B6 A Comprehensive Review of Value at Risk Methodologies

The Spanish Reviewof Financial Economics 12 (2014) 1532
The Spanish Review of Financial Economics

www. el sevi er . es/ sr f e
Article
A comprehensive review of Value at Risk methodologies
Pilar Abad
a
, Sonia Benito
b,
, Carmen Lpez
c
a
Universidad Rey Juan Carlos and IREA-RFA, Paseo Artilleros s/n, 28032 Madrid, Spain
b
Universidad Nacional de Educacin a Distancia (UNED), Senda del Rey 11, 28223 Madrid, Spain
c
Universidad Nacional de Educacin a Distancia (UNED), Spain
a r t i c l e i n f o
Article history:
Received 5 October 2012
Accepted 28 June 2013
Available online 9 September 2013
JEL classication:
G32
C14
C15
C22
Keywords:
Value at Risk
Volatility
Risk management
a b s t r a c t
In this article we present a theoretical review of the existing literature on Value at Risk (VaR) specically
focussing on the development of new approaches for its estimation. We effect a deep analysis of the State
of the Art, from standard approaches for measuring VaR to the more evolved, while highlighting their
relative strengths and weaknesses. We will also review the backtesting procedures used to evaluate VaR
approach performance. From a practical perspective, empirical literature shows that approaches based
on the Extreme Value Theory and the Filtered Historical Simulation are the best methods for forecasting
VaR. The Parametric method under skewed and fat-tail distributions also provides promising results
especially when the assumption that standardised returns are independent and identically distributed is
set aside and when time variations are considered in conditional high-order moments. Lastly, it appears
that some asymmetric extensions of the CaViaR method provide results that are also promising.
2012 Asociacin Espaola de Finanzas. Published by Elsevier Espaa, S.L. All rights reserved.
1. Introduction
Basel I, also called the Basel Accord, is the agreement reached
in 1988 in Basel (Switzerland) by the Basel Committee on Bank
Supervision (BCBS), involving the chairmen of the central banks
of Germany, Belgium, Canada, France, Italy, Japan, Luxembourg,
Netherlands, Spain, Sweden, Switzerland, the United Kingdomand
the United States of America. This accord provides recommenda-
tions on banking regulations with regard to credit, market and
operational risks. Its purpose is to ensure that nancial institu-
tions holdenoughcapital onaccount tomeet obligations andabsorb
unexpected losses.
For a nancial institution measuring the risk it faces is an essen-
tial task. In the specic case of market risk, a possible method of
measurement is the evaluation of losses likely to be incurred when
the price of the portfolio assets falls. This is what Value at Risk
(VaR) does. The portfolio VaR represents the maximum amount
an investor may lose over a given time period with a given prob-
ability. Since the BCBS at the Bank for International Settlements
requires a nancial institution to meet capital requirements on the
basis of VaR estimates, allowing them to use internal models for
This work has been funded by the Spanish Ministerio de Ciencia y Tecnologa
(ECO2009-10398/ECON and ECO2011-23959).
Corresponding author.
E-mail addresses: pilar.abad@urjc.es (P. Abad), soniabm@cee.uned.es (S. Benito).
VaR calculations, this measurement has become a basic market
risk management tool for nancial institutions.
2
Consequently, it
is not surprising that the last decade has witnessed the growth of
academic literature comparing alternative modelling approaches
and proposing new models for VaR estimations in an attempt to
improve upon those already in existence.
Although the VaR concept is very simple, its calculation is not
easy. The methodologies initially developed to calculate a portfo-
lio VaR are (i) the variancecovariance approach, also called the
Parametric method, (ii) the Historical Simulation (Non-parametric
method) and (iii) the Monte Carlo simulation, which is a Semi-
parametric method. As is well known, all these methodologies,
usually called standard models, have numerous shortcomings,
which have led to the development of new proposals (see Jorion,
2001).
Among Parametric approaches, the rst model for VaR estima-
tionis Riskmetrics, fromMorgan(1996). Themajor drawbackof this
2
When the Basel I Accord was concluded in 1988, no capital requirement was
dened for the market risk. However, regulators soon recognised the risk to a bank-
ing system if insufcient capital was held to absorb the large sudden losses from
huge exposures in capital markets. During the mid-90s, proposals were tabled for
an amendment to the 1988 accord, requiring additional capital over and above the
minimum required for credit risk. Finally, a market risk capital adequacy frame-
work was adopted in 1995 for implementation in 1998. The 1995 Basel I Accord
amendment provided a menu of approaches for determining the market risk capital
requirements.
2173-1268/$ see front matter 2012 Asociacin Espaola de Finanzas. Published by Elsevier Espaa, S.L. All rights reserved.
http://dx.doi.org/10.1016/j.srfe.2013.06.001
16 P. Abad et al. / The Spanish Review of Financial Economics 12 (2014) 1532
model is the normal distribution assumption for nancial returns.
Empirical evidence shows that nancial returns do not follow a
normal distribution. The second relates to the model used to esti-
mate nancial return conditional volatility. The third involves the
assumption that return is independent and identically distributed
(iid). There is substantial empirical evidence to demonstrate that
standardised nancial returns distribution is not iid.
Given these drawbacks research on the Parametric method
has moved in several directions. The rst involves nding a
more sophisticated volatility model capturing the characteristics
observed in nancial returns volatility. The second line of research
involves searching for other density functions that capture skew-
ness and kurtosis of nancial returns. Finally, the third line of
research considers that higher-order conditional moments are
time-varying.
In the context of the Non-parametric method, several Non-
parametric density estimation methods have been implemented,
withimprovement ontheresults obtainedbyHistorical Simulation.
In the framework of the Semi-parametric method, newapproaches
havebeenproposed: (i) theFilteredHistorical Simulation, proposed
by Barone-Adesi et al. (1999); (ii) the CaViaR method, proposed by
Engle and Manganelli (2004) and (iii) the conditional and uncon-
ditional approaches based on the Extreme Value Theory. In this
article, we will review the full range of methodologies developed
to estimate VaR, fromstandard models to those recently proposed.
We will expose the relative strengths and weaknesses of these
methodologies, from both theoretical and practical perspectives.
The articles objective is to provide the nancial risk researcher
withall themodels andproposeddevelopments for VaRestimation,
bringing himto the limits of knowledge in this eld.
The paper is structured as follows. In the next section, we
review a full range of methodologies developed to estimate VaR.
In Section 2.1, a non-parametric approach is presented. Paramet-
ric approaches are offered in Section 2.2, and semi-parametric
approaches in Section 2.3. In Section 3, the procedures for mea-
suring VaR adequacy are described and in Section 4, the empirical
results obtained by papers dedicated to comparing VaR method-
ologies are shown. In Section 5, some important topics of VaR are
discussed. The last section presents the main conclusions.
2. Value at Risk methods
According to Jorion(2001), VaRmeasure is denedas the worst
expectedloss over a givenhorizonunder normal market conditions
at a given level of condence. For instance, a bank might say that
the daily VaR of its trading portfolio is $1 million at the 99 percent
condence level. In other words, under normal market conditions,
only one percent of the time, the daily loss will exceed $1 million.
In fact the VaR just indicates the most we can expect to lose if no
negative event occurs.
The VaRis thus a conditional quantile of the asset returnloss dis-
tribution. Among the main advantages of VaR are simplicity, wide
applicability and universality (see Jorion, 1990, 1997).
3
Let r
1
, r
2
,
r
3
,. . ., r
n
be identically distributed independent random variables
representing the nancial returns. Use F(r) to denote the cumula-
tive distribution function, F(r) =Pr(r <r|
t 1
) conditionally on the
3
There is another market risk measurement, called Expected Shortfall (ES). ES
measures the expected value of our losses if we get a loss in excess of VaR. So that,
this measure tells us what to expect in a bad estate, while the VaR tells us nothing
more than to expect a loss higher than the VaR itself. In Section 5, we will formally
dene this measure besides presenting some criticisms of VaR measurement.
information set
t 1
that is available at time t 1. Assume that
{r
t
} follows the stochastic process:
r
t
= +
t
t
= z
t
o
t
z
t
iid(0, 1)
(1)
where o
2
t
= (z
2
t
|
t1
) and z
t
has the conditional distributionfunc-
tion C(z), C(z) = Pr(z
t
-z|
t1
). The VaR with a given probability
(0, 1), denoted by VaR(), is dened as the quantile of the
probability distribution of nancial returns: F(VaR()) = Pr(r
t
-
VaR()) = or VaR() = inf {v|P(r
t
v) = }.
This quantile can be estimated in two different ways: (1) invert-
ing the distribution function of nancial returns, F(r), and (2)
inverting the distribution function of innovations, with regard to
G(z) the latter, it is also necessary to estimate o
2
t
.
VaR() = F
1
() = +o
t
C
1
() (2)
Hence, a VaR model involves the specications of F(r) or G(z).
The estimation of these functions can be carried out using the
following methods: (1) non-parametric methods; (2) parametric
methods and (3) semi-parametric methods. Belowwe will describe
the methodologies, which have been developed in each of these
three cases to estimate VaR.
4
2.1. Non-parametric methods
The Non-parametric approaches seektomeasure a portfolioVaR
without makingstrongassumptions about returns distribution. The
essence of these approaches is to let data speak for themselves as
much as possible and to use recent returns empirical distribution
not some assumed theoretical distribution to estimate VaR.
All Non-parametric approaches are based on the underlying
assumption that the near future will be sufciently similar to the
recent past for us to be able to use the data fromthe recent past to
forecast the risk in the near future.
The Non-parametric approaches include (a) Historical Simula-
tion and (b) Non-parametric density estimation methods.
2.1.1. Historical simulation
Historical Simulation is the most widely implemented Non-
parametric approach. This method uses the empirical distribution
of nancial returns as an approximation for F(r), thus VaR() is
the quantile of empirical distribution. To calculate the empiri-
cal distribution of nancial returns, different sizes of samples can
be considered.
The advantages and disadvantages of the Historical Simulation
have been well documented by Down (2002). The two main advan-
tages are as follows: (1) the method is very easy to implement,
and (2) as this approach does not depend on parametric assump-
tions onthe distributionof the returnportfolio, it canaccommodate
wide tails, skewness and any other non-normal features in nan-
cial observations. Thebiggest potential weakness of this approachis
that its results are completely dependent on the data set. If our data
period is unusually quiet, Historical Simulation will often underes-
timate risk and if our data period is unusually volatile, Historical
Simulation will often overestimate it. In addition, Historical Simu-
lation approaches are sometimes slowto reect major events, such
as the increases in risk associated with sudden market turbulence.
The rst papers involving the comparison of VaR methodolo-
gies, such as those by Beder (1995, 1996), Hendricks (1996), and
Pritsker (1997), reported that the Historical Simulation performed
at least as well as the methodologies developed in the early years,
4
For a more pedagogic review of some of these methodologies (see Feria
Domnguez, 2005).
P. Abad et al. / The Spanish Review of Financial Economics 12 (2014) 1532 17
the Parametric approach and the Monte Carlo simulation. The main
conclusion of these papers is that among the methodologies devel-
oped initially, no approach appeared to perform better than the
others.
However, more recent papers such as those by Abad and Benito
(2013), Ashley and Randal, 2009, Trenca (2009), Angelidis et al.
(2007), Alonso and Arcos (2005), Gento (2001), Danielsson and de
Vries (2000) have reported that the Historical Simulation approach
produces inaccurate VaR estimates. In comparison with other
recently developed methodologies such as the Historical Simula-
tion Filtered, Conditional Extreme Value Theory and Parametric
approaches (as we become further separated from normality and
consider volatility models more sophisticated than Riskmetrics),
Historical Simulation provides a very poor VaR estimate.
2.1.2. Non-parametric density estimation methods
Unfortunately, the Historical Simulation approach does not best
utilise the information available. It also has the practical drawback
that it only gives VaR estimates at discrete condence intervals
determined by the size of our data set.
5
The solution to this prob-
lemis to use the theory of Non-parametric density estimation. The
idea behind Non-parametric density is to treat our data set as if
it were drawn fromsome unspecic or unknown empirical distri-
bution function. One simple way to approach this problem is to
draw straight lines connecting the mid-points at the top of each
histogram bar. With these lines drawn the histogram bars can be
ignoredandthearea under thelines treatedas thoughit was a prob-
ability density function (pdf) for VaR estimation at any condence
level. However, we could draw overlapping smooth curves and so
on. This approachconforms exactlytothe theoryof non-parametric
density estimation, which leads to important decisions about the
width of bins and where bins should be centred. These decisions
can therefore make a difference to our results (for a discussion, see
Butler and Schachter (1998) or Rudemo (1982)).
A kernel density estimator (Silverman, 1986; Sheather and
Marron, 1990) is a methodfor generalising a histogramconstructed
with the sample data. A histogramresults in a density that is piece-
wise constant where a kernel estimator results in smooth density.
Smoothing the data can be performed with any continuous shape
spread around each data point. As the sample size grows, the net
sum of all the smoothed points approaches the true pdf whatever
that may be irrespective of the method used to smooth the data.
The smoothing is accomplished by spreading each data point
with a kernel, usually a pdf centred on the data point, and a param-
eter called the bandwidth. A common choice of bandwidth is that
proposed by Silverman (1986). There are many kernels or curves
to spread the inuence of each point, such as the Gaussian kernel
density estimator, the Epanechnikov kernel, the biweight kernel,
an isosceles triangular kernel and an asymmetric triangular kernel.
Fromthe kernel, we can calculate the percentile or estimate of the
VaR.
2.2. Parametric method
Parametric approaches measure risk by tting probability
curves to the data and then inferring the VaR fromthe tted curve.
Among Parametric approaches, the rst model to estimate VaR
was Riskmetrics from Morgan (1996). This model assumes that
the return portfolio and/or the innovations of return follow a nor-
mal distribution. Under this assumption, the VaR of a portfolio at
5
Thus, if we have, e.g., 100 observations, it allows us to estimate VaR at the 95%
condence level but not the VaR at the 95.1% condence level. The VaR at the 95%
condence level is givenby the sixthlargest loss, but the VaRat the 95.1%condence
level is a problembecause there is no loss observation to accompany it.
an 1% condence level is calculated as VaR() = +o
t
C
1
(),
where G
1
() is the quantile of the standard normal distribution
and o
t
is the conditional standard deviation of the return portfo-
lio. To estimate o
t
, Morgan uses an Exponential Weight Moving
Average Model (EWMA). The expression of this model is as follows:
o
2
t
(1 z)
N1
j=0
z
j
(
tj
)
2
(3)
where z=0.94 and the windowsize (N) is 74 days for daily data.
The major drawbacks of Riskmetrics are related to the normal
distribution assumption for nancial returns and/or innovations.
Empirical evidence shows that nancial returns do not follownor-
mal distribution. The skewness coefcient is in most cases negative
and statistically signicant, implying that the nancial return dis-
tribution is skewed to the left. This result is not in accord with
the properties of a normal distribution, which is symmetric. Also,
empirical distribution of nancial return has been documented to
exhibit signicantly excessive kurtosis (fat tails and peakness) (see
Bollerslev, 1987). Consequently, the size of the actual losses is much
higher than that predicted by a normal distribution.
The second drawback of Riskmetrics involves the model used
to estimate the conditional volatility of the nancial return. The
EWMA model captures somenon-linear characteristics of volatility,
such as varying volatility and cluster volatility, but does not take
into account asymmetry and the leverage effect (see Black, 1976;
Pagan and Schwert, 1990). In addition, this model is technically
inferior to the GARCH family models in modelling the persistence
of volatility.
The third drawback of the traditional Parametric approach
involves the iid return assumption. There is substantial empirical
evidence that the standardised distribution of nancial returns is
not iid (see Hansen, 1994; Harvey and Siddique, 1999; Jondeau and
Rockinger, 2003; Bali and Weinbaum, 2007; Brooks et al., 2005).
Given these drawbacks research on the Parametric method has
been made in several directions. The rst attempts searched for
a more sophisticated volatility model capturing the characteris-
tics observed in nancial returns volatility. Here, three families of
volatility models have been considered: (i) the GARCH, (ii) stochas-
tic volatility and (iii) realised volatility. The second line of research
investigated other density functions that capture the skew and
kurtosis of nancial returns. Finally, the third line of research
considered that the higher-order conditional moments are time-
varying.
Using the Parametric method but with a different approach,
McAleer et al. (2010a) proposed a risk management strategy
consisting of choosing fromamong different combinations of alter-
native risk models to measure VaR. As the authors remark, given
that a combination of forecast models is also a forecast model, this
model is a novel method for estimating the VaR. With such an
approach McAleer et al. (2010b) suggest using a combination of
VaR forecasts to obtain a crisis robust risk management strategy.
McAleer et al. (2011) present cross-country evidence to support the
claimthat the median point forecast of VaR is generally robust to a
Global Financial Crisis.
2.2.1. Volatility models
The volatility models proposed in literature to capture the char-
acteristics of nancial returns can be divided into three groups:
the GARCH family, the stochastic volatility models and realised
volatility-based models. As to the GARCHfamily, Engle (1982) pro-
posed the Autoregressive Conditional Heterocedasticity (ARCH),
which featured a variance that does not remain xed but rather
varies throughout a period. Bollerslev (1986) further extended the
model by inserting the ARCH generalised model (GARCH). This
Table 1
Asymmetric GARCH.
Formulations Restrictions
EGARCH (1,1) log(o
2
t
) =
0
+
_
t1
o
t1
_
+
1
_
t1
o
t1
_
2
_
+ log(o
2
t1
)
1
+<1
GJR-GARCH
(1,1)
o
2
t
=
0
+
1
2
t1
2
t1
S
t1
+o
2
t1
S
t1
= 1 for
t1
-0 and S
t1
= 0 otherwise
0
:0,
1
, :0
1
+ -1
TS GARCH (1,1) ot =
0
+
1
|
t1
| +o
t1
0
:0,
1
, :0
1
+ -1
TGARCH
ot =
1
+
1
|
t1
| +
t1
S
t1
+o
t1
S
t1
= 1 for
t1
-0 and S
t1
= 0 otherwise
0
:0,
1
, :0
1
+ -1
PGARCH (1,1) o
t
= +|
t1
|
+o
t1
:0, 0
0, :0
APGARCH (1,1) o
t
=
0
+
1
(|
t1
| +
t1
)
+o
t1
0
:0,
1
0, :0,
:0 1 - -1
AGARCH (1,1) o
2
t
=
0
+
1
2
t1
+
2
t1
+o
2
t1
0
:0,
1
, :0
1
+ -1
SQR-GARCH o
2
t
=
0
+
1
(o
t1
+
t1
]o
t1
)
2
+o
2
t1
0
:0,
1
, :0
1
+ -1
Q-GARCH o
2
t
=
0
+
1
(
t1
+
t1
)
2
+o
2
t1
0
:0,
1
, :0
1
+ -1
VGARCH o
2
t
=
0
+
1
( +o
1
t1
t1
)
2
+o
2
t1
0
:0, 0 - -1
0 -
1
-1
NAGARCH (1,1) o
2
t
=
0
+
1
(
t1
+o
t1
)
2
+o
2
t1
0
:0,
1
, :0
1
+ -1
MS-GARCH
(1,1)
rt = s
t
+t = s
t
+ot zt with zt iid N(0, 1)
o
2
t
= s
t
+s
t
2
t1
+s
t
o
2
t1
st : state of the process at the t
s
t
:0, s
t
0, and s
t
0
RS-APARCH o
st
s
t
= s
t
+s
t
(|
t1
| +s
t
t1
)
s
t
+
s
t
2
o
s
t
t1
s
t
:0, s
t
0, s
t
:0
:0 1 -
s
t
1
-1
model species and estimates two equations: the rst depicts the
evolution of returns in accordance with past returns, whereas the
second patterns the evolving volatility of returns. The most gener-
alised formulation for the GARCHmodels is the GARCH(p,q) model
represented by the following expression:
r
t
=
t
+
t
o
2
t
=
0
+
q
i=1
2
ti
+
p
j=1
i
o
2
tj
(4)
In the GARCH(1,1) model, the empirical applications conducted
on nancial series detect that
1
+
1
is observed to be very close
to the unit. The integrated GARCH model (IGARCH) of Engle and
Bollerslev (1986)
6
is then obtained forcing the condition that the
addition is equal to the unit in expression (4). The conditional vari-
ance properties of the IGARCH model are not very attractive from
the empirical point of view due to the very slow phasing out of
the shock impact upon the conditional variance (volatility persis-
tence). Nevertheless, the impacts that fade away showexponential
behaviour, which is how the fractional integrated GARCH model
(FIGARCH) proposed by Baillie et al. (1996) behaves, with the sim-
plest specication, FIGARCH (1, d, 0), being:
o
2
t
=

