregressive Conditional Heteroskedasticity with Estimates of
the Variance of United Kingdom Inflation,” Econometrica, 50, pp. 987-1007.
Fama, E. (1965), “The Behaviour of Stock Market Prices,” Journal of Business, 38,
Greene, W. H. (2000), Econometric Analysis, Prentice Hall.
Guermat, C. & R. D. F. Harris (2002), “Robust Conditional Variance Estimation and
Value at Risk, ” Journal of Risk, 4, pp. 25-41.
Hendricks, D. (1996), “Evaluation of Value at Risk Models Using Historical Data,”
Federal Reserve Bank of New York Economic Policy Review, April, pp. 39-69.
Higgins, M. & A. Bera (1992), “A Class of Nonlinear ARCH Models,” International
Economic Review, 33, pp. 137-58.
Hill, B. (1975), “A Simple General Approach to Inference about the Tail of a
Distribution,” Annals of Statistics, 3, pp. 1163-74.
Hsieh, D. (1989), “Modeling Heteroskedasticity in Daily Foreign Exchange Rates,”
Journal of Business and Economic Statistics, 7, pp.307–317.
Huisman, R., K. G. Koedijk, C. J. M. Kool and F. Palm (2001), “Tail-Index Estimates
in Samll Samples,” Journal of Business & Economic Statistics, 19, pp. 208-16.
Hull, J. and A. White (1998), “Value at Risk When Daily Changes in Market Variables
Are Not Normally Distributed,” Journal of Derivatives, 5, pp. 9-19.
Jackson, P., D. J. Maude & W. Perraudin (1997), “Bank Capital & Value at Risk,” The
Journal of Derivatives, Spring, pp. 73-89.
Jorion, P. (1997), Value at Risk：The NewBenchmark for Controlling Market Risk,
Jorion, P. (1988), “On Jump Processes in the Foreign Exchange and Stock Markets,”
The Review of Financial Studies, 68, pp.165–176.
Jorion, P. (2000), Value at Risk: The New Benchmark for Managing Financial Risk,
Kaiser, T. (1996), “One-Factor-GARCH Models for German Stocks - Estimation and
Forecasting,” Working Paper, Universiteit Tubingen
Mandelbrot, B. (1963), “The Variation of Certain Speculative Prices,” Journal of
Business, 36, pp.394–419.
Vlaar, P., and F. Palm (1993), “The Message in Weekly Exchange Rates in the
European Monetary System: Mean Reversion, Conditional Heteroskedasticity
and Jumps,” Journal of Business and Economic Statistics, 11, pp.351–360.