Studies in Nonlinear Dynamics & Econometrics: Markov-Switching GARCH Modelling of Value-at-Risk

Studies in Nonlinear Dynamics &
Econometrics
Volume 12, Issue 3 2008 Article 7
R EGIME -S WITCHING M ODELS IN E CONOMICS AND F INANCE
Markov-Switching GARCH Modelling of

Value-at-Risk
Rasoul Sajjad∗ Jerry Coakley†
John C. Nankervis‡
∗
University of Essex, rsajja@essex.ac.uk
†
University of Essex, jcoakley@essex.ac.uk
‡
University of Essex, jcnank@essex.ac.uk
Brought to you by | Periodicals Section, Albert Sloman Libr.

Authenticated | 155.245.46.3
Download Date | 12/18/12 12:09 PM
Markov-Switching GARCH Modelling of
Value-at-Risk
Rasoul Sajjad, Jerry Coakley, and John C. Nankervis
Abstract
This paper proposes an asymmetric Markov regime-switching (MS) GARCH model to esti-
mate value-at-risk (VaR) for both long and short positions. This model improves on existing VaR
methods by taking into account both regime change and skewness or leverage effects. The perfor-
mance of our MS model and single-regime models is compared through an innovative backtesting
procedure using daily data for UK and US market stock indices. The findings from exceptions and
regulatory-based tests indicate the MS-GARCH specifications clearly outperform other models in
estimating the VaR for both long and short FTSE positions and also do quite well for S&P posi-
tions. We conclude that ignoring skewness and regime changes has the effect of imposing larger
than necessary conservative capital requirements.

Sajjad et al.: Markov-Switching GARCH Modelling of Value-at-Risk 1
1 Introduction
Value-at-Risk (VaR) is one of the most popular approaches to quantifying market
risk. It yields an estimate of the likely losses which could rise from price changes
over a pre-determined horizon at a given confidence level. It is usual that VaR is
separately computed for the left and right tails of the returns distribution
depending on the position of the risk managers or traders. Traders with long
positions are exposed to the risk of price falls (left tail VaR) while those with
short positions stand to lose when prices increase (right tail VaR). Symmetric
VaR models of the Generalised Autoregressive Conditional Heteroskedasticity
(GARCH) class have difficulties in correctly modeling the tails of the returns
distribution (Giot and Laurent, 2003) due to leverage effects.1
There are two main approaches that allow for the leverage effect in volatility
forecasting. The first is the use of conditional asymmetric models which extend
ARCH models by imposing an asymmetry parameter in the conditional variance
equation. The second approach is based on the use of asymmetric density
functions for the error term or an asymmetric confidence interval around the
predicted volatility.2 Although such approaches provide an improvement in fit
compared with symmetric models, the empirical evidence suggest that the
persistence in the conditional variance is likely to exhibit substantial upward bias.
One potential source of bias is that the means and variances are assumed fixed
rather than varying over the entire sample period (Lamoureux and Lastrapes,
1990, Timmermann, 2000, and Mikosch and Starica, 2004).
A Markov Regime-Switching (MS) approach can resolve this by
endogenising changes in the data generating process. Gray (1996) extended the
Hamilton (1994) MS model to the MS-GARCH framework by allowing within-
regime, GARCH type heteroskedasticity. This was subsequently modified by
Klaassen (2002). Marcucci (2005) compares a set of GARCH, EGARCH and
GJR-GARCH models within an MS-GARCH framework (Gaussian, Student’s t
and Generalized Error Distribution for innovations) in terms of their ability to
forecast S&P100 volatilities. Ane and Ureche-Rangau (2006) extend the regime-
switching model developed by Gray (1996) to an Asymmetric Power (AP)
GARCH model to analyze empirically Asian stock indices returns. Their
empirical results indicate that all the generalizations introduced by the MS-
APGARCH model are statistically and economically significant.
In this paper we introduce a MS-GARCH framework to take account of both
asymmetry and regime changes in returns data in forecasting VaR. Our study
1
This means that a negative shock leads to a higher conditional variance (volatility) in the
subsequent period than a positive shock would.
2
See Bond (2000) for a survey on early asymmetric conditional density functions.

2 Studies in Nonlinear Dynamics & Econometrics Vol. 12 [2008], No. 3, Article 7
builds upon the previous literature in several ways. First, we focus on VaR for
long and short positions allowing for asymmetries in both conditional variance
and distribution of error terms. This is crucially important in taking account of the
leverage effect in stock markets. Existing studies such as Marcucci (2005) assume
symmetric distributions and focus on long VaR only. Second, while Giot and
Laurent (2003) model the long and short VaR using a single-regime APARCH
model combined with the skewed Student’s t distribution, we extend the analysis
to the MS context since our results indicate that regime change matters. Third, we
evaluate out-of-sample model performance by using a novel combination of
exceptions and regulatory-based backtesting procedures. This provides a more
robust evaluation of model performance than the in-sample analysis found in
existing studies such as that of Ane and Ureche-Rangau (2006).
The rest of the paper is organized in the following way. Section 2 presents
the selected volatility models implemented to model VaR for the long and short
trading positions. The empirical results for model specification and diagnostic
tests are presented in Section 3. In Section 4, the performance of competing
models in forecasting VaR is examined. Finally, Section 5 concludes the paper.
2 VaR Models
This section presents the VaR models that are used to model the long and short
sides of daily trading positions. The GARCH model, originally introduced by
Bollerslev (1986), is the most popular model in volatility forecasting and financial
risk management. We go beyond the single-regime GARCH framework and
consider MS-GARCH models for modeling VaR for long and short trading
positions.
2.1 MS-GARCH model

The standard GARCH family models are implemented to mimic the volatility
clustering exhibited by most financial time series. However, they are not able to
capture possible regime change in the variance process since they often entail a
high volatility persistence of individual shocks (see Lamoureux and Lastrapes,
1990, and Timmermann, 2000). One possible way of modeling changing volatility
persistence is to combine MS models, as introduced in Hamilton (1994), with
GARCH type models in which volatility persistence can take different values
depending on whether it is in a high or low volatility regime (state).
Let st be a random variable that can assume only integer values {1, 2… M}.
Then, an M-state Markov chain with transition probabilities pij is a process in
which the probability that unobserved st equals some particular value j depends

on the past only through the most recent value st −1 . That is
Pr{st = i | st −1 = j , st − 2 = k ,...} = Pr{st = i | st −1 = j} = pij ,

(2.1)
for i, j = 1,2,…, M.
A M × M matrix P, known as the transition matrix, contains transition

probabilities pij giving the probability that state i will be followed by state j.
Then the dynamics of returns is given by:
φ s ( L)(rt − γ s ) = θ s ( L)ε t ,
t t t
(2.2)
ε t = h Zt 1/ 2
t , st
where the innovations Z t are i.i.d. with zero mean, unit variance and marginal
density function f z (Z ) and L is the lag operator. The conditional variance ht can be
defined either by:
η st / 2 η
β s ( L)ht ,s
t t
= ω st + [ β st ( L) − ϕ st ( L)](1 + λst S t ) | ε t | st , (2.3)
giving the ARCH class of models, or by:
β s ( L) log ht ,s = ω s + [ β s ( L) − ϕ s ( L)](ht−,s1 / 2 (| ε t | +λs ε t t )) ,

t t t t t t t
(2.4)
giving the EGARCH class, where η > 0 denotes the power parameter, st = 1 if
ε t < 0 and 0 otherwise, and λ is the leverage or asymmetry parameter.
The usual ARCH (Engle, 1982) and GARCH (Bollerslev, 1986) models are
obtained if in equation (2.3) only the parameters ω , β (L) and ϕ (L) are included
and the power parameter η is fixed at 2. Alternatively, the power parameter η
can be estimated rather than imposed as 2, yielding the Power ARCH (or
PARCH) class proposed by Ding et al. (1993). When η is fixed at 2, equation
(2.3) gives a variant of the Threshold GARCH or GJR-GARCH as introduced
independently by Zakoïan (1994) and Glosten et al. (1993). Finally, equation (2.4)
represents the Exponential GARCH model (Nelson, 1991).
Since st is an unobserved variable, the conditional variance in (2.3) and (2.4)
depends on the entire sequence of regimes up to time t. This means that, for a
sample of length T, the likelihood function requires integrating over M T
sequences of (unobserved) regime paths rendering the model essentially

intractable and practically impossible to estimate. This problem is known as path

dependence in MS-GARCH models. Gray (1996) and Klaassen (2002) suggest
using the conditional expectation of the lagged variance as a proxy for lagged
variance. In other words, the conditional variance of lagged ε t is composed of all
component variances as well as the time-varying conditional regime
probabilities.3
The probability that the observed regime at time t is j evolves according to
the filtering (updating) equation:
f (rt | st = j ,ψ t −1 ; Λ). Pr{st = j | ψ t −1}

Pr{st = j | ψ t } = . (2.5)
∑
M
i =1
f (rt | st = i,ψ t −1 ; Λ). Pr{st = i | ψ t −1}
where f (rt | st = j ,ψ t −1 ; Λ ) denotes the (conditional) probability density of the

return at time t conditional on ψ t −1 and when regime j is operating. The vector Λ
comprises parameters in the conditional mean and variance equations and
parameters characterizing the conditional density distribution. Then the maximum
likelihood estimate of Λ is obtained by maximizing
L(θ ) = ∑t =1 f (rt | ψ t −1 ; Λ ) ,
T
where
f (rt |ψ t −1 ; Λ ) = ∑ j =1 f (rt | st = j ,ψ t −1 ; Λ ). Pr{st = j |ψ t −1} .