0
1
1
+
_
1
(1 L)
d
(1
1
L)
_
r
2
t
(5)
If theparameters complywiththesettingconditions
0
:0, 0
1
-d 1, the conditional variance of the model is most likely
positive for all t cases. With this model, there is a likelihood that
the r
2
t
effect upon o
2
t+k
will trigger a decline over the hyperbolic
rate while k surges.
The models previously mentioned do not completely reect the
nature posed by the volatility of the nancial times series because,
6
The EWMA model is equivalent tothe IGARCHmodel withthe intercept
0
being
restricted to being zero, the autoregressive parameter being set at a pre-specic
value z, and the coefcient of
2
t1
being equal to 1z.
althoughtheyaccuratelycharacterisethevolatilityclusteringprop-
erties, theydonot takeintoaccount theasymmetric performanceof
yields before positive or negative shocks (leverage effect). Because
previous models depend on the square errors, the effect caused
by positive innovations is the same as the effect produced by nega-
tive innovations of equal absolute value. Nonetheless, reality shows
that in nancial time series, the existence of the leverage effect is
observed, which means that volatility increases at a higher rate
when yields are negative compared with when they are positive.
In order to capture the leverage effect several non-linear GARCH
formulations have been proposed. In Table 1, we present some of
the most popular. For a detailed review of the asymmetric GARCH
models (see Bollerslev, 2009).
In all models presented in this table, is the leverage parame-
ter. A negative value of means that past negative shocks have a
deeper impact on current conditional volatility than past positive
shocks. Thus, we expect the parameter to be negative ( <0). The
persistence of volatility is captured by the parameter. As for the
EGARCH model, the volatility of return also depends on the size of
innovations. If
1
is positive, the innovations superior to the mean
have a deeper impact on current volatility than those inferior.
Finally, it must be pointed out that there are some models
that capture the leverage effect and the non-persistence memory
effect. For example, Bollerslev and Mikkelsen (1996) insert the FIE-
GARCHmodel, which aims to account for both the leveraging effect
(EGARCH) and the long memory (FIGARCH) effect. The simplest
expression of this family of models is the FIEGARCH (1, d, 0):
(1 L)(1 L)
d
log(o
2
t
) =
0
+
_
r
t1
o
t1
_
+
1
_
r
t1
o
t1
_
2
_
(6)
Some applications of the family of GARCH models in VaR litera-
ture can be found in the following studies: Abad and Benito (2013),
Sener et al. (2012), Chen et al. (2009, 2011), Sajjad et al. (2008), Bali
andTheodossiou(2007), Angelidis et al. (2007), Haas et al. (2004), Li
and Lin (2004), Carvalho et al. (2006), Gonzlez-Rivera et al. (2004),
Giot and Laurent (2004), Mittnik and Paolella (2000), among oth-
ers. Although there is no evidence of an overpowering model, the
results obtained in these papers seemto indicate that asymmetric
GARCH models produce better outcomes.
An alternative path to the GARCH models to represent the tem-
poral changes over volatility is through the stochastic volatility
(SV) models proposed by Taylor (1982, 1986). Here volatility in t
does not depend on the past observations of the series but rather
on a non-observable variable, which is usually an autoregressive
stochastic process. To ensure the positiveness of the variance, the
volatility equation is dened following the logarithm of the vari-
ance as in the EGARCH model.
The stochastic volatility model proposed by Taylor (1982) can
be written as:
r
t
=
t
+
_
h
t
z
t
z
t
N(0, 1)
log h
t+1
= +log h
t
+q
t
q
t
N(0, o
n
)
(7)
where
t
represents the conditional mean of the nancial return,
h
t
represents the conditional variance, and z
t
and q
t
are stochastic
white-noise processes.
The basic properties of the model can be found in Taylor (1986,
1994). As in the GARCHfamily, alternative and more complex mod-
els have been developed for the stochastic volatility models to
allow for the pattern of both the large memory (see the model of
Harvey (1998) and Breidt et al. (1998)) and the leverage effect (see
the models of Harvey and Shephard (1996) and So et al. (2002)).
Some applications of the SV model to measure VaR can be found
in Fleming and Kirby (2003), Lehar et al. (2002), Chen et al. (2011)
and Gonzlez-Rivera et al. (2004).
The third group of volatility models is realised volatility (RV).
The origin of the realised volatility concept is certainly not recent.
Merton (1980) had already mentioned this concept, showing the
likelihood of the latent volatility approximation by the addition of
N intra-daily square yields over a t period, thus implying that the
addition of square yields could be used for the variance estimation.
Taylor and Xu (1997) showed that the daily realised volatility can
be easily crafted by adding the intra-daily square yields. Assuming
that a day is divided into equidistant N periods and if r
i,t
represents
the intra-daily yield of the i-interval of day t, the daily volatility for
day t can be expressed as:
RV =
_
N
i=1
r
i,t
_
2
=
N
i=1
r
2
i,t
+2
N
i=1
N
j=i+1
r
j,t
r
ji,t
(8)
In the event of yields with zero mean and no correlation what-
soever, then
_
N
i=1
r
2
i,t
_
is a consistent estimator of the daily
variance o
2
t
Andersen et al. (2001a,b) upheld that this measure
signicantly improves the forecast compared with the standard
procedures, which just rely on daily data.
Although nancial yields clearly exhibit leptokurtosis, the
standardised yields by realised volatility are roughly normal. Fur-
thermore, although the realised volatility distribution poses a clear
asymmetry to the right, the distributions of the realised volatil-
ity logarithms are approximately Gaussian (Pong et al. (2004)). In
addition, the long-term dynamics of the realised volatility loga-
rithm can be inferred by a fractionally integrated long memory
process. The theorysuggests that realisedvolatilityis a non-skewed
estimator of the volatility yields and is highly efcient. The use of
the realised volatility obtained fromthe high-frequency intra-daily
yields allows for the use of traditional procedures of temporal times
series to create patterns and forecasts.
One of the most representative realised volatility models is that
proposed by Pong et al. (2004):
(1
1
L
2
L
2
)(lnRv
t
) = (1
1
L)u
t
(9)
As in the case of GARCH family models and stochastic volatil-
ity models, some extension of the standard RV model have been
developed in order to capture the leverage effect and long-range
dependence of volatility. The former issue has been investigated
by Bollerslev et al. (2011), Chen and Ghysels (2010), Patton and
Sheppard (2009), among others. With respect to the latter point,
the autoregressive fractionally integrated model has been used by
Andersen et al. (2001a, 2001b, 2003), Koopman et al. (2005), Pong
et al. (2004), among others.
a) Empirical results of volatility models in VaR
This section lists the results obtained from research on the
comparison of volatility models in terms of VaR. The EWMA
model provides inaccurate VaR estimates. In a comparison with
other volatility models, the EWMA model scored the worst per-
formance in forecasting VaR (see Chen et al., 2011; Abad and
Benito, 2013;

