eb 022959
TRANSCRIPT
-
8/18/2019 Eb 022959
1/16
The Journal of Risk FinanceThe Effect of Asymmetries on Stock Index Return Value-at-Risk Estimates
CHRIS BROOKS GITA PERSAND
Ar ticle information:To cite this document:CHRIS BROOKS GITA PERSAND, (2003),"The Effect of Asymmetries on Stock Index Return Value-at-Risk Estimates", TheJournal of Risk Finance, Vol. 4 Iss 2 pp. 29 - 42Permanent link to this document:http://dx.doi.org/10.1108/eb022959
Downloaded on: 05 October 2015, At: 22:06 (PT)
References: this document contains references to 0 other documents.
To copy this document: [email protected]
The fulltext of this document has been downloaded 337 times since 2006*
Users who downloaded this article also downloaded:
Chu-Hsiung Lin, Shan-Shan Shen, (2006),"Can the student-t distribution provide accurate value at risk?", The Journal of Risk
Finance, Vol. 7 Iss 3 pp. 292-300 http://dx.doi.org/10.1108/15265940610664960
Stavros Degiannakis, Christos Floros, Alexandra Livada, (2012),"Evaluating value-at-risk models before andafter the financial crisis of 2008: International evidence", Managerial Finance, Vol. 38 Iss 4 pp. 436-452 http://dx.doi.org/10.1108/03074351211207563
CHRIS BROOKS, ANDREW D. CLARE, GITA PERSAND, (2002),"An Extreme Value Theory Approach to CalculatingMinimum Capital Risk Requirements", The Journal of Risk Finance, Vol. 3 Iss 2 pp. 22-33 http://dx.doi.org/10.1108/eb043485
Access to this document was granted through an Emerald subscription provided by emerald-srm:543666 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our Emerald for Authors service
information about how to choose which publication to write for and submission guidelines are available for all. Please visitwww.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society. The company manages a portfolio of more than 290 journals and over 2,350 books and book series volumes, as well as providing an extensive range of onlineproducts and additional customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the Committee on PublicationEthics (COPE) and also works with Portico and the LOCKSS initiative for digital archive preservation.
*Related content and download information correct at time of download.
http://dx.doi.org/10.1108/eb022959
-
8/18/2019 Eb 022959
2/16
The Effect of Asymmetries
on Stock Index Return
Value-at-Risk Estimates
C H R I S B R O O K S A N D G I T A P E R S A N D
T
here is widespread ag reem ent in the
relevant literature that equity return
volatility rises more following neg
ative than positive shocks. Pagan and
Schwert [1990], Nelson [1991], Campb ell and
Hentschel [1992] , Engle and N g [1993] ,
Glosten e t al. [1993], and H en ry [1998], for
example, all demonstrate the existence of asym
me tric effects in stock index return s. The re are
broadly two potential explanations for such
asymmetry in variance that have been suggested
in this litera ture . T he first is the levera ge
effect, usually associated w ith Black [1976] and
Chris tie [1982], which posits that if equity
values fall, the firm's debt-to-equity ratio will
rise,
thus inducing equityholders to perceive
the stream of future income accruing to their
positions as being relatively more risky than
previously. A second possible explanation for
observed asymm etries is term ed the volatility
feedback hypothesis. Assuming constant cash
flows, if expected returns increase when stock
price volati l i ty increases, then s tock prices
should fall when volatility rises.
An entirely separate l ine of academic
enquiry has centered around the determina
tion of wh at is term ed an insti tutions value
at risk (VaR). Va R is a calculation of the likely
losses that might occur from changes in the
mark et prices of a particular securities or po rt
fo l io pos i t ion . The minimum capi ta l r i sk
requirement (M C R R ) or position r isk require
me nt (P RR ) is then def ined as the minim um
amount of capital required to absorb all but a
prespecif ied propo r t ion of expec ted fu ture
losses. Dim son and M arsh [1995, 1997] argue
that portfolio-based approaches to determining
P R R s are mo re eff icient than a l ternat ive
approaches, since the former allow fully and
directly for the risk-reduction benefits from
having a diversified book.
Th e nu mb er of s tudies in existence that
seek to determine appropr ia te methods for
calcu la t ing and evaluat ing value-a t - r isk
methodologies has increased substantially in
the past five
years.
Jackson et al. [1998] assess
the empirical performan ce o f various mode ls
for value at risk using historical returns from
the actual portfolio of a large investment ban k.