M
(2.6)
The key probability in (2.5) and (2.6) has a first-order recursive structure
which can be written as
Pr{st = j | ψ t −1} = ∑i =1 Pr{st −1 = i | ψ t −1} Pr{st = j | st −1 = i,ψ t −1}.

M
(2.7)
2.2 Long and short VaR

Suppose that, at the time index t, we are interested in the risk of a financial
position for the next l periods. The ΔV (l) , being the change in value of the
asset(s) in the financial position from time t to t+ l , is a random variable at t. The
VaR of a long position (left tail of the distribution function) over the time horizon
3
Haas et al. (2004) present a new MS-GARCH model to overcome the path dependence problem.
In their model the regime variances only depend on past shocks and their own lagged values.

l with probability p is defined as
p = Pr[ΔV (l) ≤ VaR] = Fl (VaR) , (2.8)
where F(x) denotes the cumulative distribution function, CDF, of ΔV (l) .

Alternatively,
VaRlp = Fl−1 ( p) , (2.9)
where F −1 is the so-called quantile function defined as the inverse of the CDF.
The VaR for a short position is similarly computed where the same definition is
used for the right tail of the distribution function, i.e. 1-p substitutes for p.4
Since quantiles are direct functions of the variance in parametric models, the
ARCH class models present a dynamic measure of VaR. More precisely, the VaR
for time T+1 based on the ARCH family models can be defined as
VaRTp+1 (r ) = μT +1 + hT1 /+21Fp−1 ( z ), (2.10)
where Fp−1 ( z ) denotes the pth quantile of the distribution of variance-adjusted

residuals in (2.2). μ t +1 and ht +1 are one-step forecasts of the conditional mean and
conditional variance, respectively. Equation (2.10) shows that the conditional
variance at time T+1 and the distribution chosen for F, the innovations in (2.2),
directly affect the level of the VaRTp+1 (r ) measure.
The empirical evidence suggests that it is crucial to consider the leverage
effect in forecasting stock market volatilities (see Nelson, 1991, among others).
An asymmetric response of VaR to positive and negative shocks can be modeled
in two ways: imposing an asymmetric parameter in the conditional variance
equation or imposing a skewness parameter in the distribution of error term.
While the latter approach leads to an asymmetric quantile, Fp−1 , the former leads to
a differential response of the conditional variance, σ T +1 , to bad and good news.
The skewed Student’s t distribution can be used to obtain an asymmetric quantile
in modeling VaR for long and short positions.
Lambert and Laurent (2001) show that the quantile function with such a
density is:
4
See, for example, Dowd (2005) for a comprehensive survey of VaR methods.

⎧ 1 p 1
⎪⎪ ζ St p ,ν [ 2 (1 + ζ )] p<
2
if
1+ ζ 2
SkSt *p ,ζ ,ν =⎨ (2.11)
1− p 1
⎪− ζ St p ,ν [ (1 + ζ −2 )] if p≥
⎪⎩ 2 1+ ζ 2
where ζ is the asymmetry coefficient and St p ,ν is the quantile function of the (unit
variance) Student’s t density with ν degrees of freedom. Then the associated
quantile function is obtained from
SkSt *p ,ζ ,ν − m
SkSt p ,ζ ,ν = , (2.12)
s
where parameters m and s 2 are the mean and the variance of the non-standardized
skewed Student’s t, respectively:
ν −1
Γ( ) ν −2
2 1 1
m= (ζ − ) , s 2 = (ζ 2 + − 1) − m 2 .
π Γ(ν / 2) ζ ζ 2
For skewed Student’s t innovations, F −1 in (2.10) for long and short

positions is given by SkSt p ,ν ,ζ and SkSt1− p ,ν ,ζ indicating p% quantiles, with
ν degrees of freedom and asymmetry coefficient ζ , on the left and right tails,
respectively. If ζ < 1 (or log(ζ ) < 0 ), | SkSt p ,ν ,ζ | > | SkSt1− p ,ν ,ζ | and the VaR for
long trading positions will be larger (for the same conditional variance) than the
VaR for short trading positions. The opposite result holds when ζ > 1 .
3 Empirical Findings
The class of model developed in the previous section enables us to consider
excess kurtosis and skewness as well as possible structural changes exhibited by
most financial time series data. In the following subsections we implement this
model to examine stock market behavior.
3.1 Data
The data set analyzed in this paper comprises daily observations on two major
stock market indices returns, namely, the FTSE100 and the S&P 500 (hereafter

the FTSE and S&P). The sample covers the time interval from 1 January 1991 to
31 December 2004, resulting in 3599 daily observations.5 The first 2599
observations are used for in-sample estimation while the remaining 1000
observations are taken as the out-of- sample for forecast evaluation process in
which a sliding window (rolling) method of 1000 days is implemented. The usual
descriptive statistics of the data are given in Table 1. The moments of the stock
index returns are shown along with the results of an aggregate autocorrelation
(Ljung-Box) test for returns and their squares.
Table 1: Moments of the FTSE100 and S&P500 returns along with

aggregate autocorrelation test results.
Ljung-Box test
Moments specifications Returns Squared returns
FTSE S&P FTSE S&P FTSE S&P
Mean 0.010 0.016 Q-stat(12) 55.76 15.28 1273 633
Minimum -4.654 -3.914 {0} {0.23} {0} {0}
Maximum 2.419 2.420 Q-stat(24) 81.47 32.94 1768 888
Std. Dev. 0.479 0.492 {0} {0.11} {0} {0}
Skewness -0.371 -0.237 Q-stat(36) 91.95 53.63 211 1049
Kurtosis 7.720 7.003 {0} {0.03} {0} {0}
p-values in curly braces.
Q-stat (q) denotes a modified Ljung-Box type statistic, which combines the first q
squared normalized autocorrelation estimates.
As can be seen, for both indices, the mean return is quite small, the
skewness is significant and negative, implying a possible leverage effect in data,
and the kurtosis is significantly higher than that of a Gaussian distribution (excess
kurtosis) indicating fat-tailed returns. This suggests the need for a fat-tailed or
skewed fat-tailed distribution, for example Student’s t or skewed Student’s t, to
describe the returns’ conditional distribution. In addition, the large Q-statistics up
to 12, 24 and 36 orders strongly reject the null hypothesis of no serial correlation
in both returns and squared returns for the FTSE but only in squared returns for
the S&P index.6
We also present, in Figure 1, squared returns, rt 2 , for the last four years in
order to give an indication of high and low volatility periods. At first glance, plots
demonstrate substantial volatility clustering as periods of low volatility mix with
periods of high volatility and large positive and negative returns. This indicates
5
All data have been obtained from DataStream.
6
Other test statistics like: the BDS, McLeod–Li, Engle LM, Tsay and Bicovariance tests can be
used to examine whether the residuals are i.i.d. (see for example Panagiotidis, 2005).

the potential benefits of allowing for conditional heteroskedasticity.7 More

precisely, the plots demonstrate two crucial points: the pressure relieving effect
and volatility persistence (clustering). The latter implies that individual shocks
sometimes have a long effect on subsequent volatility, while the former implies
that a shock sometimes is followed by a period of low instead of high volatility.
S&P500 FTSE100
6 6
5 5
4 4
3 3
2 2
1 1
0 0
2002 2003 2004 2005 2002 2003 2004 2005
Figure 1: Squared returns of the S&P500 and FTSE100 rates over the sample
period January 2001 to December 2004.
It seems that the standard single-regime ARCH models cannot capture the
pressure relieving effect, since they typically imply large persistence for
individual shocks. However, it is possible that regime-switching models allowing
for a switch from a high to a low volatility regime can explain both the large
volatility persistence of individual shocks and the pressure relieving effect.
Furthermore, imposing an asymmetric parameter in the conditional variance
equation and/or distributional form of error terms enables us to capture possible
leverage effects in the data. The following subsection examines these issues.
7
The formal test results for conditional heteroskedasticity are also available upon request.