Nguez, 2008; Alonso and Arcos, 2006; Gonzlez-
Rivera et al., 2004; Huang and Lin, 2004 and among others).
The performance of the GARCH models strongly depends on
the assumption concerning returns distribution. Overall, under
a normal distribution, the VaR estimates are not very accurate.
However, when asymmetric and fat-tail distributions are con-
sidered, the results improve considerably.
There is scarce empirical evidence of the relative performance
of the SV models against the GARCH models in terms of VaR (see
Fleming and Kirby, 2003; Lehar et al., 2002; Gonzlez-Rivera et al.,
2004; Chen et al., 2011). Fleming and Kirby (2003) compared a
GARCH model with a SV model. They found that both models had
comparable performances in terms of VaR. Lehar et al. (2002) com-
pared option pricing models in terms of VaR using two family
models: GARCH and SV. They found that as to their ability to fore-
cast the VaR, there are no differences between the two. Chen et al.
(2011) compared the performance of two SV models with a range
wide of GARCH family volatility models. The comparison was con-
ductedontwodifferent samples. Theyfoundthat the SVandEWMA
models had the worst performances in estimating VaR. However,
in a similar comparison, Gonzlez-Rivera et al. (2004) found that
the SV model had the best performance in estimating VaR. In gen-
eral, with some exceptions, evidence suggests that SV models do
not improve the results obtained GARCH model family.
The models based on RV work quite well to estimate VaR (see
Asai et al., 2011; Brownlees and Gallo, 2010; Clements et al., 2008;
Giot and Laurent, 2004; Andersen et al., 2003). Some papers show
that an even simpler model (such as an autoregressive), combined
with the assumption of normal distribution for returns, yields rea-
sonable VaR estimates.
As for volatility forecasts, there are many papers in literature
showing that the models based on RV are superior to the GARCH
models. However, not many papers report comparisons on their
ability to forecast VaR. Brownlees and Gallo (2011) compared sev-
eral RV models with a GARCH and EWMA model and found that
the models based on RV outperformed both EWMA and GARCH
models. Along this same line, Giot and Laurent (2004) compared
several volatility models: EWMA, an asymmetric GARCH and RV.
The models are estimated with the assumption that returns follow
either normal or skewed t-Student distributions. They found that
under a normal distribution, the RV model performed best. How-
ever, under a skewed t-distribution, the asymmetric GARCHand RV
models provided very similar results. These authors emphasised
that the superiority of the models based on RV over the GARCH
family is not as obvious when the estimation of the latter assumes
the existence of asymmetric and leptokurtic distributions.
There is a lack of empirical evidence on the performance of frac-
tional integrated volatility models to measure VaR. Examples of
papers that report comparisons of these models are those by So
and Yu (2006) and Beltratti and Morana (1999). The rst paper
compared, in terms of VaR, a FIGARCH model with a GARCH and
an IGARCH model. It showed that the GARCH model provided
more accurate VaRestimates. Ina similar comparisonthat included
the EWMA model, So and Yu (2006) found that FIGARCH did not
outperformGARCH. Theauthors concludedthat, althoughtheir cor-
relation plots displayed some indication of long memory volatility,
this feature is not very crucial in determining the proper value of
VaR. However, in the context of the RV models, there is evidence
that models that capture long memory in volatility provide accu-
rate VaR estimates (see Andersen et al., 2003; Asai et al., 2011).
The model proposed by Asai et al. (2011) captured long memory
volatility and asymmetric features. Along this line,

Nguez (2008)
compared the ability to forecast VaR of different GARCH family
models (GARCH, AGARCH, APARCH, FIGARCH and FIAPARCH, and
EWMA) andfoundthat thecombinationof asymmetric models with
fractional integrated models provided the best results.
Although this evidence is somewhat ambiguous, the asymmet-
ric GARCHmodels seemto provide better VaR estimations than the
symmetric GARCH models. Evidence in favour of this hypothesis
can be found in studies by Sener et al. (2012), Bali and Theodossiou
(2007), Abad and Benito (2013), Chen et al. (2011), Mittnik and
Paolella (2000), Huang and Lin (2004), Angelidis et al. (2007), and
Giot and Laurent (2004). In the context of the models based on
RV, the asymmetric models also provide better results (see Asai
et al., 2011). Some evidence against this hypothesis can be found
in Angelidis et al. (2007).
Finally, some authors state that the assumption of distribution,
not the volatility models, is actually the important factor for esti-
mating VaR. Evidence supporting this issue is found in the study by
Chen et al. (2011).
2.2.2. Density functions
As previouslymentioned, theempirical distributionof thenan-
cial return has been documented to be asymmetric and exhibits
a signicant excess of kurtosis (fat tail and peakness). Therefore,
assuming a normal distribution for risk management and particu-
larly for estimating the VaR of a portfolio does not produce good
results and losses will be much higher.
As t-Student distribution has fatter tails than normal distri-
bution, this distribution has been commonly used in nance and
risk management, particularly to model conditional asset return
(Bollerslev, 1987). In the context of VaR methodology, some appli-
cations of this distribution can be found in studies by Cheng and
Hung (2011), Abad and Benito (2013), Polanski and Stoja (2010),
Angelidis et al. (2007), AlonsoandArcos (2006), Guermat andHarris
(2002), Billio and Pelizzon (2000), and Angelidis and Benos (2004).
The empirical evidence of this distribution performance in esti-
mating VaR is ambiguous. Some papers show that the t-Student
distributionperforms better thanthe normal distribution(see Abad
andBenito, 2013; Polanski andStoja, 2010; AlonsoandArcos, 2006;
SoandYu, 2006
7
). However other papers, suchas thosebyAngelidis
et al. (2007), Guermat and Harris (2002), Billio and Pelizzon (2000),
and Angelidis and Benos (2004), report that the t-Student distribu-
tion overestimates the proportion of exceptions.
The t-Student distribution can often account well for the excess
kurtosis found in common asset returns, but this distribution does
7
This last paper shows that t-Student at 1% performs better in larger positions,
although it does not in short positions.
not capture the skewness of the return. Taking this into account,
one direction for research in risk management involves searching
for other distribution functions that capture these characteris-
tics. In the context of VaR methodology, several density functions
have been considered: the Skewness t-Student Distribution (SSD)
of Hansen (1994); Exponential Generalised Beta of the Second
Kind (EGB2) of McDonald and Xu (1995); Error Generalised Dis-
tribution (GED) of Nelson (1991); Skewness Error Generalised
Distribution (SGED) of Theodossiou (2001); t-Generalised Distri-
bution of McDonald and Newey (1988); Skewness t-Generalised
distribution (SGT) of Theodossiou (1998) and Inverse Hyperbolic
Sign (IHS) of Johnson (1949). In Table 2, we present the density
functions of these distributions.
In this line, some papers such as Dufe and Pan (1997) and
Hull and White (1998) showthat a mixture of normal distributions
produces distributions with fatter tails than a normal distribution
with the same variance.
Some applications to estimate the VaR of skewed distributions
and a mixture of normal distributions can be found in Cheng
and Hung (2011), Polanski and Stoja (2010), Bali and Theodossiou
(2008), Bali et al. (2008), Haas et al. (2004), ZhangandCheng(2005),
Haas (2009), AusnandGaleano(2007), XuandWirjanto(2010) and
Kuester et al. (2006).
These papers raise some important issues. First, regarding the
normal and t-Student distributions, the skewed and fat-tail dis-
tributions seem to improve the t of nancial data (see Bali and
Theodossiou, 2008; Bali et al., 2008; Bali and Theodossiou, 2007).
Consistently, some studies found that the VaR estimate obtained
under skewedandfat-taileddistributions provides a more accurate
VaR than those obtained from a normal or t-Student distribution.
For example, Cheng and Hung (2011) compared the ability to fore-
cast the VaRof a normal, t-Student, SSDandGED. Inthis comparison
the SSDandGEDdistributions provide the best results. Polanski and
Stoja (2010) compared the normal, t-Student, SGT and EGB2 dis-
tributions and found that just the latter two distributions provide
accurate VaR estimates. Bali and Theodossiou (2007) compared a
normal distribution with the SGT distribution. Again, they found
that the SGT provided a more accurate VaR estimate. The main dis-
advantage of using some skewness distribution, such as SGT, is that
the maximisation of the likelihood function is very complicated so
that it may take a lot of computational time (see Nieto and Ruiz,
2008).
Additionally, a mixture of normal distributions, t-Student distri-
butions or GED distributions provided a better VaR estimate than
the normal or t-Student distributions (see Hansen, 1994; Zhang
and Cheng, 2005; Haas, 2009; Ausn and Galeano, 2007; Xu and
Wirjanto, 2010; Kuester et al., 2006). These studies showed that in
thecontext of theParametric method, theVaRestimations obtained
with models involving a mixture with normal distributions (and
t-Student distributions) are generally quite precise.
Lastly, to handle the non-normality of the nancial return Hull
and White (1998) develop a new model where the user is free to
choose any probability distribution for the daily return and the
parameters of the distribution are subject to an updating scheme
such as GARCH. They propose transforming the daily return into a
new variable that is normally distributed. The model is appealing
in that the calculation of VaR is relatively straightforward and can
make use of Riskmetrics or a similar database.
2.2.3. Higher-order conditional time-varying moments
The traditional parametric approach for conditional VaR
assumes that the distribution of returns standardised by condi-
tional means and conditional standard deviations is iid. However,
there is substantial empirical evidence that the distribution of
nancial returns standardised by conditional means and volatil-
ity is not iid. Earlier studies also suggested that the process of
P
.

A
b
a
d

e
t

a
l
.

/

T
h
e

S
p
a
n
i
s
h

R
e
v
i
e
w

o
f

F
i
n
a
n
c
i
a
l

E
c
o
n
o
m
i
c
s

1
2

(
2
0
1
4
)

1
5
3
2

2
1
Table 2
Density functions.
Formulations Restrictions
Skewness t-Student
distribution (SSD) of
Hansen (1994)
j (zt |v, q) =
_
_
bc
_
1 +
1
q 2
_
bzt +u
1 q
_
2
_
(q+1)]2
if zt -
_
u
b
_
bc
_
1 +
1
q 2
_
bzt +u
1 +q
_
2
_
(q+1)]2
if zt
_
u
b
_
u = 4zc
_
q 2
q 2
_
b
2
= 1 +3z
2
u
2
c =
1
_
q +1
2
_
_
(q 2) 1
_
q
2
_
zt = (rt t )]ot
|z| -1
q :2
Beta Exponential
Generalised of the
Second Kind (EGB2)
McDonald and Xu (1995)
C82(zt ; p; q) = C
c
p
(z
t
+)]0)
(1+c
p
(z
t
+)]0))
p+q
C = 1](8(p, q)0) = (4(p) 4(q))0
0 = 1]
_
4
(p) +4
(q) zt = (rt t )]ot