Alexander and Leigh [1997] offer an analysis
of the re la t ive per fo rma nce of equal ly
we igh ted , exponen t i a l ly we igh ted mov ing
ave rage (EWMA) , and GAR C H mode l f o r e
casts of volatility, evaluated using traditional
s tatis t ical and operational adequacy criteria.
Th e G A R C H mod el is found to be preferable
to E W M A in t e rms o f min im iz ing the
number of exceedances in a backtest, although
the s imple unw eigh ted average is supe rior to
both. The issue of sample length is discussed
in Hoppe [1998], who argues that, for all asset
classes and holding periods tested, the use of
(unweighted) shorter samples of data yields
more accurate VaRs than longer runs. Kupiec
[1995] , on the other hand, argues that long
samples are always required in order to eval
uate the effectiveness of those estimates using
exception tests . Berkowitz [2001] also exam-
CHRIS
BROOKS
is professor of finance at the
ISMA
Centre,
University
of
Reading,
in the UK.
GITA PERSAND
is a
lecturer
in
finance
at the
Department
o f
Economics,
University
of Bristol, in the UK .
W I N T E R 2003
THE J OUR NAL OF RISK FINANCE 2 9
mailto:[email protected]:[email protected]:[email protected]:[email protected]
-
8/18/2019 Eb 022959
3/16
ines the frequency and the ma gnitu de of exc ept ion s
(days w he n th e VaR is insufficient to cover actual trad ing
losses), although it is in the spirit of VaR modeling as
per the Bas le recomme ndat ions that on ly the num ber of
exceptions is considered and not their sizes. Other recent
studies that compare different models for computing VaR
include those o f Brooks and Persand [2000a, 2000b] ,
Brooks et al . [2000], Longin [2000], Vlaar [2000], and
Lopez and Walter [2001]. Mu ch research has relied upo n
an assumption of normally dis tr ibuted returns, while in
practice almost all asset return distributions are fat tailed.
Two studies that have proposed methods to deal with
leptokurtosis are Huisman et al . [1998] and Hull and
White [1998] .
This article extends recent research and considers th e
effect of any asymmetries that may be present in the data
on the evaluation and accuracy of value-at-risk estimates.
A number of recent s tudies have examined the use of
extreme value dis tr ibutions for computing value-at-r isk
estimates. Such models can explicitly allow for lepto kurtic
return distributions, but can also account for asymmetries
in the unconditional distributions by fitting separate models
for the upper and lower tails. McNeil and Frey [2000], for
example, propose a new method for estimating VaR based
on a combinat ion of G A R C H m odel ing wi th ex treme
value dis tr ibutions. Other s tudies employing EVT have
included Longin [2000], Neftci [2000], and Brooks et al.
[2002].
However, since the extreme value theory (EVT)
approach to com puting position risk requirements has been
considered elsewhere, we do not discuss it further here.
Instead, we examine a number of alternative models that
may be used to capture asymmetries in asset return distri
butions. Our analysis is conducted in the context of the
stock markets of five Southeast Asian economies, and the
S&P 500 index, which is employed as a benchmark. We
compare the performance of various models, some of which
do, and some of which do not, allow for asymmetries. The
models are evaluated in the context of the Basle Com
mittee rules , and their proposed method for determining
whether models for the calculation of VaR are adequate
using a holdout sample. To anticipate our main finding,
we conclude that allowing for asymmetries can lead to
improved V aR estimates, and that a simple asymmetric risk
measure proves to be the most stable and reliable method
for calculating VaR.
The remainder of this article is organized into six
sections. Section 2 presents and describes the data em ployed,
while the various symmetric and asymmetric volati l i ty
models uti l ized are displayed in section 3. Section 4
describes the methodology employed for the calculation of
value at risk, while the value-at-risk estimates from the
various models are presented and evaluated in section 5.
Finally, section 6 offers some concluding remarks.