3.2 Specification and diagnostic results

We estimate all models using the Maximum Likelihood (ML) method, assuming
normal (N), Student’s t (t) and skewed Student’s t (skt) innovations.8 Table 2
summarizes the estimation results from the single-regime models. The p-values
are associated with critical values corrected for skewness using a wild bootstrap
simulation.9 Using the Likelihood Ratio (LR) test and the Schwarz criterion,10 we
try different orders of the ARMA process for the conditional mean equations (not
shown here) and conclude that no dynamics is preferred to model the conditional
mean in our data (only a constant is included). The results also indicate that the
asymmetry parameters in both the conditional variance and the skewed Student’s t
innovations of the EGARCH specification are highly significant for both indices.
These results resemble those of Marcucci (2005) in modeling the S&P100 returns.
However, the asymmetry parameters are insignificant for both the FTSE and the
S&P returns with GARCH specifications. In other words, it seems that imposing
an asymmetry parameter in either the conditional variance equation or the
Student’s t distribution of the GARCH specification is unnecessary fully to model
the dynamics of our return series.
Regime-switching models
The ML estimation results assuming normal and Student’s t innovations for the
FTSE and the S&P returns are respectively shown in Tables 3-6. In order to find
out the appropriate MS structure for the returns series, three different MS models
are analyzed: a partial MS model with no dynamics in the mean equation, a partial
MS and a full MS model with appropriate dynamics (preferred dynamics as
indicated by the single-regime models) in the mean equation. In the full MS
model all parameters in the mean and variance equations are allowed to switch
between regimes while in the partial MS model only the variance equation
parameters differ across regimes.
According to the information criterion and the LR test, the null of the partial
MS structure can be rejected in favor of the full MS structure for the S&P returns
with Student’s t innovations. However, allowing the mean equation’s parameters
to differ across regimes (full MS model) results in some insignificant parameters
in the mean or variance equations. Consequently, it seems the partial MS model in
conjunction with the ARMA (1, 1) structure for the mean equation, with either
8
All estimations are performed in the TSMod package developed by James Davidson (see
http://www.timeseriesmodelling.com/)
9
See Arghyrou and Gregoriuo (2007) for a very recent application of the wild bootstrap technique.
10
We use the Schwarz Criterion, which provides consistent order-estimation in the context of
linear ARMA models (see Hannan, 1980).

Table 2: ML estimation results from the single-regime models.

ARCH specifications
FTSE100 S&P500
N t skt N t skt
Variance parameters:
GARCH Intercept: ω 1/ 2 0.044 0.072 0.072 0.059 0.049 0.048
[0.012] [0.018] [0.018] [0.020] [0.019] [0.021]
ARCH term: α 0.054 0.047 0.047 0.072 0.045 0.044
{0} {0} {0.002} {0} {0.001} {0.111}
GARCH term: β 0.954 0.949 0.949 0.935 0.949 0.948
{0} {0} {0.001} {0} {0} {0}
GARCH asymmetry: λ -0.198 0.129 0.132 0.133 0.576 0.655
{0.283} {0.515) {0.193} {0.205} {0.194} {0.076}
Power GARCH term: η 1.451 1.117 1.120 1.148 1.395 1.442
[0.312] [0.209] [0.212] [0.307] [0.508] [0.585]
Log skewness term: Ln(ζ ) -0.008 -0.047
{0.504} {0.452}
Student's t d. f. 6.492 6.508 5.165 5.309
Log Likelihood: -1439.11 -1368.96 -1368.92 -1487.23 -1388.24 -1386.78
Ljung-Box Q(12): 19.1 19.2 19.2 6.8 8.1 7.9
{0.085} {0.083} {0.083} {0.872} {0.778} {0.793}
EGARCH specifications
FTSE100 S&P500
N t skt N t skt
EGARCH Intercept: ω 1 / 2 3.686 3.744 3.743 3.439 3.447 3.441
[0.63] [0.60] [0.60] [0.57] [0.48] [0.48]
ARCH term: α 0.063 0.062 0.062 0.140 0.126 0.127
{0} {0} {0} {0} {0} {0}
GARCH term: β 0.976 0.976 0.976 0.941 0.949 0.948
{0} {0} {0} {0} {0} {0}
EGARCH asymmetry: λ -0.293 -0.593 -0.607 -0.196 -0.314 -0.354
{0} {0} {0} {0} {0} {0.045}
Log skewness term: Ln(ζ ) -0.021 -0.053
{0} {0}
Student's t d. f. 6.274 6.306 4.684 4.773
Log Likelihood: -1465.53 -1380.85 -1380.58 -1550.04 -1437.97 -1436.12
Ljung-Box Q(12): 24.987 22.63 22.48 21.346 23.60 23.12
{0.015} {0.031} {0.032}{0.046} {0.023} {0.027}
Standard errors are given in square brackets and p-values in curly braces.
The ML results assuming Normal, Student’s t and skewed Student’s t for the error terms are
presented in the columns labeled N, t and skt, respectively. The upper part of the table reports the
estimated coefficients for the ARCH specifications, while the lower part reports those for the
EGARCH specifications. Q(12) indicates the aggregate autocorrelation test for the squared
normalized residuals up to lag 12.

normal or Student’s t, can be considered as the appropriate dynamics for the S&P
returns, while no dynamics is preferred for the FTSE returns. We note that
Marcucci (2005) and Ane and Ureche-Rangau (2006) do not compare the partial
and full MS and instead use the full MS structure in their studies. By contrast,
Haas et al. (2005) implement the partial MS structure in analyzing foreign
exchange data.
Table 3: ML estimation results from the two-regime GARCH models with normal
innovations for the FTSE100 returns.
Mean parameters:
Intercept: γ s 0.020 0.020 -0.035 0.027
{0.006} {0.010} {0.457} {0.003}
AR1: θ s 0.045 0.084 0.041
{0.015} {0.601} {0.063}
GARCH Intercept: ω 1s / 2 0.048 0.000 0.049 0.000 0.048 0.000
[0.009] [0] [0.009] [0] [0.009] [0]
ARCH term: α s 0.029 0.083 0.029 0.096 0.030 0.100
{0} {0} {0} {0} {0} {0}
GARCH term: β s 0.939 0.971 0.939 0.967 0.939 0.966
{0} {0} {0} {0} {0} {0}
Stay probabilities: ( p11 , p22 ) 0.817 0.001 0.823 0.001 0.827 0.001
Test summary:
Log Likelihood -1375.44 -1373.25 -1372.29
Schwarz Criterion -1406.86 -1408.60 -1419.42
Ljung-Box Q(12) 16.673 15.879 16.986
{0.162} {0.197} {0.15}
Standard errors are given in square brackets and p-values in curly braces.
The upper and middle parts of the table report the estimated coefficients for the mean and variance
equations. In each pair of columns, left and right columns report parameter estimates for regime 1
and 2, respectively. Non-switching parameters are reported under regime 1 (left columns). The
lower part reports diagnostic test results, where the second row shows the increase in the Log-
likelihood value compared to the one for the corresponding single-regime model. For each of the
criteria, boldface entries indicate the best regime switching model for the particular criterion. Q
(12) indicates the aggregate autocorrelation test for the squared normalized residuals up to lag 12.
Tables 3-6 also present the results of the Q-test to test for autocorrelation in
the standardized residuals. The p-values are high for both series, indicating that
the models are rich enough to remove all traces of autocorrelation in the
normalized residuals up to lag 12. Finally, our results (available upon request)
indicate that the skewness parameter in the skewed Student’s t distribution is
insignificant for both series, similar to the single-regime GARCH models. These

results contrast with those of Ane and Ureche-Rangau (2006) who find significant
asymmetry in an MS-APARCH specification for different (Asian) stock market
indices.
Table 4: ML estimation results from the two-regime GARCH models with

Student’s t innovations for the FTSE100 returns.
Mean parameters:
Intercept: γs 0.021 0.021 -0.015 0.022
{0.004} {0.006} {0.007} {0.005}
AR1: θs 0.039 -0.918 0.051
{0.031} {0} {0.006}
[0.009] [0.033] [0.008] [0.022] [0.027] [0.010]
ARCH term: α s 0.007 0.043 0.007 0.044 0.001 0.034
{0} {0} {0.006} {0.013} {0.444} {0}
GARCH term: β s 0.980 0.927 0.979 0.925 0.911 0.956
{0} {0} {0} {0} {0} {0}
Student's t d. f. 6.896 6.863 5.331
Test summary:
Log Likelihood -1366.66 -1364.77 -1363.12
Ljung-Box Q(12) 15.024 15.453 17.791
{0.24} 0.218} {0.122}
See the legend of Table 3 for explanations.
3.3 Comparing single-regime and MS models

Unlike single-regime models, MS-GARCH models distinguish two sources of
volatility persistence to capture the clustering of large changes as well as the
pressure relieving effect (see Klaassen, 2002). These are within-regime volatility
persistence with different unconditional variances and regime shifts with different
periods of persistence (different regime persistence). High (within-regime)
volatility persistence manifests itself in a significant large persistence term, as
measured by the sum αˆ + βˆ , similar to the single-regime model. This implies that
the effect of an individual shock takes a long time to dissipate. The persistence of
regimes can be illustrated by p11 and p22 which are usually referred to as the
staying probabilities of regimes. The expected duration of regimes is also utilized
to get a better idea about regime persistence.