p = qEGB2symmetric
p :qpositivelyskewed
p :qnegativelyskewed
Error Generalised (GED) of
Nelson (1991)
j (zt ) =
q
z2(1+1]q)1(1]q)
exp
_
_
1
2
z
t
z
q
__
z
_
2
(2]v)
1
_
1
q
_
]1
_
3
q
__
zt = (rt t )]ot
0.5
-zt -
0 -q -(thickness of tail)
q -2Thinner tail thannormal
q = 2Normal distribution
q :2Excess kurtosis
Skewness Generalised
Error (SGED) of
Theodossiou (2001)
j (zt |z, k) =
C
o
exp
_
|z
t
+|
k
(1+sign(z
t
+)z)
k
0
k
_
zt = (rt t )]ot
C = k](201 (1]k)) = 2z/S(z)
1
0 = 1(1]k)
0.5
1(3]k)
0.5
S(z)
1
= 2z/S(z)
1
S(z) =
_
1 +3z
2
4/
2
z
2
|z| -1skewedparameter
k = kurtosis parameter
t-Generalised Distribution
(GT) of McDonald and
Newey (1988)
j (zt |z, h, k) =
k1(h)
2z1(1]k)1(h1]k)
_
1 +
_
|z
t
|
z
_
k
_
h
z :0, k -0, h :0
-zt -
Skewness t-Generalised
Distribution (SGT) of
Theodossiou (1998)
j (zt |z, q, k) = C
_
1 +
|z
t
+|
k
((q+1)]k)(1+sign(z
t
+)z)
k
0
k
_
q+1]k
C = 0, 5k
_
q +1
k
_
1]k
8
_
q
k
,
1
k
_
1
0
1
0 =
1
_
g
2
= 2z8
_
q
k
,
1
k
_
1
_
q +1
k
_
1]k
8
_
q 1
k
,
2
k
_
g = (1 +3z
2
)8
_
q
k
,
1
k
_
1
_
q +1
k
_
1]k
8
_
q 1
k
,
3
k
_
= 0
|z| -1skewedparameter
q :2tail-thickness parameter
k :0peakedness parameter
Pearsons skewness
zt = (rt t )]ot
Inverse Hyperbolic sine
(IHS) of Johnson (1949)
lHS(zt |z, k) =
k
_
2(0
2
+(z
t
+)
2
)
exp
_
k
2
2
_
ln
_
(zt +) +
_
0
2
+(zt +)
2
_
(z +ln(0)
_
2
_
0 = 1]ow = w]ow
ow = 0.5(c
2z+k
2
+c
2z+k
2
+2)
0.5
(c
k
2
1)
w mean
ow standarddeviation
w = sinh(z +x]k)
x standardnormal variable
negative extreme returns at different quantiles may differ fromone
another (Engle and Manganelli, 2004; Bali and Theodossiou, 2007).
Thus, given the above, some studies developed a new approach to
calculate conditional VaR. This new approach considered that the
higher-order conditional moments are time-varying (see Bali et al.,
2008; Polanski and Stoja, 2010; Ergun and Jun, 2010).
Bali et al. (2008) introduced a new method based on the
SGT with time-varying parameters. They allowed higher-order
conditional moment parameters of the SGT density to depend
on the past information set and hence relax the conventional
assumption in the conditional VaR calculation that the distri-
bution of standardised returns is iid. Following Hansen (1994)
and Jondeau and Rockinger (2003), they modelled the conditional
high-order moment parameters of the SGT density as an autore-
gressive process. The maximum likelihood estimates show that
the time-varying conditional volatility, skewness, tail-thickness,
and peakedness parameters of the SGT density are statistically sig-
nicant. In addition, they found that the conditional SGT-GARCH
models with time-varying skewness and kurtosis provided a better
t or returns than the SGT-GARCH models with constant skew-
ness and kurtosis. In their paper, they applied this newapproach to
calculate the VaR. The in-sample and out-of-sample performance
results indicated that the conditional SGT-GARCH approach with
autoregressive conditional skewness and kurtosis provided very
accurate and robust estimates of the actual VaR thresholds.
In a similar study, Ergun and Jun (2010) considered the SSD dis-
tribution, which they called the ARCD model, with a time-varying
skewness parameter. They found that the GARCH-based models
that take conditional skewness and kurtosis into account pro-
vided an accurate VaR estimate. Along this same line, Polanski and
Stoja (2010) proposed a simple approach to forecast a portfolio
VaR. They employed the Gram-Charlier expansion (GCE) augment-
ing the standard normal distribution with the rst four moments,
whichareallowedtovaryover time. Thekeyidea was toemploythe
GCE of the standard normal density to approximate the probability
distribution of daily returns in terms of cumulants.
8
This approach
provides a exible tool for modelling the empirical distribution of
nancial data, which, inadditiontovolatility, exhibits time-varying
skewness and leptokurtosis. This method provides accurate and
robust estimates of the realised VaR. Despite its simplicity, their
dataset outperformed other estimates that were generated by both
constant and time-varying higher-moment models.
All previously mentioned papers compared their VaR estimates
with the results obtained by assuming skewed and fat-tail distri-
butions with constant asymmetric and kurtosis parameters. They
found that the accuracy of the VaR estimates improved when time-
varying asymmetric and kurtosis parameters are considered. These
studies suggest that within the context of the Parametric method,
techniques that model the dynamic performance of the high-order
conditional moments (asymmetry and kurtosis) provide better
results than those considering functions with constant high-order
moments.
2.3. Semi-parametric methods
The Semi-parametric methods combine the Non-parametric
approach with the Parametric approach. The most important
Semi-parametric methods are Volatility-weight Historical Simula-
tion, Filtered Historical Simulation (FHS), CaViaR method and the
approach based on Extreme Value Theory.
8
Although in different contexts, approximating the distribution of asset returns
via the GCE has been previously employed in the literature (e.g., Jarrow and Rudd,
1982; Baillie and Bollerslev, 1992; Jondeau and Rockinger, 2001; Leon et al., 2005;
Christoffersen and Diebold, 2006).
2.3.1. Volatility-weight historical simulation
Traditional Historical Simulation does not take any recent
changes in volatility into account. Thus, Hull and White (1998)
proposed a new approach that combines the benet of Historical
Simulation with volatility models. The basic idea of this approach
is to update the return information to take into account the recent
changes in volatility.
Let r
t,i
be the historical return on asset i on day t in our historical
sample, o
t,i
be the forecast of the volatility
9
of the return on asset i
for dayt made at the endof t 1, ando
T,i
be our most recent forecast
of the volatility of asset i. Then, we replace the return in our data
set, r
t,i
, with volatility-adjusted returns, as given by:
r
t,i
=
o
1,i
r
t,i
o
t,i
(10)
According to this newapproach, the VaR() is the quantile of
the empirical distribution of the volatility adjusted return (r
t,i
).
This approach directly takes into account the volatility changes,
whereas the Historical Simulation approach ignores them. Further-
more, this method produces a risk estimate that is appropriately
sensitive to current volatility estimates. The empirical evidence
presented by Hull and White (1998) indicates that this approach
produces a VaR estimate superior to that of the Historical Simula-
tion approach.
2.3.2. Filtered Historical Simulation
Filtered Historical Simulation was proposed by Barone-Adesi
et al. (1999). This method combines the benets of Historical
Simulation with the power and exibility of conditional volatility
models.
Suppose we use Filtered Historical Simulation to estimate the
VaRof a single-asset portfolioover a 1-dayholdingperiod. Inimple-
menting this method, the rst step is to t a conditional volatility
model to our portfolio return data. Barone-Adesi et al. (1999) rec-
ommended an asymmetric GARCH model. The realised returns
are then standardised by dividing each one by the corresponding
volatility, z
t
= (
t
]o
t
). These standardised returns should be inde-
pendent and identically distributed and therefore be suitable for
Historical Simulation. The third step consists of bootstrapping a
large number L of drawings fromthe above sample set of standard-
ised returns.
Assuming a 1-day VaR holding period, the third stage involves
bootstrapping fromour data set of standardised returns: we take a
large number of drawings from this data set, which we now treat
as a sample, replacing each one after it has been drawn and mul-
tiplying each such randomdrawing by the volatility forecast 1 day
ahead:
r
t
=
t
+z
o
t+1
(11)
where z* is the simulated standardised return. If we take M draw-
ings, we therefore obtain a sample of M simulated returns. With
this approach, the VaR() is the % quantile of the simulated return
sample.
Recent empirical evidence shows that this approachworks quite
well in estimating VaR (see Barone-Adesi and Giannopoulos, 2001;
Barone-Adesi et al., 2002; Zenti and Pallotta, 2001; Pritsker, 2001;
Giannopoulos and Tunaru, 2005). As for other methods, Zikovic
and Aktan (2009), Angelidis et al. (2007), Kuester et al. (2006) and
Marimoutou et al. (2009) provide evidence that this method is the
best for estimating the VaR. However, other papers indicate that
this approach is not better than any other (see Nozari et al., 2010;
Alonso and Arcos, 2006).
9
To estimate the volatility of the returns, several volatility models can be used.
Hull and White (1998) proposed a GARCH model and the EWMA model.
2.3.3. CAViaR model
Engle and Manganelli (2004) proposed a conditional autore-
gressive specication for VaR. This approach is based on a quantile
estimation. Instead of modelling the entire distribution, they pro-
posed directly modelling the quantile. The empirical fact that the
volatilities of stock market returns cluster over time may be trans-
lated quantitatively in that their distribution is autocorrelated.
Consequently, the VaR, which is tightly linked to the standard
deviation of the distribution, must exhibit similar behaviour. A
natural way to formalise this characteristic is to use some type
of autoregressive specication. Thus, they proposed a conditional
autoregressive quantile specication that they called the CAViaR
model.
Let r
t
be a vector of time t observable nancial return and
a p-
vector of unknownparameters. Finally, let VaR
t
() VaR
t
(r
t1
,
)
bethequantileof thedistributionof theportfolioreturnformedat
timet 1, wherewesuppress thesubscript from
for notational
convenience.
A generic CAViaR specication might be the following:
VaR
t
() =
0
+
q
i=1
i
vuR
ti
() +
r
j=1
j
l(x
tj
) (12)
where p=q+r +1 is the dimension of and l is a function of a nite
number of lagged observable values. The autoregressive terms
i
VaR
ti
() i =1,. . ., q ensure that the quantile changes smoothly
over time. The role of l(r
t 1
) is to link VaR
t
() to observable vari-
ables that belong to the information set. A natural choice for x
t 1
is lagged returns. The advantage of this method is that it makes no
specic distributional assumption on the return of the asset. They
suggested that the rst order is sufcient for practical use:
VaR
t
() =
0
+
1
VaR
t1
() +
2
l(r
ti
, VaR
t1
) (13)
In the context of CAViaR model, different autoregressive speci-
cations can be considered
- Symmetric absolute value (SAV):
VaR
t
() =
0
+
1
VaR
t1
() +
2
|r
t1
| (14)
- Indirect GARCH(1,1) (IG):
VaR
t
() = (
0
+
1
VaR
2
t1
() +
2
(r
t1
)
2
)
1]2
(15)
In these two models the effects of the extreme returns and
the variance on the VaR measure are modelled symmetrically. To
account for nancial market asymmetry, via the leverage effect
(Black, 1976), the SAVmodel was extendedinEngle andManganelli
(2004) to asymmetric slope (AS):
VaR
t
() =
0
+
1
VaR
t1
() +
2
(r
t1
)
+
+
3
(r
t1)
)
(16)
In this representation, (r)
+
=max(r
t
,0) and (r
t
)
=min(r
t
,0) are
used as the functions.
The parameters of the CAViaR models are estimated by regres-
sion quantiles, as introduced by Koenker and Basset (1978). They
showed how to extend the notion of a sample quantile to a lin-
ear regression model. In order to capture leverage effects and other
nonlinear characteristics of the nancial return, some extensions of
the CAViaR model have been proposed. Yu et al. (2010) extend the
CAViaR model to include the Threshold GARCH (TGARCH) model
(an extension of the double threshold ARCH (DTARCH) of Li and
Li (1996)) and a mixture (an extension of Wong and Li (2001)s
mixture ARCH). Recently, Chen et al. (2011) proposed a non-linear
dynamic quantile family as a natural extension of the AS model.
Although empirical literature on CAViaR method is not exten-
sive, the results seemto indicate that the CAViaR model proposed
by Engle and Manganelli (2004) fails to provide accurate VaR esti-
mate although it may provide accurate VaR estimates in a stable
period (see Bao et al., 2006; Polanski and Stoja, 2009). However,
somerecent extensions of theCaViaRmodel suchas thoseproposed
by Gerlach et al. (2011) and Yu et al. (2010) work pretty well in esti-
mating VaR. As inthe case of the Parametric method, it appears that
when use is made of an asymmetric version of the CaViaR model
the VaR estimate notably improves. The paper of Sener et al. (2012)
supports this hypothesis. In a comparison of several CaViaR models
(asymmetric and symmetric) they nd that the asymmetric CaViaR
model outperforms the result from the standard CaViaR model.
Gerlach et al. (2011) compared three CAViaR models (SAV, AS and
Threshold CAViaR) with the Parametric method using different
volatility GARCHfamily models (GARCH-Normal, GARCH-Student-
t, GJR-GARCH, IGARCH, Riskmetric). At the 1% condence level, the
Threshold CAViaR model performs better than any other.
2.3.4. Extreme Value Theory
The Extreme Value Theory (EVT) approach focuses on the
limiting distribution of extreme returns observed over a long time
period, which is essentially independent of the distribution of the
returns themselves. The two main models for EVT are (1) the
block maxima models (BM) (McNeil, 1998) and (2) the peaks-over-
threshold model (POT). The second is generally considered to be
the most useful for practical applications due to the more efcient
use of the data at extreme values. In the POT models, there are two
types of analysis: the Semi-parametric models built around the Hill
estimator and its relatives (Beirlant et al., 1996; Danielsson et al.,
1998) and the fully Parametric models based on the Generalised
Pareto distribution (Embrechts et al., 1999). In the coming sections
each one of these approaches is described.
2.3.4.1. Block Maxima Models (BMM). This approach involves split-
ting the temporal horizon into blocks or sections, taking into
account the maximumvalues in each period. These selected obser-
vations formextreme events, also called a maximumblock.
The fundamental BMM concept shows howto accurately choose
the length period, n, and the data block within that length. For
values greater than n, the BMM provides a sequence of maximum
blocks M
n,1
,. . ., M
n,m
that can be adjusted by a generalised distri-
bution of extreme values (GEV). The maximumloss within a group
of n data is dened as M
n
=max(X
1
, X
2
,. . ., X
n
).
For a group of identically distributed observations, the distribu-
tion function of M
n
is represented as:
P(M
n
x) = P(X
1
x, . . ., X
n
x) =
n
i=1
F(x) = F
n
(x) (17)
The asymptotic approach for F
n
(x) is based on the maximum
standardised value
Z
n
=
M
n
n
o
n
(18)
where
n
and o
n
are the location and scale parameters, respec-
tively. The theorem of Fisher and Tippet establishes that if Z
n
converges toanon-degenerateddistribution, this distributionis the
generalised distribution of the extreme values (GEV). The algebraic
expression for such a generalised distribution is as follows:
H
,,o
(x) =
_
exp(1 +(x )]o)
1]
/ = 0 and (1 +(x )]o) :0
exp(c
x
) = 0
(19)
where o >0, <<, and < <. The parameter is known
as the shape parameter of the GEV distribution, and q=
1
is the
index of the tail distribution, H. The prior distribution is actually a
generalisation of the three types of distributions, depending on the
value taken by : Gumbel type I family ( =0), Frchet type II family
( >0) and Weibull type III family ( <0). The , o and parameters
are estimated using maximum likelihood. The VaR expression for
the Gumbel and Frchet distribution is as follows:
VaR =
_
o
n
n
(1 (n ln())
n
) to :0 (Fr chet)
n
o
n
ln(nln()) to = 0 (Gumbel)
(20)
In most situations, the blocks are selected in such a way that
their length matches a year interval and n is the number of obser-
vations within that year period.
This method has been commonly used in hydrology and engi-
neering applications but is not very suitable for nancial time
series due to the cluster phenomenon largely observed in nancial
returns.
2.3.4.2. Peaks over threshold models (POT). The POT model is gen-
erally considered to be the most useful for practical applications
due to the more efcient use of the data for the extreme values. In
this model, we can distinguish between two types of analysis: (a)
the fully Parametric models based on the Generalised Pareto dis-
tribution (GPD) and (b) the Semi-parametric models built around
the Hill estimator.
(a) Generalised Pareto Distribution
Among the random variables representing nancial returns
(r
1
, r
2
, r
3
,. . ., r
n
), we choose a low threshold u and examine all
values (y) exceeding u: (y
1
, y
2
, y
3
, . . ., y
Nu
), where y
i
=r
i
u and
N
u
are the number of sample data greater than u. The distribu-
tion of excess losses over the threshold u is dened as:
F
u
(y) = P(r u -y|r :u) =
F(y +u) F(u)
1 F(u)
(21)
Assuming that for a certain u, the distribution of excess
losses above the threshold is a Generalised Pareto Distribu-
tion, C
k,
(y) = 1 [1 +(k])y]
1]k
, the distribution function of
returns is given by:
F(r) = F(y +u) = [1 F(u)]C
k,o
(y) +F(u) (22)
To construct a tail estimator from this expression, the only
additional element we need is an estimation of F(u). For this
purpose, we take the obvious empirical estimator (uN
u
)/u.
We then use the historical simulation method. Introducing the
historical simulation estimate of F(u) and setting r =y +u in the
equation, we arrive at the tail estimator
F(r) = 1
N
u
n
_
1 +
k
(r u)
_1]k
r :u (23)
For a given probability >F(u), the VaR estimate is calculated
by inverting the tail estimation formula to obtain
VaR() = u +

k
_
_
n
N
u
(1 )
_
k
1
_
(24)
None of the previous Extreme Value Theory-based meth-
ods for quantile estimation yield VaR estimates that reect
the current volatility background. These methods are called
Unconditional Extreme Value Theory methods. Given the condi-
tional heteroscedasticity characteristic of most nancial data,
McNeil and Frey (2000) proposed a new methodology to esti-
mate the VaR that combines the Extreme Value Theory with
volatility models, known as the Conditional Extreme Value The-
ory. These authors proposed GARCH models to estimate the
current volatility and Extreme Value Theory to estimate the
distributions tails of the GARCH model shocks.
If the nancial returns are a strictly stationary time series and
follows a Generalised Pareto Distribution, denoted by G
k,o
(),
the conditional quantile of the returns can be estimated as
VaR() = +o
t
C
1
k,o
() (25)
where o
2
t
represents the conditional variance of the nancial
returns and C
1
k,o
() is the quantile of the GPD, which can be
calculated as:
C
1
k,o
() = u +