D A T A
Th e analysis unde rtaken in this article is based on daily
closing prices of five Southeast Asian stock market indices:
the Ha ng Seng Price Index, N ikkei 225 Stock Average Price
Index, Singapore Straits Times Price Index, South Korea
SE Composite Price Index, and Bangkok Book Club Price
Index. We also employ the U.S. S&P 500 composite index
E X H I B I T 1
Sum mary Statist ics: Stock Inde x Returns for the Period January 1, 1985-April 29, 1999
(3,737 Observations)
Ho ng Ko ng
Japan
Sing a po re
South Korea
Thai land
Portfolio
S P 500
M ea n
0.00077
0.00019
0.00037
0.00056
0.00042
0.00046
0.00024
Variance
0.00030
0.00018
0.00020
0.00027
0.00030
0.00009
0.00010
S k e w n e s s
- 2 . 2 7 6 1 3 *
0.10919*
-1 .18975*
0.42596*
0.15243*
- 0 . 6 2 7 9 2 *
- 3 . 6 2 4 9 3 *
Kurtosis
51 .96710*
10.03701*
44 .72603*
5.55209*
9.85702*
14.59893*
80.59497*
Normali ty
Test**
422368*
15643*
311360*
4897*
15095*
33324*
1019595*
* denotes significance
at the 1% level.
** Bera-Jarque Normality Test.
Th e
portfolio
is an equally
weighted combination
of the five Southeast Asian market returns.
3 0 TH E EFFECT OF ASYMMETRIES ON STOCK INDEX RET UR N VALU E-AT-RISK ESTIMATES
WINTER 2003
-
8/18/2019 Eb 022959
4/16
returns as a bench mark for comparison. T he data, obtained
from Primark Datastream, run from January 1, 1985, to
April 29, 1999, giving a total of 3,737 observations. All sub
sequent analysis is performed on the daily log returns, with
the sum ma ry statistics being given in E xhib it 1. All six returns
series exhibit the standard property of asset return data that
they have fat-tailed distribution s as indicated by the sig
nificant coefficient of excess kurtosis. These characteristics
are also shown by the highly significant Jarque-Bera nor
mality test statistics. All series are also are, eit her significantly
skewed to the left (U.S., Hong Kong, and Singapore) or to
the right (Japan, South Korea, and Thailand). Unconditional
skewness is an arguably important but neglected feature of
many asset return series (see, for example, Harvey and Sid-
dique [1999]), which if ignored could lead to misspecified
risk manage men t models. An illustration of where skewness
may arise is given by the view that equity price movem ents
go up the stairs and dow n in the elevator, imply ing that
upward movements are smaller but more frequent than
downward movements,
ceteris
paribus. T his being the case, a
symmetrical measure such as variance will no longer b e an
appropriate measure of the total risk of an asset.
We also form, and perform subsequent analysis on,
an equally weighte d portfolio consisting of the five So uth
east Asian market indices listed above. The advantage of
diversification is clearly recognized in this case: the vari
ance of the portfolio is only around one-half of the vari
ance of even the least volatile of the individual com pon ent
series—see the penultimate colum n o f Exhibit 1.
On
the
other hand, the distr ibution of portfolio returns is st i l l
skewed (to the left) and is leptokurtic . T he skewness in all
five series of returns data, and in the portfolio returns (albeit
in different directions), is one manifestation of an asym
metry— in othe r words, this is prima acie evidence that the
unconditional distribution of returns is not symmetric. We
shall retu rn to this feature of the data in the following sec
tion in the context of value-at-risk estimation.
S Y M M E T R I C A N D A S Y M M E T R I C
V O L A T IL I TY M O D E L S
There exist a number of different methods for deter
mining an insti tution 's value at r isk. The most popular
me tho ds can be usefully classified as be ing eit her p ara
metric or non-parametric. In the former category comes
the volatilities and correlations app roach popularize d by
J.P. M orga n [1996]; this me tho d involves the estimation of
a volatility parameter, and, conditional upon an assump
tion of normality, the volatility estimate is multiplied by the
appropriate critical value from the no rma l distribution and
by the value of the asset or portfolio, to o btain an estimate
of the VaR in money terms. There are broadly two ways
that VaR could be ca lcu la ted under the parametr ic
umb rella. First , one could estimate a mo del for ret urn
volatility, and, conditioned upon this, employ the appro
priate critical value from a relevant distribution. Second,
one could, rather than estimating a conditional parametric
model, employ the unconditional distr ibution of returns,
fitting one of many other available parametric distribu
tions. Press [1967] and Kon [1984], for example, both sug
gest the use of a mixture of norma l distributions to explain
the observed kurtosis and skewness in the distribution of
stock returns. Madan and Seneta [1990], on the o ther hand,
propose an entirely new class of models, known as variance-
gamma processes, as a model for stock prices.