Table 5: ML estimation results from the two-regime GARCH models with normal
innovations for the S&P500 returns.
Mean parameters:
Intercept: γ s 0.031 0.032 -0.381 0.044
{0} {0} {0.546} {0}
AR1: θ s 0.795 0.752 0.757
{0} {0.058} {0}
MA1: φ s 0.839 0.320 0.822
{0} {0} {0}
[0.011] [0.070] [0.012] [0.065] [0] [0.007]
ARCH term: α s 0.027 0.271 0.027 0.330 0.130 0.034
{0.024} {0} {0.006} {0} {0.506} {0}
GARCH term: β s 0.953 0.896 0.954 0.870 0.942 0.946
{0} {0} {0} {0} {0} {0}
Test summary:
Log Likelihood -1404.84 -1396.51 -1387.38
Ljung-Box Q(12) 8.96 9.567 10.675
{0.706} {0.654} {0.557}
Table 7 presents the unconditional probabilities, π j , the expected durations,

δ j , and the unconditional variances, Eσ 2jt , for the preferred MS-GARCH models.
Assuming normal or Student’s t innovations leads to different regime-switching
structures in our data. The first column of Table 7 assumes normal innovations.
The unconditional probability, π 1 , of being in the first (lower volatility) regime is
85% and 80% with expected duration of 6 and 4 trading days for the FTSE and
S&P returns, respectively. The unconditional probability of being in the second
(high-volatility) regime is 15% and 20% for the FTSE and S&P, respectively,
with an expected duration of around one day for both series. Thus the low
volatility periods are generally longer lasting. This is known as the mean-
reverting phenomenon and is first addressed by Dueker (1997) in equity markets.
This result is consistent with Tables 3 and 5 where αˆ 2 + β̂ 2 > 1 (1.05 for the
FTSE and 1.17 for the S&P returns) indicates that the process is non-stationary in
high volatility periods. However, the probability of staying in this non-stable
regime ( p22 ) is small for both series. Consequently, as noted by Yang (2000), the

process endogenously collapses back from its explosive state en route to a stable
regime and is stationary in the long run.
Table 6: ML estimation results from the two-regime GARCH models with

Student’s t innovations for the S&P500 returns.
Mean parameters:
Intercept: γ s 0.031 0.032 0.052 -2.941
{0} {0} {0} {)}
AR1: θ s 0.795 0.760 0.953
{0} {0} {0}
MA1: φ s 0.837 0.830 0.526
{0} {0} {0.004}
[0.014] [0.016] [0.013] [0.015] [0.012] [0.099]
ARCH term: α s 0.039 0.043 0.038 0.042 0.048 0.014
{0.005} {0} {0.005} {0} {0} {0.527}
GARCH term: β s 0.926 0.926 0.927 0.929 0.946 0.972
{0} {0} {0} {0} {0} {0}
Student's t d. f. 5.331 5.309 5.565
Test summary:
Log Likelihood -1388.45 -1382.03 -1373.20
Ljung-Box Q(12) 10.493 9.970 6.062
{0.573} {0.619} {0.913}
The second column of Table 7 for the Student’s t case indicates persistence
in the two regimes with a staying probability, p11 and p22 , both exceeding 0.99.
The regimes are also characterized by different unconditional variances, Eσ 2jt : the
unconditional variances in the high-volatility regime are about twice as large as
those in the low-volatility regime. Consequently, the degree of persistence due to
the Markov effects is close to one, i.e. p11 + p22 − 1 ≈ 1 . As noted by Timmermann
(2000) and Morana (2002), these values of the staying probabilities representing
infrequent mixing of regimes may be interpreted as closely resembling structural
break models. In this case, estimates of the GARCH parameters from models
ignoring the switching may be overwhelmed by substantial upward bias.

Table 7: Regime switching properties in FTSE100 and S&P500 markets.

FTSE100 S&P500
Normal Student’s t Normal Student’s t
P11 , P22 0.817 0.001 0.998 0.999 0.771 0.103 0.998 0.999
π1 , π 2 0.845 0.155 0.420 0.580 0.797 0.203 0.243 0.757
δ1 , δ 2 5.471 1.001 541 746 4.370 1.114 617 1923
Eσ 1t2 , Eσ 2t2 0.115 0.503 0.176 0.262 0.301 1.371 0.156 0.288
The table shows the unconditional properties of estimated MS-GARCH models with normal and
Student’s t innovations. Pjj , j=1, 2, are the staying probabilities and give the probability that state j
will be followed by state j . π j , j=1, 2, are the unconditional probabilities of being in regime j, that
is π j = (1 − Pii ) /(2 − Pii − Pjj ) , j=1, 2 and i ≠ j . δ j , j=1 2, are the expected duration times for
regime j, that is δ j = 1 /(1 − Pjj ) . Eσ 2jt , j=1, 2, denotes the unconditional expectation of the variance.
The above results resemble those in which regime-switching GARCH

models are implemented in modeling the dynamics of stock returns. For instance
they are similar to those of Marcucci (2005) for the S&P100 and of Ane and
Ureche-Rangau (2006) for Asian stock market indices.
Finally we note that the FTSE 100 moved from an auction to an electronic
trading system on October 10, 1997. This may affect the regime change and
leverage effects in our FTSE100 data. Therefore, the models are also estimated
for the pre- and post- October 10, 1997 periods. The results indicate (available
upon request) that the data follows similar structures in the sub-periods as well.
The asymmetry coefficient for the sub-period before (after) October 10, 1997 is
smaller (bigger) than the estimated coefficient for the whole sample (in absolute
value). Therefore, it seems that under an electronic trading system the market
reacts more strongly to bad news as compared to the previous trading system.
Furthermore, the MS results reveal that the FTSE100 market follows a mean
reverting feature under the old trading system while it exhibits a structural break
feature after October 10, 1997. Using the rolling method in the next subsection,
we consider these effects in evaluating the out-of-sample performance of models
in forecasting VaR.
4 Performance of Models in Forecasting VaR

The diagnostic tests in the previous section show that standard econometric tests
for model specification may not be appropriate for choosing the best model
among different GARCH models. In particular, as demonstrated by Hansen
(1996) and McLachlan and Peel (2000), the standard likelihood test cannot be

employed for testing the single-regime versus the MS model.11 Furthermore,

Sarma et al. (2003) show that different methodologies can yield different VaR
measures for the same portfolio and can sometimes lead to significant errors in
risk measurement.
One can compare the out-of-sample performance in forecasting VaR by the
competing models to overcome the above problem. The approach focusing on the
past performance of VaR models is referred to as backtesting which checks
whether a model’s risk estimates are consistent with its assumptions. Furthermore,
special attention is devoted to the validation of internal risk assessment models
within the Basle Accord (1996) framework.
Having considered alternative methods for backtesting VaR,12 we utilize a
set of exceptions and regulatory-based backtesting methods to evaluate the
performance of our competing VaR models. The first set comprises the
Christoffersen (2003) LR test and the Hurlin and Tokpavi (2006) Multivariate
Portmanteau (MP) test. The second set is based on the traffic light regulations
proposed by the Basle Committee on Banking Supervision (1996). While the first
set (the LR and MP tests) evaluates the statistical accuracy of the competing VaR
models, the regulatory-based backtest measures the loss to the economic agent
using the model.
Based on the diagnostic results in the preceding section, we compare three
different groups of VaR models: the GARCH, EGARCH and MS-GARCH
models. For each group, three types of innovations are considered: the Normal,
Student’s t and skewed Student’s t, resulting in nine GARCH family VaR models.
The models are used to estimate one day ahead VaR of both long and short
trading positions (left and right tails of returns distribution) with different
probabilities (at different tail quantiles): 0.5%, 1%, 5%, 95%, 99% and 99.5%.
4.1 Exceptions-based backtesting results