k
_
_
n
N
u
(1 )
_
k
1
_
(26)
(b) Hill estimator
The parameter that collects the features of the tail distribu-
tion is the tail index, q =
1
. Hill proposed a denition of the
tail index as follows:
q
H
=
_
1
u
_
u
i=1
log(r
i
) log r
u+1
__
1
(27)
where r
u
represents the threshold return and u is the num-
ber of observations equal to or less than the threshold return.
Thus, the Hill estimator is the mean of the most extreme u
observations minus u+1 observations (r
u+1
). Additionally, the
associated quantile estimator is (see Danielsson and de Vries,
2000):
vuR() = r
u+1
_
1
u]n
_
1]q
(28)
The problemposed by this estimator is the lack of any analytical
means to choose the threshold value of u in an optimum manner.
Hence, as an alternative, the procedure involves using the feature
known as Hill graphics. Different values of the Hill index are cal-
culated for different u values; the Hill estimator values become
represented in a chart or graphic based on u, and the u value is
selected from the region where the Hill estimators are relatively
stable (Hill chart leaning almost horizontally). The underlying intu-
itive idea posed in the Hill chart is that as u increases, the estimator
variance decreases, and thus, the bias is increased. Therefore, the
ability to foresee a balance between both trends is likely. When this
level is reached, the estimator remains constant.
Existing literature on EVT models to calculate VaR is abundant.
Regarding BMM, Silva and Melo (2003) considered different maxi-
mumblock widths, with results suggesting that the extreme value
method of estimating the VaR is a more conservative approach
for determining the capital requirements than traditional meth-
ods. Bystrm (2004) applied both unconditional and conditional
EVT models to the management of extreme market risks in the
stock market and found that conditional EVT models provided par-
ticularly accurate VaR measures. In addition, a comparison with
traditional Parametric (GARCH) approaches to calculate the VaR
demonstratedEVTas beingthe superior approachfor bothstandard
and more extreme VaR quantiles. Bekiros and Georgoutsos (2005)
conducted a comparative evaluation of the predictive performance
of various VaR models, with a special emphasis on two method-
ologies related to the EVT, POT and BM. Their results reinforced
previous results and demonstrated that some traditional meth-
ods might yieldsimilar results at conventional condencelevels but
that the EVT methodology produces the most accurate forecasts of
extreme losses at very high condence levels. Tolikas et al. (2007)
compared EVT with traditional measures (Parametric method, HS
and Monte Carlo) and agreed with Bekiros and Georgoutsos (2005)
onthe outperformance of the EVT methods comparedwiththe rest,
especially at very high condence levels. The only model that had
a performance comparable with that of the EVT is the HS model.
Some papers showed that unconditional EVT works better than
the traditional HS or Parametric approaches when a normal dis-
tribution for returns is assumed and a EWMA model is used to
estimate the conditional volatility of the return (see Danielsson
and de Vries, 2000). However, the unconditional version of this
approach has not been profusely used in the VaR estimation
because such an approach has been overwhelmingly dominated
by the conditional EVT (see McNeil and Frey, 2000; Ze-To, 2008;
Velayoudoumet al., 2009; Abad and Benito, 2013). Recent compar-
ative studies of VaRmodels, suchas Nozari et al. (2010), Zikovic and
Aktan (2009) and Genc ay and Selc uk (2004), showthat conditional
EVT approaches perform the best with respect to forecasting the
VaR.
Within the POT models, an environment has emerged in
which some studies have proposed some improvements on cer-
tain aspects. For example, Brooks et al. (2005) calculated the VaR
by a semi-nonparametric bootstrap using unconditional density, a
GARCH(1,1) model and EVT. They proposed a Semi-nonparametric
approach using a GPD, and this method was shown to generate
a more accurate VaR than any other method. Marimoutou et al.
(2009) used different models and conrmed that the ltering pro-
cess was important for obtaining better results. Ren and Giles
(2007) introduced the media excessive function concept as a new
way to choose the threshold. Ze-To (2008) developed a new con-
ditional EVT-based model combined with the GARCH-Jump model
to forecast extreme risks. He utilised the GARCH-Jump model to
asymmetrically provide the past realisation of jump innovation to
the future volatility of the return distribution as feedback and also
used the EVT to model the tail distribution of the GARCH-Jump-
processed residuals. The model is compared with unconditional
EVT and conditional EVT-GARCH models under different distri-
butions, normal and t-Student. He shows that the conditional
EVT-GARCH-Jump model outperforms the GARCH and GARCH-t
models. Chan and Gray (2006) proposed a model that accommo-
dates autoregression and weekly seasonals in both the conditional
mean and conditional volatility of the returns as well as leverage
effects via an EGARCH specication. In addition, EVT is adopted to
explicitly model the tails of the return distribution.
Finally, concerning the Hill index, some authors used the men-
tioned estimator, such as Bao et al. (2006), whereas others such as
Bhattacharyya and Ritolia (2008) used a modied Hill estimator.
2.3.5. Monte Carlo
The simplest Monte Carlo procedure to estimate the VaRondate
t on a one-day horizon at a 99% signicance level consists of sim-
ulating N draws from the distribution of returns on date t +1. The
VaR at a 99% level is estimated by reading off element N/100 after
sorting the N different draws from the one-day returns, i.e., the
VaR estimate is estimated empirically as the VaR quantile of the
simulated distribution of returns.
However, applyingsimulations toa dynamic model of riskfactor
returns that capture path dependent behaviour, such as volatility
clustering, and the essential non-normal features of their multi-
variate conditional distributions is important. With regard to the
rst of these, one of the most important features of high-frequency
returns is that volatility tends to come in clusters. In this case, we
canobtainthe GARCHvariance estimate at time t( o
t
) using the sim-
ulated returns in the previous simulation and set r
1
= o
t
z
t
, where
z
t
is a simulation from a standard normal variable. With regard
to the second item, we can model the interdependence using the
standard multivariate normal or t-Student distribution or use cop-
ulas instead of correlation as the dependent metric.
Monte Carlo is an interesting technique that can be used to esti-
mate the VaR for non-linear portfolios (see Estrella et al., 1994)
because it requires no simplifying assumptions about the joint dis-
tribution of the underlying data. However, it involves considerable
computational expenses. This cost has been a barrier limiting its
application into real-world risk containment problems. Srinivasan
andShah(2001) proposedalternative algorithms that require mod-
est computational costs and, Antonelli and Iovino (2002) proposed
a methodology that improves the computational efciency of the
Monte Carlo simulation approach to VaR estimates.
Finally, the evidence shown in the studies on the comparison
of VaR methodologies agree with the greater accuracy of the VaR
estimations achievedby methods other thanMonte Carlo(see Abad
andBenito, 2013; Huang, 2009; Tolikas et al., 2007; Baoet al., 2006).
To sum up, in this section we have reviewed some of the most
important VaR methodologies, from the standard models to the
more recent approaches. From a theoretical point of view, all of
these approaches show advantages and disadvantages. In Table 3,
we resume these advantages and disadvantages. In the next sec-
tions, we will reviewthe results obtained for these methodologies
froma practical point of view.
3. Back-testing VaR methodologies
Many authors are concerned about the adequacy of the VaR
measures, especially when they compare several methods. Papers,
which compare the VaR methodologies commonly use two alter-
native approaches: the basis of the statistical accuracy tests and/or
loss functions. As to the rst approach, several procedures based on
statistical hypothesis testing have been proposed in the literature
and authors usually select one or more tests to evaluate the accu-
racy of VaR models and compare them. The standard tests about
the accuracy VaR models are: (i) unconditional and conditional
coverage tests; (ii) the back-testing criterion and (iii) the dynamic
quantile test. To implement all these tests an exception indicator
must be dened. This indicator is calculated as follows:
l
t+1
=
_
1 if r
t+1
-VaR()
0 if r
t+1
:VaR()
(29)
Kupiec (1995) shows that assuming the probability of an excep-
tion is constant, then the number of exceptions x =
l
t+1
follows
a binomial distribution B(N, ), where N is the number of observa-
tions. Anaccurate VaR() measure shouldproduce anunconditional
coverage ( =
l
t+1
]N) equal to percent. The unconditional cov-
erage test has as a null hypothesis = , with a likelihood ratio
statistic:
LR
UC
= 2[log(
x
(1
)
Nx
) log(
x
(1 )
Nx
)] (30)
which follows an asymptotic y
2
(1) distribution.
Christoffersen (1998) developed a conditional coverage test. This
jointly examines whether the percentage of exceptions is statis-
tically equal to the one expected and the serial independence of
I
t +1
. He proposed an independence test, which aimed to reject VaR
models with clustered violations. The likelihood ratio statistic of
the conditional coverage test is LR
cc
=LR
uc
+LR
ind
, which is asymp-
totically distributed y
2
(2), and the LR
ind
statistic is the likelihood
ratio statistic for the hypothesis of serial independence against
rst-order Markov dependence.
10
10
The LR
ind
statistic is LR
ind
=2 [logL
A
logL
0
] and has an asymptotic y
2
(1)
distribution. The likelihood function under the alternative hypothesis is L
/
=
(1
01
)
N
00
N01
01
(1
11
)
N
10
N
11
11
where N
ij
denotes the number of observations in
state j after having been in state i in the previous period,
01
= N
01
](N
00
+N
01
)
and
11
= N
11
](N
10
+N
11
). The likelihood function under the null hypothesis is
(
01
=
11
= = (N
11
+N
01
)]N) is L
0
= (1 )
N
00
+N
01
N
01
+N
11
.
Table 3
Advantages and disadvantages of VaR approaches.
Advantages Disadvantages
Non Parametric approach (HS)
Minimal assumptions made about the error distribution,
nor the exact formof the dynamic specications
Not making strong assumptions
about the distribution of the returns
portfolio, they can accommodate wide
tails, skewness and any other
non-normal features.
It is very easy to implement.
Its results are completely dependent
on the data set.
It is sometimes slowto reect major
events.
It is only allows us to estimate VaR at
discrete condence intervals
determined by the size of our data set.
Parametric approach
Makes a full parametric distributional and model form
assumption. For example AGARCH model with Gaussian
errors
Its ease of implementation when a
normal or Student-t distributions is
assumed.
It ignores leptokurtosis and skewness
when a normal distribution is assumed.
Difculties of implementation when a
skewed distributions is assumed
a
.
Riskmetrics
A kind of parametric approach
Its ease of implementation can be
calculated using a spreadsheet.
It assumes normality of return
ignoring fat tails, skewness, etc.
This model lack non linear property
which is a signicant of the nancial
return.
Semiparametric approach
Some assumptions are made, either
about the error distribution, its
extremes, or the model dynamics
Filter Historical
Simulation
This approach retains the
nonparametric advantage (HS) and at
the same time addresses some of HSs
inherent problems, i.e. FHS take
volatility background into account.
Its results slightly dependent on the
data set.
ETV Capture curtosis and changes in
volatility (conditional ETV).
It depends on the extreme return
distribution assumption.
Its results depend on the extreme
data set.
CaViaR It makes no specic distributional
assumption on the return of the asset.
It captures non linear characteristics
of the nancial returns.
Difculties of implementation.
Monte Carlo The large number of scenarios
generated provide a more reliable and
comprehensive measure of risk than
analytical method.
It captures convexity of non linear
instruments and changes in volatility
and time.
Its reliance on the stochastic process
specied or historical data selected to
generate estimations of the nal value
of the portfolio and hence of the VaR.
It involves considerable
computational expenses.
a
In addition, as the skewness distributions are not included in any statistical package, the user of this methodology have to programtheir code of estimation To do that,
several programlanguage can be used: MATLAB, R, GAUSS, etc. It is in this sense we say that the implementation is difcult. As the maximisation of the likelihood based on
several skewed distributions, such as, SGT is very complicated so that it can take a lot of computational time.
A similar test for the signicance of the departure of from is
the back-testing criterion statistic:
Z =
(N
N)
_
N(1 )
(31)
which follows an asymptotic N(0,1) distribution.
Finally, the Dynamic Quantile (DQ) test proposed by Engle and
Manganelli (2004) examines whether the exception indicator is
uncorrelated with any variable that belongs to the information set
t1
available when the VaR was calculated. This is a Wald test of
the hypothesis that all slopes in the regression model
l
t
=
0
+
p
i=1
i
l
t1
+
q
j=i
j
X
j
+
t
(32)
are equal to zero, where X
j
are explanatory variables contained in
t1
.VaR() is usually an explanatory variable to test if the proba-
bility of an exception depends on the level of the VaR.
The tests described above are based on the assumption that the
parameters of the models tted to estimate the VaR are known,
although they are estimations. Escanciano and Olmo (2010) show
that the use of standard unconditional and independence backtest-
ing procedures can be misleading. They propose a correction of the
standard backtesting procedures. Additionally, Escanciano and Pei
(2012) propose correction when VaR is estimated with HS or FHS.
On a different route, Baysal and Staum (2008) provide a test on
the coverage of condence regions and intervals involving VaR and
Conditional VaR.
The second approach is based on the comparison of loss func-
tions. Some authors compare the VaR methodologies by evaluating
the magnitude of the losses experienced when an exception occurs
in the models. The magnitude loss function that addresses the
magnitude of the exceptions was developed by Lopez (1998, 1999).