Non-parametr ic approaches to the ca lcu la t ion of
value at risk can take many forms, bu t the simplest is derived
from an estimate of the unconditional density of the sample
of returns. So, for example, the VaR expressed as a pro
por tion of the initial value of the asset or portfolio, wh ich
is required to cover 99% of expected losses, would simply
be given by the absolute value of the first percen tile of th e
return distribution. Jorio n [1995] shows that the param etric
approach to VaR can be preferable, even in situations where
the returns are not normally distributed. In any case, non-
parametric approaches are beyond the scope of this article
and are thus not considered further (but see, for example,
Jorion [1996] or Dowd [1998] for extended descriptions
of the various methods available under the non-parametric
umbrella). Under the parametric approach, the normal dis
tribution is employed almost universally, and the VaR (in
money terms) for model i is given by
wh er e α ( N
5%
) is the relevant value from the standard
normal tables, a. is the volatility estimate, and V is the
value of the portfolio. VaR is also commonly expressed
as a proportion of the asset or portfolio value, and this
convention will also be adopted in this study. Once a view
is taken that a parametric approach will be em ployed for
the calculation of VaR, the issue simply boils down to
estimation of the volatility parameter that describes the
asset or portfolio;
1
we now present a variety of models
which can be used for estimating and modeling σ
i
.
WINTER
2003
THE JOURNAL OF RISK FINANCE 3 1
-
8/18/2019 Eb 022959
5/16
T h e Un co nd i t io na l Va r ia nce
a n d t h e E W M A M o d e l
An estimate of the unconditional variance provides
the simplest method for forecasting future volatility, also
known as an equally weighted moving average model, as
forecasts are constructed by simply calculating the standard
deviation from the most recent three years of historical
data. This estimated historical standard deviation is then
the prediction of the standard deviation over the evalua
tion period. The sample period is then rolled forward by
60 observations (one trading quarter), σ
i
re-estimated, and
so on. To ensure consistency and a fair comparison, we
also use the same framework (a three-year rolling sample
updated every 60 observations) for the estimation of the
parameters for all other approaches employed.
We similar ly employ the exponent ia l ly weighted
mo ving average (EW M A) m ode l , popu lar ized by J .P.
Morgan, and which has been found to produce accurate
volatil i ty forecasts. U nd er the E W M A specif ication, t he
fit ted variance from the model, which becomes the fore
cast for multi-step ahead forecasts, is an expo nentially
declining function of previous squared values. The decay
factor, λ , is set to 0.94 following the
J.P.
Mo r g an r ec
ommendation and previous research in this area.
T he GARCH( 1 ,1 ) M o de l
Recent research (see Alexander and Leigh [1997],
for example) has suggested tha t GARCH-type models
may be preferable for modeling volatility in a risk man
agement context, and hence our conditional model anal
y sis co m men ces w i th th e p la in v ani ll a GA R C H( 1 ,1 )
model, given as follows:
wh er e x
t
= Log(P
t
/P
t-1
) wi th P
t
being the value of the
stock index at time t and ε
t
~ N(0, h
t
) .
2 , 3
The volatility
forecasts a re const ruc ted by i te ra t ing the condi t ional
expectations operator in the usual fashion.
To determine whether an asymmetr ic condi t ional
volatility model is necessary and to ensure that the model
does not und erpre dict or overpredict volatil i ty in periods
where there are large innovations in returns, the stan
dardized residuals from the GARCH(1,1) specif ication
are examined for sign and size bias, which is carried out
using the tests of Engle a nd N g [1993]. T he test for asym
metry in return volatility is given as follows:
E X H I B I T 2
Engle and
N g
[1993] Test for the GAR CH(1,1) Mo del: Stock Index Returns
for the P eriod Janu ary 1, 1985-Ap ril 29, 1999
(3,737 Observations)
Hong Kong
Japan
Sing a po re
South Korea
Thailand
Portfolio
S P 500
Φ
0
0.913
(0.138)**
0 937
(0.112)**
0 705
(0 .145)**
0 979
(0 .075)**
0 846
(0.113)**
1.085
(0.110)**
0 960
(0.115)**
Φ
1
0.189
(0.176)
-0.042
(0.148)
0.159
(0.188)
0 082
(0.100)
0 322
(0.108)**
-0.168
(0.151)
0 037
(0.151)
Φ
2
-0.128
(0.103)
-0.304
(0.095)**
-0.429
(0.119)**
0.061
(0.072)
0 085
(0.100)
-0.173
(0.101)
-0.222
(0.093)*
Φ
3
-0.088
(0.142)
-0.151
(0.117)
0.149
(0.146)
-0.037
(0.070)
-0.018
(0.108)
-0.179
(0.115)
-0.178
(0.122)
χ
2
3)
10.141*
19.723**
20.342**
1.991
10.222*
10.992*
17.590**
Standard
errors
are in parentheses; * and
**
indic te signific nce
at the 5% and 1%
levels respectively.