A rolling window method of 1000 days is used to estimate the daily VaR for each
model. Thus the indicator variable I t contains the last 1000 (2001 t0 2004) hit
sequences of the VaR violations. The motivation behind the rolling window
technique is to consider dynamic time-varying characteristics of the data in
different time periods.
The results of Christoffersen’s LR test for the FTSE and the S&P are
summarized in Tables 8 and 9, respectively. The number of failures is shown
along with p-values for an unconditional coverage test, LRuc , an independence
11
It is also demonstrated by Dacco and Satchell (1999) that the evaluation of forecasts from non-
linear models like regime-switching models based on statistical measures might be misleading.
12
See Dowd (2005) for a survey of backtesting VaR models.

test, LRind , and a conditional coverage test, LRcc . It is obvious that there are some
big discrepancies between the number of failures for the long and short positions
(hereafter f l and f s , respectively) obtained from the symmetric models (normal
and Student’s t innovations). For the FTSE returns f l is higher than f s at all
levels of VaR, and vice-versa for the S&P returns. This can be considered as clear
evidence of the asymmetry in our returns data.
The specification results showed it was not necessary to impose a skewness
parameter in the Student’s t distribution in modeling the distribution of our returns
series. We check whether the skewed Student’s t may be able to improve the out-
of-sample results for both the negative and positive returns. In fact, it is expected
that skewed Student’s t innovations with ζ < 1 can decrease f l and increase f s ,
compared to f l and f s obtained from a symmetric Student’s t. The opposite result
holds when ζ > 1 .
Tables 8 and 9 indicate that, in general, the skewed Student’s t improves on
the out-of-sample performance of the corresponding GARCH and MS-GARCH
models with Student’s t innovations. The overall improvement for both negative
and positive returns is 50% and 42% for the FTSE and S&P returns,
respectively.13
On the other hand, the EGARCH specification with symmetric innovations
generally leads to an acceptable performance for out-of-sample VaR prediction.
Consequently, the EGARCH model with skewed Student’s t innovations
generally fails to improve on the number of failures compared to those obtained
with symmetric Student’s t innovations.
The results unsurprisingly show that the VaR models based on normal
innovations have difficulties in modeling large returns. In particular, the normal
MS-GARCH model consistently underestimates the return (risk) of both series at
different tails, specifically at the 0.5% and 99.5% tails. In other words, failure
numbers are much greater than the expected one at a given quantile, in the case of
normal innovations. This leads to low p-values for the LRuc test, indicating an
insignificant model for volatility forecasting and the VaR estimation. On the other
hand, the models with Student’s t innovations perform very well, irrespective of
the model and the tail one takes into account. Thus, the LR test results show that a
switch from normal to Student’s t innovations yields a significant improvement in
the VaR performance.
13
“Improvement” implies to the situation where the number of failures obtained from an
asymmetric model is closer to the expected one, compared with the corresponding symmetric
model.

Table 8: LR test results for different confidence level of VaR for the FTSE100
returns.
p-values NF p-values
VaR models NF
LRcc LRind LRcc LRcc LRind LRcc
VaR 5% VaR 95%

GARCH-N 54 0.562 0.960 0.844 45 0.465 0.039 0.092
GARCH-t 55 0.470 0.571 0.656 48 0.776 0.028 0.085
GARCH-skt 53 0.661 0.479 0.707 50 1.000 0.260 0.531
EGARCH-N 56 0.389 0.301 0.404 56 0.389 0.469 0.530
EGARCH-t 56 0.389 0.301 0.404 55 0.470 0.508 0.619
EGARCH-skt 56 0.389 0.620 0.610 59 0.201 0.362 0.292
MS-GARCH-N 63 0.068 0.987 0.190 53 0.661 0.015 0.047
MS-GARCH-t 58 0.254 0.721 0.490 53 0.661 0.194 0.390
MS-GARCH-skt 57 0.316 0.670 0.553 52 0.768 0.214 0.443
VaR 1% VaR 99%

GARCH-N 14 0.229 0.528 0.398 9 0.749 0.686 0.875
GARCH-t 11 0.752 0.621 0.842 6 0.171 0.788 0.377
GARCH-skt 11 0.752 0.621 0.842 8 0.512 0.719 0.756
EGARCH-N 19 0.011 0.391 0.027 8 0.512 0.719 0.756
EGARCH-t 13 0.360 0.558 0.555 8 0.512 0.719 0.756
EGARCH-skt 12 0.536 0.589 0.713 8 0.512 0.719 0.756
MS-GARCH-N 18 0.022 0.416 0.052 13 0.360 0.558 0.555
MS-GARCH-t 11 0.752 0.621 0.842 8 0.512 0.719 0.756
MS-GARCH-skt 12 0.536 0.589 0.713 9 0.749 0.686 0.875
VaR 0.5% VaR 99.5%

GARCH-N 11 0.020 0.621 0.060 3 0.334 0.893 0.621
GARCH-t 7 0.397 0.753 0.664 2 0.126 0.929 0.309
GARCH-skt 8 0.215 0.719 0.435 3 0.334 0.893 0.621
EGARCH-N 9 0.107 0.686 0.251 8 0.215 0.719 0.435
EGARCH-t 4 0.644 0.858 0.884 5 1.000 0.823 0.975
EGARCH-skt 2 0.126 0.929 0.310 6 0.662 0.788 0.877
MS-GARCH-N 11 0.020 0.621 0.060 6 0.662 0.788 0.877
MS-GARCH-t 10 0.048 0.653 0.129 5 1.000 0.823 0.975
MS-GARCH-skt 10 0.048 0.653 0.129 2 0.126 0.929 0.310
The table shows the backtesting results from the Likelihood Ratio test of Christoffersen (1998).
P-values for unconditional coverage, LRuc , independence, LRind , and conditional coverage,
LRcc , tests along with the number of failures, N F , for both long and short positions are
reported in the left and right panel of the table, respectively. The models are successively the
GARCH, EGARCH and MS-GARCH specifications, where N, t and skt denote normal,
Student’s t and skewed Student’s t error terms, respectively.

Table 9: LR test results for different confidence level of VaR for the S&P500
returns.
p-values p-values
VaR models NF NF
LRcc LRind LRcc LRcc LRind LRcc
VaR 5% VaR 95%

GARCH-N 43 0.302 0.042 0.074 46 0.561 0.931 0.842
GARCH-t 48 0.776 0.030 0.091 51 0.879 0.800 0.957
GARCH-skt 46 0.561 0.020 0.055 55 0.470 0.508 0.619
EGARCH-N 47 0.666 0.252 0.472 57 0.316 0.126 0.188
EGARCH-t 47 0.666 0.596 0.791 56 0.389 0.141 0.233
EGARCH-skt 45 0.465 0.502 0.611 56 0.389 0.141 0.233
MS-GARCH-N 66 0.026 0.423 0.061 64 0.050 0.224 0.070
MS-GARCH-t 58 0.254 0.165 0.199 59 0.201 0.100 0.115
MS-GARCH-skt 53 0.661 0.076 0.189 65 0.036 0.503 0.090
VaR 1% VaR 99%

GARCH-N 9 0.749 0.056 0.175 11 0.752 0.621 0.842
GARCH-t 4 0.030 0.858 0.094 5 0.079 0.823 0.209
GARCH-skt 4 0.030 0.858 0.094 7 0.315 0.753 0.575
EGARCH-N 8 0.512 0.719 0.756 12 0.536 0.589 0.713
EGARCH-t 6 0.171 0.788 0.377 9 0.749 0.686 0.875
EGARCH-skt 6 0.171 0.788 0.377 11 0.752 0.621 0.842
MS-GARCH-N 22 0.001 0.091 0.001 16 0.079 0.470 0.165
MS-GARCH-t 12 0.536 0.130 0.263 12 0.536 0.589 0.713
MS-GARCH-skt 11 0.752 0.106 0.258 16 0.079 0.470 0.165
VaR 0.5% VaR 99.5%