It deals with a specic concern expressed by the Basel Committee
on Banking Supervision, which noted that the magnitude, as well
as the number of VaR exceptions is a matter of regulatory con-
cern. Furthermore, the loss function usually examines the distance
between the observed returns and the forecasted VaR() values if
an exception occurs.
Lopez (1999) proposed different loss functions:
lj
t+1
=
_
z(r
t+1
, VaR()) if r
t+1
-VaR()
0 if r
t+1
:VaR()
(33)
where the VaR measure is penalised with the exception indi-
cator (z(.) =1), the exception indicator plus the square distance
(z(.) =1+(r
t +1
VaR())
2
or using weight (z(r
t +1
,VaR()x) =k,
where x, being the number of exceptions, is divided into several
zones and k is a constant which depends on zone) based on what
regulators consider to be a minimum capital requirement reect-
ing their concerns regarding prudent capital standards and model
accuracy.
More recently, other authors have proposed loss func-
tion alternatives, such as Abad and Benito (2013) who
consider z(.) =(r
t +1
VaR())
2
and z(.) =|r
t +1
VaR()| or
Caporin (2008) who proposes z(.) =
r
t+1
]VaR()
and
z(.) = ((|r
t+1
| |vuR()|)
2
]|vuR()|). Caporin (2008) also designs a
measure of the opportunity cost.
In this second approach, the best VaR model is that which min-
imises the loss function. In order to knowwhich approach provides
minimumlosses, different tests can be used. For instance Abad and
Benito (2013) use a non-parametric test while Sener et al. (2012)
use the Diebold and Mariano (1995) test as well as that of White
(2000).
Another alternative to compare VaR models is to evaluate the
loss in a set of hypothetical extreme market scenarios (stress test-
ing). Linsmeier and Pearson (2000) discuss the advantages of stress
testing.
4. Comparison of VaR methods
Empirical literature on VaR methodology is quite extensive.
However, there are not many papers dedicated to comparing the
performance of a large range of VaR methodologies. In Table 4,
we resume 24 comparison papers. Basically, the methodologies
compared in these papers are HS (16 papers), FHS (8 papers),
the Parametric method under different distributions (22 papers
included the normal, 13 papers include t-Student and just 5 papers
include any kind of skewness distribution) and the EVT based
approach (18 papers). Only a few of these studies include other
methods, such as the Monte Carlo (5 papers), CaViaR (5 papers) and
the Non-Parametric density estimation methods (2 papers) in their
comparisons. For each article, we marked the methods included in
the comparative exercise with a cross and shaded the method that
provides the best VaR estimations.
The approach based on the EVT is the best for estimating the
VaR in 83.3% of the cases in which this method is included in
the comparison, followed closely by FHS, with 62.5% of the cases.
This percentage increases to 75.0% if we consider that the differ-
ences between ETV and FHS are almost imperceptible in the paper
of Giamouridis and Ntoula (2009), as the authors underline. The
CaViaR method ranks third. This approach is the best in 3 out of 5
comparison papers (it represents a percentage of success of 60%,
which is quite high). However, we must remark that only in one of
these 3 papers, ETV is included in the comparison and FHS is not
included in any of them.
The worst results are obtained by HS, Monte Carlo and Riskmet-
rics. None of those methodologies rank best in the comparisons
where they are included. Furthermore, in many of these papers
HS and Riskmetrics perform worst in estimating VaR. A similar
percentage of success is obtained by Parametric method under a
normal distribution. Only in 2 out of 18 papers, does this method-
ology rank best in the comparison. It seems clear that the new
proposals to estimate VaR have outperformed the traditional ones.
Taking this into account, we highlight the results obtained by
Berkowitz and OBrien (2002). In this paper the authors compare
some internal VaR models used by banks with a parametric GARCH
model estimated under normality. They nd that the bank VaR
models are not better than a simple parametric GARCH model. It
reveals that internal models work very poorly in estimating VaR.
The results obtained by the Parametric method should be
taken into account when the conditional high-order moments are
time-varying. The two papers that include this method in the com-
parison obtained a 100% outcome success (see Ergun and Jun,
2010; Polanski and Stoja, 2010). However, only one of these papers
included EVT in the comparison (Ergun and Jun, 2010).
Although not shown in Table 4, the VaR estimations obtained
by the Parametric method with asymmetric and leptokurtic distri-
butions and in a mixed-distribution context are also quite accurate
(see Abad and Benito, 2013; Bali and Theodossiou, 2007, 2008; Bali
et al., 2008; Chen et al., 2011; Polanski and Stoja, 2010). However
this method does not seemto be superior to EVT and FHS (Kuester
et al., 2006; Cifter and zn, 2007; Angelidis et al., 2007). Never-
theless, there are not many papers including these three methods
in their comparison. In this line, some recent extensions of the
CaViaR method seem to perform quite well, such as those pro-
posed by Yu et al. (2010) and Gerlach et al. (2011). This last paper
compared three CAViaR models (SAV, AS and Threshold CAViaR)
with the Parametric model under some distributions (GARCH-N,
GARCH-t, GJR-GARCH, IGARCH, Riskmetric). They nd that at 1%
condence level, the ThresholdCAViaRmodel performs better than
the Parametric models considered. Sener et al. (2012) carried out a
comparison of a large set of VaR methodologies: HS, Monte Carlo,
EVT, Riskmetrics, Parametric method under normal distribution
and four CaViaR models (symmetric and asymmetric). They nd
that the asymmetric CaViaR model joined to the Parametric model
with an EGARCH model for the volatility performs the best in esti-
mating VaR. Abad and Benito (2013), in a comparison of a large
range of VaR approaches that include EVT, nd that the Parametric
method under an asymmetric specication for conditional volatil-
ity and t-Student innovations performs the best in forecasting VaR.
Both papers highlight the importance of capturing the asymmetry
in volatility. Sener et al. (2012) state that the performance of VaR
methods does not depend entirely on whether they are parametric,
non-parametric, semi-parametric or hybrid but rather on whether
they can model the asymmetry of the underlying data effectively
or not.
In Table 5, we reconsider the papers of Table 4 to show which
approachtheyusetocompareVaRmodels. Most of thepapers (62%)
evaluate the performance of VaR models on the basis of the fore-
casting accuracy. To do that not all of them used a statistical test.
There is a signicant percentage (25%) comparing the percentage
of exceptions with that expected without using any statistical test.
38% of the papers in our sample consider that both the number of
exceptions and their size are important and include both dimen-
sions in their comparison.
Although there are not many articles dedicated to the com-
parison of a wide range of VaR methodologies, the existing ones
offer quite conclusive results. These results showthat the approach
based on the EVT and FHS is the best method to estimate the VaR.
We also note that VaR estimates obtained by some asymmetric
extensions of CaViaRmethodandthe Parametric methodunder the
skewed and fat-tail distributions lead to promising results, espe-
cially when the assumption that the standardised returns is iid is
abandoned and the conditional high-order moments are consid-
ered to be time-varying.
5. Some important topics of VaR methodology
As we stated in the introduction, VaR is by far the leading mea-
sure of portfolioriskinuse inmajor commercial banks andnancial
institutions. However, this measurement is not exempt from crit-
icism. Some researchers have remarked that VaR is not a coherent
market measure (see Artzner et al., 1999). These authors dene a
set of criteria necessary for what they call a coherent risk mea-
surement. These criteria include homogeneity (larger positions are
associated with greater risk), monotonicity (if a portfolio has sys-
tematically lower returns thananother for all states of the world, its
risk must be greater), subadditivity (the risk of the sumcannot be
greater than the sumof the risk) and the risk free condition (as the
proportion of the portfolio invested in the risk free asset increases,
portfolio risk should decline). They showthat VaR is not a coherent
risk measure because it violates one of their axioms. In particu-
lar VaR does not satisfy the subadditivity condition and it may
discourage diversication. On this point Artzner et al. (1999) pro-
posedanalternative risk measure relatedtoVaRwhichis calledTail
Table 4
Overviewof papers that compare VaR methodologies: what methodologies compare?
HS FHS RM Parametric approaches ETV CF CaViaR MC N-P
N T SSD MN HOM
Abad and Benito (2013)
Gerlach et al. (2011)
Sener et al. (2012)
Ergun and Jun (2010)
Nozari et al. (2010)
Polanski and Stoja (2010)
Brownlees and Gallo (2010)
Yu et al. (2010)
Ozun et al. (2010)
Huang (2009)
Marimoutou et al. (2009)
Zikovic and Aktan (2009)
Giamouridis and Ntoula (2009)
Angelidis et al. (2007)
Tolikas et al. (2007)
Alonso and Arcos (2006)
Bao et al. (2006)
Bhattacharyya and Ritolia (2008)
Kuester et al. (2006)
Bekiros and Georgoutsos (2005)
Genc ay and Selc uk (2004)
Genc ay et al. (2003)
Darbha (2001)
Danielsson and de Vries (2000)
Note: In this table, we present some empirical papers involving comparisons of VaR methodologies. The VaR methodologies are marked with a cross when they are included
in a paper. A shaded cell indicates the best methodology to estimate the VaR in the paper. The VaR approaches included in these paper are the following: Historical Simulation
(HS); Filtered Historical Simulation (FHS); Riskmetrics (RM); Parametric approaches estimated under different distributions, including the normal distribution (N), t-Student
distribution (T), skewed t-Student distribution (SSD), mixed normal distribution (MN) and high-order moment time-varying distribution (HOM); Extreme Value Theory
(EVT); CaViaR method (CaViaR); Monte Carlo Simulation (MC); and non-parametric estimation of the density function (NP).
Conditional Expection, also called Conditional Value at Risk (CVaR).
The CVaR measures the expected loss in the % worst cases and is
given by
CVaR
t
=
t1
{R
t
|R
t
vuR
t
} (34)
The CVaR is a coherent measure of risk when it is restricted
to continuous distributions. However, it can violate sub-additivity
with non-continuous distributions. Consequently, Acerbi and
Tasche (2002) proposed the Expected Shortfall (ES) as a coherent
measure of risk. The ES is given by
S
t
= CVaR
t
+(z 1)(CVaR
t
VaR
t
) (35)
where z ( P
t1
[R
t
vuR
t
]])1. Note that CVaR=ES when the
distribution of returns is continuous. However, it is still coherent
Table 5
Overviewof papers that compare VaR methodologies: howdo they compare?
The accuracy Loss function
Abad and Benito (2013) LRuc-ind-cc, BT, DQ Quadratic
Gerlach et al. (2011) LRuc-cc, DQ
Sener et al. (2012) DQ Absolute
Ergun and Jun (2010) LRuc-ind-cc
Nozari et al. (2010) LRuc
Polanski and Stoja (2010) LRuc-ind
Brownlees and Gallo (2010) LRuc-ind-cc, DQ Tick loss function
Yu et al. (2010) %
Ozun et al. (2010) LRuc-ind-cc Quadratics Lopez
Huang (2009) LRuc
Marimoutou et al. (2009) LRuc-cc Quadratics Lopez
Zikovic and Aktan (2009) LRuc-ind-cc Lopez
Giamouridis and Ntoula (2009) LRuc-ind-cc
Angelidis et al. (2007) LRuc Quadratic
Tolikas et al. (2007) LRcc
Alonso and Arcos (2006) BT Quadratics Lopez
Bao et al. (2006) % Predictive quantile loss
Bhattacharyya and Ritolia (2008) LRuc
Kuester et al. (2006) LRuc-ind-cc, DQ
Bekiros and Georgoutsos (2005) LRuc-cc
Genc ay and Selc uk (2004) %
Genc ay et al. (2003) %
Darbha (2001) %
Danielsson and de Vries (2000) %
Note: In this table, we present some empirical papers involving comparisons of VaR methodologies. We indicate the test to evaluate the accuracy of VaR models and/or the
loss function used in the comparative exercise. LRuc is the unconditional coverage test. LRind is the statistic for the serial independence. LRcc is the conditional coverage test.
BT is the back-testing criterion. DQis the Dynamic Quantile test. % denotes the comparison of the percentage of exceptions with the expected percentage without a statistical
test.
when the distribution of returns is not continuous. The ES has also
several advantages when compared with the more popular VaR.
First of all, the ES is free of tail risk in the sense that it takes into
account information about the tail of the underlying distribution.
The use of a risk measure free of tail risk avoids extreme losses in
the tail. Therefore, the ES is anexcellent candidate for replacing VaR
for nancial risk management purposes.
Despite the advantages of ES, it is still less used than VaR.
The principal reason for this pretermission is that the ES back-
test is harder than VaR one. In that sense, in the last years
some ES backtesting procedures have been developed. We can
cite here the residual approach introduced by McNeil and Frey
(2000), the censored Gaussian approach proposed by Berkowitz
(2001), the functional delta approach of Kerkhof and Melenberg
(2004), and the saddlepoint technique introduced by Wong (2008,
2010).
However, these approaches present some drawbacks. The back-
tests of McNeil and Frey (2000), Berkowitz (2001) and Kerkhof
and Melenberg (2004) rely on asymptotic test statistics that might
be inaccurate when the sample size is small. The test proposed
by Wong (2008) is robust to these questions; nonetheless, it has
some disadvantages (as with the Gaussian distribution assump-
tion).
Regardless of the sector in which a nancial institution par-
ticipates, all such institutions are subject to three types of risk:
market, credit and operational. So, to calculate the total VaR of
a portfolio it is necessary to combine these risks. There are dif-
ferent approximations to carry this out. First, an approximation
that sums up the three types of risk (VaR). As VaR is not a sub-
additivity measure this approximation overestimates total risk or
economic capital. Second, assuming joint normality of the risk
factors, this approximation imposes tails that are thinner than
the empirical estimates andsignicantly underestimates economic
capital and the third approach to assess the risk aggregation is
based on using copulas. To obtain the total VaR of a portfolio it
is necessary to obtain the joint return distribution of the port-
folio. Copulas allow us to solve this problem by combining the
specic marginal distributions with a dependence function to cre-
ate this joint distribution. The essential idea of the copula approach
is that a joint distribution can be factored into the marginals and
a dependence function, called copula. The term copula is based
on the notion of coupling: the copula couples the marginal dis-
tributions together to form a joint distribution. The dependence
relation is entirely determined by the copula, while location, scal-
ing and shape are entirely determined by the marginals. Using a
copula, marginal risk that is initially estimated separately can then
be combined in a joint risk distribution preserving the marginals
original characteristics. This is sometimes referred to as obtaining a
joint density with predetermined marginals. The joint distribution
can then be used to calculate the quantiles of the portfolio return
distribution, since the portfolio returns are a weighted average
on individual returns. Embrechts et al. (1999, 2002) were among
the rst to introduce this methodology in nancial literature.
Some applications of copulas focussing on cross-risk aggregation
for nancial institutions can be found in Alexander and Pezier
(2003), Ward and Lee (2002) and Rosenberg and Schuermann
(2006).
6. Conclusion
In this article we reviewthe full range of methodologies devel-
oped to estimate the VaR, from standard models to the recently
proposedandpresent their relative strengths andweaknesses from
both theoretical and practical perspectives.
The performance of the parametric approach in estimating the
VaRdepends ontheassumeddistributionof thenancial returnand
on the volatility model used to estimate the conditional volatility
of the returns. As for the return distribution, empirical evidence
suggests that when asymmetric and fat-tail distributions are con-
sidered, the VaR estimate improves considerably. Regardless of
the volatility model used, the results obtained in the empirical
literature indicate the following: (i) The EWMA model provides
inaccurate VaR estimates, (ii) the performance of the GARCH mod-
els strongly depends on the assumption of returns distribution.
Overall, under a normal distribution, the VaR estimates are not
very accurate, but when asymmetric and fat-tail distributions are
applied, the results improve considerably, (iii) evidence suggests
with some exceptions that SV models do not improve the results
obtained by the family of GARCH models, (iv) the models based
on the realised volatility work quite well to estimate VaR, outper-
forming the GARCHmodels estimated under a normal distribution.
Additionally, Markov-Switching GARCH outperforms the GARCH
models estimated under normality. In the case of the realised
volatility models, some authors indicate that its superiority com-
pared with the GARCH family is not as high when the GARCH
models are estimated assuming asymmetric and fat-tail returns
distributions, (v) in the GARCH family, the fractional-integrated
GARCH models do not appear to be superior to the GARCH mod-
els. However, in the context of the realised volatility models,
there is evidence that models, which capture long memory in
volatility, provide more accurate VaR estimates, and (vi) although
evidence is somewhat ambiguous, asymmetric volatility models
appear to provide a better VaR estimate than symmetric mod-
els.
Althoughtherearenot manyworks dedicatedtothecomparison
of a wide range of VaR methodologies, the existing ones offer quite
conclusive results. These results show that the approach based on
the EVT and FHS is the best method to estimate the VaR. We also
note that VaRestimates obtained by some asymmetric extension of
CaViaR method and the Parametric method under the skewed and
fat-tail distributions lead to promising results, especially when the
assumption that the standardised returns is iid is abandoned and
that theconditional high-order moments areconsideredtobetime-
varying. It seems clear that the newproposals to estimate VaR have
outperformed the traditional ones.
To further the research, it would be interesting to explore
whether in the context of an approach based on the EVT and
FHS considering asymmetric and fat-tail distributions to model the
volatility of the returns could help to improve the results obtained
by these methods. Along this line, results may be further improved
by applying the realised volatility model and Markov-switching
model.
References
Abad, P., Benito, S., 2013. A detailed comparison of value at risk in inter-
national stock exchanges. Mathematics and Computers in Simulation,
http://dx.doi.org/10.1016/j.matcom.2012.05.011 (forthcoming).
Acerbi, C., Tasche, D., 2002. Onthecoherenceof expectedshortfall. Journal of Banking
and Finance 26, 14871503.
Alexander, C., Pezier, J., 2003. On the Aggregation of Firm-Wide Market and Credit
Risks. ISMA Centre Discussion Papers in Finance 13.
Alonso, J., Arcos, M., 2005. VaR: Evaluacin del Desempe no de Diferentes
Metodologas para Siete Pases Latinoamericanos. Borradores de economa y
nanzas.
Alonso, J., Arcos, M., 2006. Cuatro Hechos Estilizados de las Series de Rendimientos:
una Ilustracin para Colombia. Estudios Gerenciales 22, 103123.
Andersen, T., Bollerslev, T., Diebold, F., Ebens, H., 2001a. The distribution of realised
stock returns volatility. Journal of Financial Economics 61, 4376.
Andersen, T., Bollerslev, T., Das, A., 2001b. The distribution of realized exchange rate
volatility. Journal of the American Statistical Association 96, 4252.
Andersen, T., Bollerslev, T., Diebold, F., Labys, P., 2003. Modeling and forecasting
realized volatility. Econometrica 71, 529626.
Angelidis, T., Benos, A., 2004. Market risk incommodity markets: a switching regime
approach. Economic and Financial Modelling 11, 103148.
Angelidis, T., Benos, A., Degiannakis, S., 2007. A robust VaR model under differ-
ent time periods and weighting schemes. Review of Quantitative Finance and
Accounting 28, 187201.
Antonelli, S., Iovino, M.G., 2002. Optimization of Monte Carlo procedures for value
at risk estimates. Economic Notes 31, 5978.
Artzner, P., Delbaen, F., Eber, J.M., Heath, D., 1999. Coherent measures of risk. Math-
ematical Finance 9, 203228.
Asai, M., McAleer, M., Medeiros, M., 2011. Asymmetry and Leverage in
Realized Volatility, Available at SSRN: http://ssrn.com/abstract=1464350 or
http://dx.doi.org/10.2139/ssrn.1464350 (30.08.09).
Ashley, R., Randal, V., 2009. Frequency dependence in regression model coefcients:
an alternative approach for modeling nonlinear dynamics relationships in time
series. Econometric Reviews 28, 420.
Ausn, M., Galeano, P., 2007. Bayesian estimation of the Gaussian mixture GARCH
model. Computational Statistics & Data Analysis 51, 26362652.
Baillie, R., Bollerslev, T., 1992. Prediction in dynamic models with time-dependent
conditional variances. Journal of Econometrics 52, 91113.
Baillie, R., Bollerslev, T., Mikkelsen, H., 1996. Fractionally integrated generalized
autoregressiveconditional heteroskedasticity. Journal of Econometrics 74, 330.
Bali, T., Theodossiou, P., 2007. A conditional-SGT-VaR approach with alternative
GARCH models. Annals of Operations Research 151, 241267.
Bali, T., Weinbaum, D., 2007. A conditional extreme value volatility estimator
based on high-frequency returns. Journal of Economic Dynamics & Control 31,
361397.
Bali, T., Theodossiou, P., 2008. Risk measurement performance of alternative distri-
bution functions. Journal of Risk and Insurance 75, 411437.
Bali, T., Hengyong, M., Tang, Y., 2008. Theroleof autoregressiveconditional skewness
and kurtosis in the estimation of conditional VaR. Journal of Banking & Finance
32, 269282.
Bao, Y., Lee, T., Saltoglu, B., 2006. Evaluating predictive performance of value-at-
risk models in emerging markets: a reality check. Journal of Forecasting 25,
101128.
Barone-Adesi, G., Giannopoulos, K., 2001. Non-parametric VaR techniques. Myths
and realities. Economic Notes by Banca Monte dei Paschi di Siena, SpA. 30,
167181.
Barone-Adesi, G., Giannopoulos, K., Vosper, L., 1999. VaR without correlations for
nonlinear portfolios. Journal of Futures Markets 19, 583602.
Barone-Adesi, G., Giannopoulos, K., Vosper, L., 2002. Backtesting derivative portfo-
lios with ltered historical simulation (FHS). European Financial Management
8, 3158.
Baysal, R.E., Staum, J., 2008. Empirical likelihood for value-at-risk and expected
shortfall. Journal of Risk 11, 332.
Beder, T., 1995. VaR: seductive but dangerous. Financial Analysts Journal, 1224.
Beder, T., 1996. Report card on value at risk: high potential but slow starter. Bank
Accounting & Finance 10, 1425.
Beirlant, J., Teugels, J.L., Vyncker, P., 1996. Practical Analysis of Extreme Values.
Leuven University Press, Leuven, Belgium.
Bekiros, S., Georgoutsos, D., 2005. Estimation of value at risk by extreme value
and conventional methods: a comparative evaluation of their predictive per-
formance. Journal of International Financial Markets, Institutions & Money 15
(3), 209228.
Beltratti, A., Morana, C., 1999. Computing value at risk with high frequency data.
Journal of Empirical Finance 6, 431455.
Berkowitz, J., 2001. Testing density forecasts, with applications to risk management.
Journal of Business and Economic Statistics 19, 465474.
Berkowitz, J., OBrien, J., 2002. Howaccurate are value-at-risk models at commercial
banks? Journal of Finance 57, 10931111.
Bhattacharyya, M., Ritolia, G., 2008. Conditional VaR using ETV. Towards a
planned margin scheme. International Review of Financial Analysis 17,
382395.
Billio, M., Pelizzon, L., 2000. Value-at-risk: amultivariateswitchingregimeapproach.
Black, F., 1976. Studies in stock price volatility changes. In: Proceedings of the 1976
Business Meeting of the Business and Economics Statistics Section, American
Association, pp. 177181.
Bollerslev, T., 1986. Generalizedautoregressive conditional heteroscedasticity. Jour-
nal of Econometrics 21, 307327.
Bollerslev, T., 1987. A conditionally heteroskedastic time series model for spec-
ulative prices and rates of return. Review of Economics and Statistics 69,
542547.
Bollerslev, T., 2009. Glossary to ARCH(GARCH), Available at http://public.econ.duke.
edu/boller/
Bollerslev, T., Mikkelsen, H., 1996. Modeling and pricing long memory in stock mar-
ket volatility. Journal of Econometrics 73, 151184.
Bollerslev, T., Gibson, M., Zhou, H., 2011. Dynamics estimation of volatility risk pre-
mia and investor risk aversion from option-implied and realized volatilities.
Journal of Econometrics 160, 235245.
Breidt, F., Crato, N., de Lima, P., 1998. The detection and estimation of long memory
in stochastic volatility. Journal of Econometrics 83, 325348.
Brooks, C., Clare, A., Dalle Molle, J., Persand, G., 2005. A comparison of extreme value
theory approaches for determining value at risk. Journal of Empirical Finance
12, 339352.
Brownlees, C., Gallo, G., 2010. Comparisonof volatility measures: a risk management
perspective. Journal of Financial Econometrics 8, 2956.
Brownlees, C., Gallo, G., 2011. Shrinkageestimationof semiparametric multiplicative
error models. International Journal of Forecasting 27, 365378.
Butler, J.S., Schachter, B., 1998. Estimating value at risk with a precision measure by
combining kernel estimation with historical simulation. Review of Derivatives
Research 1, 371390.
Bystrm, H., 2004. Managing extreme risks in tranquil and volatile markets using
conditional extreme value theory. International Review of Financial Analysis,
133152.
Caporin, M., 2008. Evaluatingvalue-at-riskmeasures inthepresenceof longmemory
conditional volatility. Journal of Risk 10, 79110.
Carvalho, M., Freire, M., Medeiros, M., Souza, L., 2006. Modeling and forecasting
the volatility of Brazilian asset returns: a realized variance approach. Revista
Brasileira de Finanzas 4, 321343.
Cifter, A., zn, A., 2007. Nonlinear Combination of Financial Forecast with Genetic
Algorithm. MPRA Paper no 2488, posted 07. November 2007/02:31, Available
from: http://mpra.ub.uni-muenchen.de/2488/
Clements, M., Galvao, A., Kim, J., 2008. Quantile forecasts of daily exchange rate
returns from forecasts of realized volatility. Journal of Empirical Finance 15,
729750.
Chan, K., Gray, P., 2006. Using extreme value theory to measure value-at-risk for
daily electricity spot prices. International Journal of Forecasting, 283300.
Chen, X., Ghysels, E., 2010. News good or bad and its impact on volatility predic-
tions over multiple horizons. Reviewof Financial Studies 24, 4680.
Chen, C., So, M., Lin, E., 2009. Volatility forecasting with double Markov switching
GARCH models. Journal of Forecasting 28, 681697.
Chen, C., Gerlach, R., Lin, E., Lee, W., 2011. Bayesian forecasting for nancial risk
management, pre and post the global nancial crisis. Journal of Forecasting,
http://dx.doi.org/10.1002/for.1237 (published online in Wiley Online Library
(wileyonlinelibrary.com)).
Cheng, W., Hung, J., 2011. Skewness and leptokurtosis in GARCH-typed VaR esti-
mation of petroleum and metal asset returns. Journal of Empirical Finance 18,
160173.
Christoffersen, P., 1998. Evaluating interval forecasting. International Economic
Review39, 841862.
Christoffersen, P., Diebold, F., 2006. Financial asset returns, direction-of-changefore-
casting and volatility dynamics. Management Science 52, 12731287.
Danielsson, J., de Vries, C., 2000. Value-at-risk and extreme returns. Annales d
Economie et de Statistique 60, 239270.
Danielsson, J., Hartmann, P., de Vries, C., 1998. The cost of conservatism. Risk 11,
101103.
Darbha, G., 2001. Value-at-Risk for Fixed Income Portfolios A Comparison of Alter-
native Models. National Stock Exchange, Mumbai, India.
Diebold, F.X., Mariano, R.S., 1995. Comparingpredictive accuracy. Journal of Business
& Economic Statistics 13, 253263.
Down, K., 2002. Measuring Market Risk. John Wiley & Sons, Chichester.
Dufe, D., Pan, J., 1997. An overviewof value at risk. Journal of Derivates 4 ((Spring)
3), 749.
Embrechts, P., Resnick, S., Samorodnitsky, G., 1999. Extreme value theory as a risk
management tool. North American Actuarial Journal 26, 3041.
Embrechts, P., McNeil, A.J., Straumann, D., 2002. Correlation and dependence in
risk management: properties and pitfalls. In: Dempster, M. (Ed.), Risk Man-
agement: Value at Risk and Beyond. Cambridge University Press, Cambridge,
pp. 176223.
Engle, R., 1982. Autoregressive conditional heteroskedasticity with estimates of the
variance of UK ination. Econometrica 50, 9871008.
Engle, R., Bollerslev, T., 1986. Modeling the persistence of conditional variances.
Econometric Reviews 5, 150.
Engle, R., Manganelli, S., 2004. CAViaR: conditional autoregressive value at risk by
regression quantiles. Journal of Business & Economic Statistics 22, 367381.
Ergun, A., Jun, J., 2010. Time-varying higher-order conditional moments and fore-