GARCH-N 4 0.644 0.858 0.884 5 1.000 0.823 0.975
GARCH-t 4 0.644 0.858 0.884 2 0.126 0.929 0.310
GARCH-skt 4 0.644 0.858 0.884 3 0.334 0.893 0.621
EGARCH-N 7 0.397 0.753 0.665 4 0.644 0.858 0.884
EGARCH-t 3 0.334 0.893 0.621 2 0.126 0.929 0.310
EGARCH-skt 3 0.334 0.893 0.621 3 0.334 0.893 0.621
MS-GARCH-N 11 0.020 0.106 0.018 11 0.020 0.621 0.060
MS-GARCH-t 6 0.662 0.788 0.877 7 0.397 0.753 0.664
MS-GARCH-skt 4 0.644 0.858 0.884 10 0.048 0.653 0.129

We also examine the statistical accuracy of the VaR models using Hurlin
and Tokpavi’s (2006) MP test. The results for the FTSE and S&P return are
reported in Tables 10 and 11, respectively. Following their suggestion concerning
the choice of lag order K and number of coverage rates m, the following sets are
considered for each VaR model, k ∈ {1,3,5} , Θ ={5%, 1%} and Θ ={5%, 0.5%}
for m = 2 and Θ = {5%, 1%, 0.5%} for m = 3. The p-values corresponding to
Qm (k ) for long and short positions are reported in the left and right panel of the
tables, respectively. Consider the total number of violations of the no
autocorrelation null. The MS-GARCH models with a total of just 6 (out of a total
of 54 possible) violations outperform the single-regime GARCH and EGARCH
models with 13 and 16 violations, respectively, for the FTSE 100. However, the
EGARCH model (8 violations) outperforms the MS-GARCH (14 violations) and
single-regime GARCH (19 violations) for the S&P500.
Comparing the MP test results with those of the LR test, we conclude that
the MS-GARCH models outperform single-regime models as there is no case in
which the p-values for all lag orders are less than 5%. The only small exception
are the results on the left tail of the S&P returns with Θ = {5%, 1%, 0.5%} where
the MS-GARCH model with skewed Student’s t innovations is rejected at all lag
orders. On the other hand, the single-regime models are more likely to be rejected
by the MP test, especially those with large lag orders and high coverage rates. For
instance, the null of no autocorrelation in the VaR violation sequences left by
GARCH and EGARCH models is rejected when the MP test with the coverage set
Θ = {95%, 99%, 99.5%} and lags orders k=3,5 is implemented for the FTSE
returns. The same result is achieved for the GARCH model in forecasting VaR for
the S&P returns with the coverage set Θ = {5%, 1%, 0.5%} at all lag orders, k=1,
3, 5.
Overall, the MS-GARCH-skt model is favored by our exceptions-based tests
in forecasting of both the long and short VaR for FTSE returns. The same results
hold for the MS-GARCH-t model in the case of S&P returns. Our findings are an
improvement on those in Marcucci (2005) in which no model clearly outperforms
the others in forecasting the long VaR of the S&P100 returns.
4.2 Regulatory-based backtesting results

We implement the Basle traffic light regulation to compute the capital
requirements imposed by the previously introduced VaR methods using our 1000
daily VaR numbers previously estimated for exceptions-based backtesting. Since
the multiplication factor is determined based on the number of exceptions over the
previous 250 trading days, our regulatory-based backtesting sample contains the
last 751 daily capital requirements imposed by each VaR model.

Table 10: MP test results for different multivariate coverage rates of VaR for the
FTSE100 returns.
Θ ={5%, 1%} Θ ={95%, 99%}
VaR models
K=1 K=3 K=5 K=1 K=3 K=5
GARCH-N 0.61 0.08 0.05 0.70 0.11 0.32
GARCH-t 0.47 0.19 0.11 0.64 0.00 0.01
GARCH-skt 0.38 0.13 0.06 0.88 0.10 0.35
EGARCH-N 0.38 0.12 0.04 0.91 0.14 0.38
EGARCH-t 0.77 0.16 0.14 0.89 0.17 0.42
EGARCH-skt 0.77 0.11 0.10 0.89 0.13 0.38
MS-GARCH-N 0.80 0.64 0.37 0.54 0.35 0.70
MS-GARCH-t 0.58 0.29 0.06 0.83 0.22 0.46
MS-GARCH-skt 0.79 0.07 0.04 0.85 0.92 0.98
Θ ={5%, 0.5%} Θ ={95%, 99.5%}

K=1 K=3 K=5 K=1 K=3 K=5
GARCH-N 0.75 0.28 0.13 0.70 0.51 0.90
GARCH-t 0.71 0.03 0.09 0.63 0.27 0.74
GARCH-skt 0.57 0.05 0.10 0.90 0.66 0.95
EGARCH-N 0.57 0.02 0.01 0.93 0.14 0.38
EGARCH-t 0.61 0.01 0.04 0.95 1.00 0.99
EGARCH-skt 0.93 0.00 0.00 0.92 0.99 0.98
MS-GARCH-N 0.76 0.25 0.09 0.54 0.00 0.00
MS-GARCH-t 0.66 0.25 0.03 0.85 0.83 0.94
MS-GARCH-skt 0.82 0.02 0.07 0.87 0.39 0.81
Θ ={5%, 1%, 0.5%} Θ ={95%, 99%, 99.5}

K=1 K=3 K=5 K=1 K=3 K=5
GARCH-N 0.97 0.66 0.55 0.99 0.00 0.01
GARCH-t 0.94 0.45 0.01 0.98 0.00 0.00
GARCH-skt 0.90 0.55 0.00 1.00 0.00 0.00
EGARCH-N 0.78 0.00 0.00 0.00 0.00 0.00
EGARCH-t 0.94 0.10 0.10 1.00 0.00 0.00
EGARCH-skt 0.99 0.00 0.00 1.00 0.00 0.00
MS-GARCH-N 0.93 0.17 0.07 0.96 0.05 0.55
MS-GARCH-t 0.95 0.92 0.77 1.00 0.79 0.99
MS-GARCH-skt 0.99 0.49 0.00 1.00 0.97 1.00
This table shows the p-values for the MP test of Hurlin and Tokpavi (2006). The models are
successively the GARCH, EGARCH and MS-GARCH specifications, where N, t and skt denote
normal, Student’s t and skewed Student’s t error terms, respectively. For each model, the sliding
window (rolling) method with a size of 1000 days is implemented to estimate the one day ahead
VaR. Θ denotes discrete set of coverage rate in testing the null hypothesis corresponding to the
joint null by the autocorrelations of order 1 in k = 1,3,5, for the hit sequences of VaR violations.

Table 11: MP test results for different multivariate coverage rates of VaR for the
S&P500 returns.
Θ ={5%, 1%} Θ ={95%, 99%}
VaR models
K=1 K=3 K=5 K=1 K=3 K=5
GARCH-N 0.00 0.01 0.01 0.83 0.16 0.45
GARCH-t 0.05 0.11 0.20 0.91 0.09 0.28
GARCH-skt 0.03 0.06 0.13 0.93 0.38 0.75
EGARCH-N 0.04 0.33 0.68 0.76 0.04 0.15
EGARCH-t 0.55 0.87 0.91 0.78 0.01 0.08
EGARCH-skt 0.55 0.92 0.97 0.78 0.12 0.45
MS-GARCH-N 0.09 0.27 0.24 0.79 0.55 0.78
MS-GARCH-t 0.03 0.34 0.35 0.73 0.23 0.56
MS-GARCH-skt 0.01 0.05 0.00 0.90 0.47 0.80
Θ ={5%, 0.5%} Θ ={95%, 99.5%}

K=1 K=3 K=5 K=1 K=3 K=5
GARCH-N 0.05 0.16 0.26 0.97 0.80 0.95
GARCH-t 0.04 0.10 0.18 0.98 0.59 0.81
GARCH-skt 0.02 0.06 0.12 0.96 0.48 0.85
EGARCH-N 0.02 0.25 0.59 0.78 0.21 0.40
EGARCH-t 0.92 0.98 1.00 0.80 0.00 0.00
EGARCH-skt 0.88 0.99 1.00 0.80 0.01 0.10
MS-GARCH-N 0.02 0.07 0.20 0.85 0.45 0.78
MS-GARCH-t 0.38 0.85 0.53 0.75 0.00 0.00
MS-GARCH-skt 0.14 0.22 0.02 0.90 0.14 0.40
Θ ={5%, 1%, 0.5%} Θ ={95%, 99%, 99.5}

K=1 K=3 K=5 K=1 K=3 K=5
GARCH-N 0.00 0.00 0.00 1.00 0.00 0.01
GARCH-t 0.00 0.00 0.00 1.00 0.08 0.59
GARCH-skt 0.00 0.00 0.00 1.00 0.93 1.00
EGARCH-N 0.20 0.95 1.00 0.99 0.43 0.88
EGARCH-t 0.64 1.00 1.00 0.99 0.00 0.07
EGARCH-skt 0.58 1.00 1.00 0.99 0.16 0.75
MS-GARCH-N 0.06 0.21 0.06 1.00 0.00 0.06
MS-GARCH-t 0.00 0.11 0.16 0.99 0.02 0.33
MS-GARCH-skt 0.00 0.01 0.00 0.95 0.01 0.13

Figure 2 shows the capital requirement for the FTSE and SP imposed by
three selected GARCH, EGARCH-t and MS-GARCH models with Student’s t
innovations. One striking feature is that the capital requirements of these three
VaR models display a very similar pattern. However, the MS-GARCH models
usually impose lower daily capital requirements, compared to the single-regime
models. This helps risk managers avoid over-conservative estimation of VaR and
so save on their minimum capital requirements.
FTSE100
25 25
20 20
15 15
10 10
Daily c apital requirem ents
5 5
0 200 400 600 0 200 400 600
S&P500
25 25
20 20
15 15
10 10
5 5
0 200 400 600 0 200 400 600
GARCH-t EGARCH-t MS-GARCH-t
Figure 2: Comparison of capital requirement imposed by different

VaR models over 751 days for the left and right tails of the FTSE100
and S&P500 returns.
Tables 12 and 13 summarize the results of regulatory-based backtesting for

the FTSE and the S&P data, respectively. The numbers under the green, yellow
and red columns indicate how many times during the 751 days each model has
been placed in that particular zone by the Basle traffic light regulation. The
average daily capital requirement is also reported along with its variance over the
sample period. We implement the Hansen (2005) Superior Predictive Ability
(SPA) test to compare the performance of the VaR models in terms of regulatory-
based backtesting. The last column of the Tables presents the ranking between
comparable models based on the SPA test, where the benchmark model is

GARCH-t. The ‘best’ model is that which has the most significant performance
relative to the benchmark model. The other four pair-wise comparisons are those
models with a performance that corresponded to the 75% (second best), 50%
(median), 25% (second worst) and 0% (worst) quantile of model performance.
Table 12: Results of regulatory-based backtesting over 751 days for the FTSE100.
Areas Capital Requirements
VaR Models
Green Yellow Red Average Variance SPA Rank
Long position
GARCH-N 660 91 0 11.97 14.06 Median
GARCH-t 751 0 0 12.30 14.32 Benchmark
GARCH-skt 751 0 0 12.50 14.01
EGARCH-N 452 299 0 12.57 25.97 Worst
EGARCH-t 733 18 0 11.69 15.34 Second best
EGARCH-skt 751 0 0 12.01 16.08 Second worst
MS-GARCH-N 480 271 0 11.84 19.81
MS-GARCH-t 751 0 0 11.49 14.59 Best
MS-GARCH-skt 699 52 0 11.62 14.42
Short position
GARCH-N 619 132 0 11.72 13.02 Worst
GARCH-skt 751 0 0 11.70 14.73 Second worst
EGARCH-N 741 10 0 11.12 12.33 Best
EGARCH-t 741 10 0 11.47 14.99
EGARCH-skt 741 10 0 11.17 14.27
MS-GARCH-N 452 299 0 11.36 11.99
MS-GARCH-t 751 0 0 11.18 14.19 Second best
MS-GARCH-skt 741 10 0 11.31 15.20 Median
This table summarizes the capital requirements imposed by different VaR models. For each
model, the numbers under the green, yellow and red columns indicate how many times during
the 751 days the model has been placed in that particular zone by the Basle traffic light. Average
daily capital requirement is also reported along with its variance over the sample period. The last
column reports the ranking between comparable models based on the Superior Predictive Ability
(SPA) approach, where the benchmark model is GARCH-t. The ‘best’ model is that model had
the most significant performance relative to the benchmark model. The other four pair-wise
comparisons are those models with a performance that corresponded to the 75% (second best),
50% (median), 25% (second worst) and 0% (worst) quantile of model performance. There is no
ranking for the models which have some red zone record, as a placing in the red zone implies a
problem within the VaR model. This is also the case for those models which have weak
performance in the exception-based testing.

Table 13: Results of regulatory-based backtesting over 751 days for the
S&P500.
Areas Capital Requirements
VaR Models
Green Yellow Red Average Variance SPA Rank
Long position
GARCH-N 749 2 0 11.23 10.24
GARCH-skt 751 0 0 12.25 13.13 Worst
EGARCH-N 749 2 0 10.81 10.34 Second best
EGARCH-t 751 0 0 11.35 12.56 Median
EGARCH-skt 751 0 0 11.55 14.89 Second worst
MS-GARCH-N 352 396 3 11.73 20.83
MS-GARCH-t 734 17 0 10.76 12.71 Best
MS-GARCH-skt 746 5 0 10.95 11.62
Short position
GARCH-N 504 247 0 11.59 16.42 Worst
GARCH-skt 720 31 0 11.26 12.00 Second worst
EGARCH-N 699 52 0 10.57 10.81 Best
EGARCH-t 751 0 0 11.01 12.54
EGARCH-skt 751 0 0 10.80 12.81 Second best
MS-GARCH-N 435 287 29 11.10 19.09
MS-GARCH-t 504 247 0 11.17 19.63 Median
MS-GARCH-skt 435 287 29 10.87 17.90
Overall, the MS-GARCH-t model is ranked best for long positions in both
indices whereas EGARCH specifications are best for short positions. However,
the actual difference in performance between both sets of models is not large in
economic terms. This means that the null hypothesis of the SPA test (the
benchmark model imposes minimum capital requirements) is strictly rejected
when the benchmark is one of the standard GARCH models. This result support
those found in earlier studies. Marcucci (2005) obtains similar results with the
S&P100 data, finding that MS-GARCH and EGARCH specifications perform
better than the GARCH specification.
5 Conclusions
It is usually found that the GARCH family of models is a good candidate for
estimating conditional VaR over short-term time horizons. In this paper, we
extend this analysis to take account of both skewness and regime changes in

forecasting VaR for long and short positions. The empirical study shows the MS-
GARCH models lead to considerable improvements in correctly forecasting one-
day-ahead VaR for long and short positions of the FTSE100 and S&P500 indices.
We use a novel combination of exceptions and regulatory-based backtesting
procedures to evaluate out-of-sample model performance. This indicates that the
MS-GARCH-t clearly outperforms other models in estimating the VaR for both
long and short positions of the FTSE returns data. The exceptions-based (LR and
MP) tests indicate that it is an acceptable VaR model and it imposes lower capital
requirements according to the regulatory-based test. By contrast, the MS-
GARCH-t and EGARCH-t models outperform others in the case of the S&P
returns data. Imposing lower capital requirements, on either short or long position,
they are acceptable VaR models for both long and short positions. Furthermore,
analogous to the findings of Giot and Laurent (2003), the LR test confirms that
assuming skewed Student’s t innovations improves on the out-of-sample
performance of the corresponding GARCH and MS-GARCH models with
Student’s t innovations.
Finally, we believe further research could endeavour to replicate these
results for different firm sizes (e. g. the FTSE 250, FTSE Small Cap and FTSE
All Share indices). This would be to investigate if smaller firms have greater
leverage effects due to the additional volatility imposed upon them as a
consequence of lower trading volume. This may shed further light on quantifying
the effects of skewness and regime change in forecasting VaR for long and short
positions.
References
Ane, T., Ureche-Rangau, L., 2006. Stock market dynamics in a regime-switching

asymmetric power GARCH model. International Review of Financial
Analysis. 15, 109-129.
Arghyrou, M., Gregoriou, A., 2007. Testing for Purchasing Power Parity
correcting for non normality using the wild bootstrap. Economics Letters,
95(2), 285-290.
Basle Committee on Banking Supervision , 1996. Supervisory Framework for the

use of Backtesting in Conjunction with the Internal Models Approach to
Market Risk Capital Requirements. January.
Bollerslev, T., 1986. Generalized Autoregressive Conditional Heteroskedasticity.

Journal of Econometrics, 31, 307-327.