casting intraday VaR and expected shortfall. Quarterly Reviewof Economics and
Finance 50, 264272.
Escanciano, J.C., Olmo, J., 2010. Backtesting parametric value-at-risk withestimation
risk. Journal of Business & Economic Statistics 28, 3651.
Escanciano, J.C., Pei, P., 2012. Pitfalls inbacktesting historical simulationVaRmodels.
Journal of Banking & Finance 36, 22332244.
Estrella, A., Hendricks, D., Kambhu, J., Shin, S., Walter, S., 1994. The price risk of
options positions: measurement andcapital requirements. Federal ReserveBank
of NewYork. Quarterly Review, 2743.
Feria Domnguez, J.M., 2005. El riesgo de mercado: su medicin y control. Ed. Delta.
Fleming, J., Kirby, C., 2003. Acloser look at the relation between GARCHand stochas-
tic autoregressive volatility. Journal of Financial Econometrics 1, 365419.
Genc ay, R., Selc uk, F., 2004. Extreme value theory and value-at-risk: relative perfor-
mance in emerging markets. International Journal of Forecasting 20, 287303.
Genc ay, R., Selc uk, F., Uluglyagci, A., 2003. High volatility, thick tails and extreme
value theory in value-at-risk estimation. Insurance: Mathematics and Eco-
nomics 33, 337356.
Gento, P., 2001. Comparacin entre Mtodos Alternativos para la Estimacin del
Valor en Riesgo. Documentos de trabajo. Doc. 1/2001/1. Facultad de Ciencias
Econmicas y Empresariales de Albacete, Universidad de Castilla-La Mancha.
Gerlach, R., Chen, C., Chan, N., 2011. Bayesian time-varying quantile forecasting for
value-at-risk in nancial markets. Journal of Business & Economic Statistics 29,
481492.
Giamouridis, D., Ntoula, I., 2009. A comparison of alternative approaches for deter-
mining the downside risk of hedge fund strategies. Journal of Futures Markets
29, 244269.
Giannopoulos, K., Tunaru, R., 2005. Coherent risk measures under ltered historical
simulation. Journal of Banking and Finance 29, 979996.
Giot, P., Laurent, S., 2004. Modelling daily value-at-risk using realized volatility and
ARCH type models. Journal of Empirical Finance 11, 379398.
Gonzlez-Rivera, G., Lee, T., Mishra, S., 2004. Forecasting volatility: a reality check
basedonoptionpricing, utility function, value-at-risk, andpredictive likelihood.
International Journal of Forecasting 20, 629645.
Guermat, C., Harris, R., 2002. Forecasting value-at-risk allowing for time variation
in the variance and kurtosis of portfolio returns. International Journal of Fore-
casting 18, 409419.
Haas, M., 2009. Modelling skewness and kurtosis with the skewed Gauss-Laplace
sumdistribution. Applied Economics Letters 16, 12771283.
Haas, M., Mittnik, S., Paolella, M., 2004. Mixednormal conditional heteroskedasticity.
Journal of Financial Econometrics 2, 211250.
Hansen, B., 1994. Autoregressive conditional density estimation. International Eco-
nomic Review35, 705730.
Harvey, A., 1998. Long memory in stochastic volatility. In: Kight, J., Satchell, E. (Eds.),
Forescasting Volatility in Financial Markets. Butterworth-Haineman, London,
pp. 307320.
Harvey, A., Shephard, N., 1996. Estimation of an asymmetric stochastic volatility
model for asset returns. Journal of Business andEconomic Statistics 14, 429434.
Harvey, C., Siddique, A., 1999. Autoregressiveconditional skewness. Journal of Finan-
cial and Quantitative Analysis 34, 465487.
Hendricks, D., 1996. Evaluation of value-at-risk models using historical data. Federal
Reserve Bank of NewYork Economic Police Review2, 3970.
Huang, A., 2009. A value-at-risk approach with kernel estimator. Applied Financial
Economics, 379395.
Huang, Y., Lin, B., 2004. Value-at-risk analysis for Taiwan stock index futures: fat
tails and conditional asymmetries in return innovations. Reviewof Quantitative
Finance and Accounting 22, 7995.
Hull, J., White, A., 1998. Incorporating volatility updating into the historical simula-
tion method for value-at-risk. Journal of Risk 1, 519.
Jarrow, R., Rudd, A., 1982. Approximate option valuation for arbitrary stochastic
processes. Journal of Financial Economics 3, 347369.
Johnson, N.L., 1949. Systems of frequency curves generated by methods of transla-
tions. Biometrika 36, 149176.
Jondeau, E., Rockinger, N., 2001. Conditional Dependency of Financial Series: an
Application of Copulas. Working Paper, Banque de France, NER#82.
Jondeau, E., Rockinger, N., 2003. Conditional Volatility, skewness andkurtosis: exist-
ence, persistence, and comovements. Journal of Economic Dynamics & Control
27, 16991737.
Jorion, P., 1990. The exchange rate exposure of U.S. multinationals. Journal of Busi-
ness 63, 331345.
Jorion, P., 1997. Value at Risk: The New Benchmark for Controlling Market Risk.
Irwin, Chicago, IL.
Jorion, P., 2001. Value at Risk: The New Benchmark for Managing Financial Risk.
McGraw-Hill.
Kerkhof, J., Melenberg, B., 2004. Backtesting for risk-basedregulatorycapital. Journal
of Banking & Finance 28, 18451865.
Koenker, R., Basset, G., 1978. Regression quantiles. Econometrica 46, 3350.
Koopman, S., Jungbacker, B., Hol, E., 2005. Forecasting daily variability of the S&P
100 stock index using historical, realized and implied volatility measurements.
Kuester, K., Mittnik, S., Paolella, M., 2006. Value-at-risk prediction: a comparison of
alternative strategies. Journal of Financial Econometrics 4, 5389.
Kupiec, P., 1995. Techniques for verifying the accuracy of risk measurement models.
Journal of Derivatives 2, 7384.
Lehar, A., Scheicher, M., Schittenkopf, C., 2002. GARCH vs. stochastic volatil-
ity: option pricing and risk management. Journal of Banking & Finance 26,
323345.
Leon, A., Gonzalo, R., Serna, G., 2005. Autoregressive conditional volatility,
skewness and kurtosis. Quarterly Review of Economics and Finance 45,
599618.
Li, C.W., Li, W.K., 1996. On a double-threshold autoregressive heteroscedastic time
series model. Journal of Applied Econometrics 11, 253274.
Li, M., Lin, H., 2004. Estimating value-at-risk via Markov switching ARCHmodels: an
empirical study on stock index returns. Applied Economics Letters 11, 679691.
Linsmeier, T.J., Pearson, N.D., 2000. Value at risk. Financial Analysts Journal 56,
4767.
Lopez, J.A., 1998. Testing your risk tests. Financial Survey, 1820.
Lopez, J.A., 1999. Methods for evaluating value-at-risk estimates. Federal Reserve
Bank of San Francisco Economic Review2, 317.
Marimoutou, V., Raggad, B., Trabelsi, A., 2009. Extreme value theory and value at
risk: application to oil market. Energy Economics 31, 519530.
McAleer, M., Jimenez, J.A., Prez, A., 2011. International Evidence on GFC-robust
Forecast for Risk Management under the Basel Accord. Econometric Institute
Report EI 2011-04. Erasmus University Rotterdam, Econometric Institute.
McAleer, M., Jimnez-Martin, J., Prez-Amaral, T., 2010a. A decision rule to mini-
mize daily capital charges in forecasting value-at-risk. Journal of Forecasting 29,
617634.
McAleer, M., Jimnez-Martin, J., Prez-Amaral, T., 2010b. What happened to risk
management during the 200809 nancial crisis? In: Kolb, R.W. (Ed.), Lessons
from the Financial Crisis: Causes, Consequences and Our Economic Future.
Wiley, NewYork, pp. 307316.
McDonald, J., Newey, W., 1988. Partially adaptive estimation of regression models
via the generalized t distribution. Econometric Theory 4, 428457.
McDonald, J., Xu, Y., 1995. Ageneralizationof the beta distributionwithapplications.
McNeil, A., 1998. Calculating Quantile Risk Measures for Financial Time Series Using
Extreme Value Theory. Department of Mathematics, ETS. Swiss Federal Techni-
cal University E-Collection, http://e-collection.ethbib.etchz.ch/
McNeil, A., Frey, R., 2000. Estimationof tail-relatedriskmeasures for heteroscedastic
nancial time series: an extreme value approach. Journal of Empirical Finance,
271300.
Merton, R.C., 1980. Onestimating the expectedreturnonthe market: anexploratory
investigation. Journal of Financial Economics 8, 323361.
Mittnik, S., Paolella, M., 2000. Conditional density and value-at-risk prediction of
Asian currency exchange rates. Journal of Forecasting 19, 313333.
Morgan, J.P., 1996. Riskmetrics Technical Document, 4th ed. J.P. Morgan, NewYork.
Nelson, D., 1991. Conditional heteroskedasticity in asset returns: a new approach.
Econometrica 59, 347370.
Nieto, M.R., Ruiz, E., 2008. Measuring Financial Risk: Comparison of Alternative Pro-
cedures to Estimate VaR and ES. W.P. 08-73. Statistics and Econometrics Series
26.
Nozari, M., Raei, S., Jahanguin, P., Bahramgiri, M., 2010. Acomparisonof heavy-tailed
estimates and ltered historical simulation: evidence from emerging markets.
International Reviewof Business Papers 64, 347359.
Nguez, T., 2008. Volatility and VaR forecasting in the madrid stock exchange. Span-
ish Economic Review10, 169196.
Ozun, A., Cifter, A., Yilmazer, S., 2010. Filteredextreme-value theory for value-at-risk
estimation: evidence fromTurkey. Journal of Risk Finance 11, 164179.
Pagan, A., Schwert, G., 1990. Alternative models for conditional stock volatility.
Patton, A., Sheppard, K., 2009. Evaluating volatility and correlation forecasting. In:
Andersen, T.G., Davis, R.A., Kreiss, J.-P., Mikosch, T. (Eds.), The Handbook of
Financial Time Series. Springer Verlag.
Polanski, A., Stoja, E., 2009. Dynamics Density Forecasts for Multivariate
Asset Returns. Department of Economics. University of Bristol, Discussion
Paper no 09/616, available at SSRN: http://ssrn.com/abstract=1655767 or
http://dx.doi.org/10.2139/ssrn.1655767
Polanski, A., Stoja, E., 2010. Incorporating higher moments into value-at-risk fore-
casting. Journal of Forecasting 29, 523535.
Pong, S., Shakelton, M., Taylor, S., Xu, X., 2004. Forecasting currency volatility: a
comparison of implied volatilities and AR(FI)MA models. Journal of Banking &
Finance 28, 25412563.
Pritsker, M., 1997. Evaluating value at risk methodologies: accuracy versus compu-
tational time. Journal of Financial Services Research 12, 201242.
Pritsker, M., 2001. The Hidden Dangers of Historical Simulation. Board of Governors
the Federal Reserve System (U.S.), Finance and Economics Discussion Series:
20012027.
Ren, F., Giles, D., 2007. Extreme Value Analysis of Daily Canadian Crude Oil. Econo-
metrics Working Paper EWP0708, ISSN 1485-6441.
Rosenberg, J.V., Schuermann, T., 2006. Ageneral approachtointegratedriskmanage-
ment with skewed, fat-tailed risks. Journal of Financial Economics 79, 569614.
Rudemo, M., 1982. Empirical choice of histograms and kernel density estimators.
Scandinavian Journal of Statistics 9, 6578.
Sajjad, R., Coakley, J., Nankervis, J., 2008. Markov-switching GARCH modelling of
value-at-risk. Studies in Nonlinear Dynamics & Econometrics 12, Art. 7.
Sener, E., Baronyan, S., Menguturk, L.A., 2012. Ranking the predictive performances
of value-at-risk estimation methods. International Journal of Forecasting 28,
849873.
Sheather, S., Marron, J., 1990. Kernel quantile estimator. Journal of American Statis-
tical Association 85, 410416.
Silva, A., Melo, B., 2003. Value-at-risk and extreme returns in Asian stock markets.
International Journal of Business 8, 1740.
Silverman, B., 1986. Density Estimation for Statistics and Data Analysis. Chapman
and Hall, London.
Srinivasan, A., Shah, A., 2001. Improved Techniques for Using Monte Carlo in VaR
estimation. National Stock Exchange Research Initiative, Working Paper.
So, M., Yu, P., 2006. Empirical analysis of GARCH models in value at risk esti-
mation. Journal of International Financial Markets, Institutions & Money 16,
180197.
So, M., Li, W., Lam, K., 2002. On a threshold stochastic volatility model. Journal of
Forecasting 22, 473500.
Taylor, S., 1982. Financial returns modeledbytheproduct of two stochastic processes
a study of daily sugar prices. In: Anderson, O.D. (Ed.), Time Series Analysis:
Theory and Practice 1. North Holland, Amsterdam, pp. 203226.
Taylor, S., 1986. Modeling Financial Time Series. John Wiley and Sons, Chichester,
UK.
Taylor, S., 1994. Modelling stochastic volatility: a review and comparative study.
Mathematical Finance 4, 653667.
Taylor, S., Xu, X., 1997. The Incremental volatility information in one million foreign
exchange quotations. Journal of Empirical Finance, 317340.
Theodossiou, P., 1998. Financial data and skewed generalized t distribution. Man-
agement Science 44, 16501661.
Theodossiou, P., 2001. Skewness and Kurtosis in Financial Data and the Pricing of
Options. Working Paper. Rutgers University.
Tolikas, K., Koulakiotis, A., Brown, R., 2007. Extreme risk and value-at-risk in the
German stock market. European Journal of Finance 13, 373395.
Trenca, I., 2009. The use in banks of VaR method in market risk management. Scien-
tic annals of the Alexandru Ioan Cuza, University of Iasi. Economic Sciences
Series 2009, 186196.
Velayoudoum, M., Raggad, B., Trabelsi, A., 2009. Extreme value theory and value at
risk: application to oil market. Energy Economics 31, 519530.
Ward, J.D., Lee, C.L., 2002. Areviewof problem-based learning. Journal of Family and
Consumer Sciences Education 20, 1626.
White, H., 2000. A reality check for data snooping. Econometrica 68, 10971126.
Wong, W.K., 2008. Backtesting trading risk of commercial banks using expected
shortfall. Journal of Banking & Finance 32, 14041415.
Wong, W.K., 2010. Backtesting value-at-risk basedontail losses. Journal of Empirical
Finance 17, 526538.
Wong, C., Li, W., 2001. On a mixture autoregressive conditional heteroscedastic
model. Journal of the American Statistical Association 96, 982995.
Xu, D., Wirjanto, T., 2010. An empirical characteristic function approach to VaR
under a mixture-of-normal distribution with time-varying volatility. Journal of
Derivates 18, 3958.
Yu, P.L.H., Li, W.K., Jin, S., 2010. On some models for value-at-risk. Econometric
Reviews 29, 622641.
Ze-To, S.Y.M., 2008. Value at risk and conditional extreme value theory via Markov
regime switching models. Journal of Futures Markets 28, 155181.
Zenti, R., Pallotta, M., 2001. Risk Analysis for Asset Manager: Historical Scenarios
Based Methods and the Bootstrap Approach. Mineo. RAS Asset Management,
Milan, Italia.
Zhang, M.H., Cheng, Q.S., 2005. AnapproachtoVaRfor capital markets withGaussian
mixture. Applied Mathematics and Computations 168, 10791085.
Zikovic, S., Aktan, B., 2009. Global nancial crisis and VaR performance in emerg-
ing markets: a case of EU candidate states Turkey and Croatia. Proceedings
of Rijeka faculty of economics. Journal of Economics and Business 27,
149170.

B6 A Comprehensive Review of Value at Risk Methodologies

Uploaded by

Copyright:

Available Formats

B6 A Comprehensive Review of Value at Risk Methodologies

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

B6 A Comprehensive Review of Value at Risk Methodologies

Uploaded by

Copyright:

Available Formats

The Spanish Reviewof Financial Economics 12 (2014) 1532

The Spanish Review of Financial Economics

(q) zt = (rt t )]ot

You might also like