Bond, S. A., 2000. Asymmetry and downside risk in foreign exchange markets.
Mimeo, Department of Land Economy, University of Cambridge,
Cambridge, UK.
Box, G. E. P., Jenkins, G. M., Reinsel, G. C., 1994. Time Series Analysis:
Forecasting and Control. Third edition, Prentice Hall.
Christoffersen, P. F., 2003. Elements of financial risk management. Academic

Press, San Diego.
Dacco, R., Satchell, S., 1999. Why do regime-switching models forecast so

badly? Journal of Forecasting, 18, 1-16.
Ding, Zhuanxin, Granger, C. W. J., Engle, R. F., 1993. A long memory property
of stock market returns and a new model. Journal of Empirical Finance, 1,
83-106.
Dowd, K., 2005. Measuring Market Risk. John Wiley & Sons Ltd.
Dueker, M. J., 1997. Markov switching in GARCH process and Mean-Reverting

stock market volatility. Journal of Business and Economic Statistics, 15(1),
26-34.
Engle, R. F., 1982. Autoregressive Conditional Heteroskedasticity with estimate

the variance of UK Inflation. Econometrica, 50, 987-1008.
Giot, P., Laurent, S., 2003. Value-at-Risk for long and short positions. Journal of
Applied Econometrics, 18, 641-664.
Glosten, L. R., Jaganathan, R., Runkle, D., 1993. On the relation between the
expected value and the volatility of the normal excess return on stocks.
Journal of Finance, 48, 1779-1801.
Gray, S. F., 1996. Modelling the conditional distribution of interest rates as a

Regime-Switching process. Journal of Financial Economics, 42, 27-62.
Haas, M., Mittnik, S., Paollela, M. S., 2004. A new approach in Markov-
switching GARCH models. Journal of Financial Econometrics, 2(4), 493-
530.

Hamilton, J. D., 1994. Time series analysis. Princeton University Press, Princeton,
New Jersey.
Hansen, B. E., 1996. Erratum: the Likelihood Ratio test under nonstandard
conditions: Testing the Markov switching model of GNP. Journal of
Applied Econometrics, 11, 195-198.
Hansen, P. R., 2005. A test for superior predictive ability. Journal of Business and
Economic Statistics, 23, 365-380.
Hannan, E. J., 1980. The estimation of the order of an ARMA process. The
Annals of Statistics, 8(5), 1071-1081.
Hurlin, C., Tokpavi, S., 2006. Backtesting VaR accuracy: A new simple test.
University of Orleans, Working Paper.
Klaassen, F., 2002. Improving GARCH volatility forecasts with Regime-

Switching GARCH. Empirical Economics, 27, 363-394.
Lambert, P., Laurent, S., 2001. Modelling financial time series using GARCH-
type models and a skewed Student density. Mimeo, Universite de Liege.
Laurent, P., Peter, J-P., 2002. G@RCH 2.2: An Ox package for estimating and
forecasting various ARCH models, Journal of Economic Surreys, 16, 447-
485.
Lamoureux, C. G., Lastrapes, W. D., 1990. Persistence in variance, structural

change and the GARCH model. Journal of Business and Economic
Statistics, 8, 225-234.
Lunde, A., Hansen, P. R., 2005. A forecast comparison of volatility models: does
anything beat a GARCH(1,1)?. Journal of Applied Econometrics, 20, 873-
889.
Marcucci, J., 2005. Forecasting stock market volatility with Regime-Switching

GARCH models. Studies in Nonlinear Dynamics & Econometrics: Vol. 9,
No. 4, Article 6.
McLachlan, G. J., Peel, D., 2000. Finite Mixture Models. New York: John Wiley
& Sons.

Mikosch, T., Starica, C., 2004. Nonstationarities in financial time series, the long-
range dependence, and the IGARCH effects. Review of Econometrics and
Statistics, 86, 378-390.
Morana, C., 2002. GARCH effects: An interpretation. Applied Econometrics

Letters, 11, 745-748.
Nelson, D. B., 1991. Conditional heteroskedasticity in asset returns: A New

Approach. Econometrica, 59, 347-370.
Panagiotidis, T., 2005. Market capitalization and efficiency. Does it matter?

Applied Financial Economics, 15(9), 707-713.
Sarma, M., Thomas, S., Shah, A., 2003. Selection of Value-at-Risk models.
Journal of Forecasting, 22, 337-358.
Timmermann, A., 2000. Moments of Markov switching models. Journal of

Econometrics, 96, 75-111.
Yang, M., 2000. Some properties of autoregressive Processes with Markov-

Switching coefficients. Econometric Theory, 16, 23-43.
Zakoïan, J. M., 1994. Threshold heteroskedastic models. Journal of Economic

Dynamics and Control, 18, 931-944.


Studies in Nonlinear Dynamics & Econometrics: Markov-Switching GARCH Modelling of Value-at-Risk

Uploaded by

Copyright:

Available Formats

Studies in Nonlinear Dynamics & Econometrics: Markov-Switching GARCH Modelling of Value-at-Risk

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Studies in Nonlinear Dynamics & Econometrics: Markov-Switching GARCH Modelling of Value-at-Risk

Uploaded by

Copyright:

Available Formats

Studies in Nonlinear Dynamics &

Markov-Switching GARCH Modelling of

Brought to you by | Periodicals Section, Albert Sloman Libr.

Brought to you by | Periodicals Section, Albert Sloman Libr.

Brought to you by | Periodicals Section, Albert Sloman Libr.

2.1 MS-GARCH model

Brought to you by | Periodicals Section, Albert Sloman Libr.

on the past only through the most recent value st −1 . That is

Pr{st = i | st −1 = j , st − 2 = k ,...} = Pr{st = i | st −1 = j} = pij ,

A M × M matrix P, known as the transition matrix, contains transition

giving the ARCH class of models, or by:

β s ( L) log ht ,s = ω s + [ β s ( L) − ϕ s ( L)](ht−,s1 / 2 (| ε t | +λs ε t t )) ,

Brought to you by | Periodicals Section, Albert Sloman Libr.

intractable and practically impossible to estimate. This problem is known as path

f (rt | st = j ,ψ t −1 ; Λ). Pr{st = j | ψ t −1}

where f (rt | st = j ,ψ t −1 ; Λ ) denotes the (conditional) probability density of the

f (rt |ψ t −1 ; Λ ) = ∑ j =1 f (rt | st = j ,ψ t −1 ; Λ ). Pr{st = j |ψ t −1} .

Pr{st = j | ψ t −1} = ∑i =1 Pr{st −1 = i | ψ t −1} Pr{st = j | st −1 = i,ψ t −1}.

2.2 Long and short VaR

Brought to you by | Periodicals Section, Albert Sloman Libr.

l with probability p is defined as

p = Pr[ΔV (l) ≤ VaR] = Fl (VaR) , (2.8)

where F(x) denotes the cumulative distribution function, CDF, of ΔV (l) .

VaRlp = Fl−1 ( p) , (2.9)

VaRTp+1 (r ) = μT +1 + hT1 /+21Fp−1 ( z ), (2.10)

where Fp−1 ( z ) denotes the pth quantile of the distribution of variance-adjusted

Brought to you by | Periodicals Section, Albert Sloman Libr.

For skewed Student’s t innovations, F −1 in (2.10) for long and short

Brought to you by | Periodicals Section, Albert Sloman Libr.

Table 1: Moments of the FTSE100 and S&P500 returns along with

Brought to you by | Periodicals Section, Albert Sloman Libr.

the potential benefits of allowing for conditional heteroskedasticity.7 More

Brought to you by | Periodicals Section, Albert Sloman Libr.

3.2 Specification and diagnostic results

Brought to you by | Periodicals Section, Albert Sloman Libr.

Table 2: ML estimation results from the single-regime models.

Brought to you by | Periodicals Section, Albert Sloman Libr.

Brought to you by | Periodicals Section, Albert Sloman Libr.

Table 4: ML estimation results from the two-regime GARCH models with

3.3 Comparing single-regime and MS models

Brought to you by | Periodicals Section, Albert Sloman Libr.

Table 7 presents the unconditional probabilities, π j , the expected durations,

Brought to you by | Periodicals Section, Albert Sloman Libr.

Table 6: ML estimation results from the two-regime GARCH models with

Brought to you by | Periodicals Section, Albert Sloman Libr.

Table 7: Regime switching properties in FTSE100 and S&P500 markets.

π1 , π 2 0.845 0.155 0.420 0.580 0.797 0.203 0.243 0.757

δ1 , δ 2 5.471 1.001 541 746 4.370 1.114 617 1923

The above results resemble those in which regime-switching GARCH

4 Performance of Models in Forecasting VaR

Brought to you by | Periodicals Section, Albert Sloman Libr.

employed for testing the single-regime versus the MS model.11 Furthermore,

4.1 Exceptions-based backtesting results

Brought to you by | Periodicals Section, Albert Sloman Libr.

Brought to you by | Periodicals Section, Albert Sloman Libr.

VaR 5% VaR 95%

VaR 1% VaR 99%

VaR 0.5% VaR 99.5%

Brought to you by | Periodicals Section, Albert Sloman Libr.