Th e
portfolio
is an equally
weighted combin tion
of
the
five Southeast Asian market returns.
3 2 TH E EFFECT OF ASYMMETRIES ON STOCK INDEX RE TU RN VALU E-AT-RISK ESTIMATES
W I N T E R
2003
-
8/18/2019 Eb 022959
6/16
wh er e
ω
t
is a
wh i t e n o i se d i s tu r b an ce t e r m .
Z
t-1
is
defined as an indicator dummy that takes the value 1 if
ε
t
/√ h
t
-
8/18/2019 Eb 022959
7/16
wh er e C
0
, A
1
, and B
1
are parameter matr ices to be esti
mated , ε
t-1
is a vector of lagged errors, and
The BEKK parameterization requires estimation of
only 55 free parameters in the conditional v ariance-cov ari-
ance s t ruc ture (compared wi th 255 in the comple te ly
unrestr icted
vec
model) , and guarantees H
t
positive defi
nite. Simple modifications to (6) can be made to allow for
asymmetr ies under the GJR and EG A R CH formula tions.
The modif ied model inc lud ing asymmetry te rms in a
quadratic form could be written
where ξ
j,t
= min{ε
t
, 0} and D
1
is a 5
× 5 parameter matrix
including all of the asymmetry coefficients. In all cases,
the estimated asymmetry terms are significant at the 5%
level or better.
T h e S e m i - V a r i a n c e
Modern Portfolio Theory (MPT), which underlies
many theoretical models in f inance, carries with i t the
explicit assu mp tion that asset retu rns are joi ntl y elliptically
dis tr ibuted. Such an assumption m ay be considered u nd e
sirable since it rules ou t the possibility of asym metric retu rn
distributions. Risk in the M P T framew ork is measured in
term s of surpr ises rath er than in term s of losses or failure
to achieve an expected or benchmark re turn . Such a
description of risk does not tie in well with m ost investors'
notions of what constitutes a risk.
Another approach to estimating a volati l i ty param
eter, which may be used in the present context, and which
has been given broad theoretical consideration in an asset
allocation framew ork, is the co nce pt of the lower partial
momen t (LPM)— s ee , f o r example , C hoob ineh and
Bran t ing [1986] , Ha r low [1991] , and Grootveld and
Hallerbach [1999]. T he lower partial m om en t of order a
around
τ is defined as
wh ere F(X) is the cum ulative distr ibution function of the
return X. Setting α = 2, and τ = (I in Equation (8) above
defines the semi-variance of a random variable X with
mean μ .
Fo l lo win g J .P . Mo r g an R i sk M et r i c s an d man y
empirical studies, assuming
4
that μ = 0, an asym ptotically
unbiased and strongly consistent estimator of the semi-
varian ce for a sam ple of size T is given by (see Jos eph y and
Aczel [1993])
5
whe re T denotes the num ber of te rms for which x
t
< 0,
and the scaled semi-standard deviation which we employ
for the com pu tati on of value at risk is given by th e square
root of (9).
The square root of 9) can thus replace directly the
usual standard deviation in (1). The scaled semi-standard
deviation ( the square root of (9)) can account for the
uncondi t ional skewness in the d is t r ibu t ion of re turns .
Expressed in this way, we can still employ the standard
normal critical value since we now implicitly assume that
x
t
< 0 follows a half-n orm al d istribu tion. If desired, the
formula in (9) could be trivially modified to estimate the
upper scaled semi-standard deviation, which would be of
interest for calculating the VaR of a short posit ion.
M E T H O D O L O G I E S F O R C A L C U L A T I N G
A N D E V A L U A T I N G T H E V A L UE AT R I S K
Th e regulatory environment under the Basle Co m
mittee Rules requires that VaR should be calculated as
the higher of 1) the firm's previous day's value at risk m ea
sured according to the parameters given below and 2) an
average of the daily VaR measures on each of the pre
ceding 60 business days, with the latter subjected to a
multiplication factor. Value at risk is to be computed on
a daily basis over a mi nim um holding period of 10 days.
However, shorter holding periods can be used, but they
have to be scaled up to 10 days, and in general this is
achieved using the square root of t ime rule. Moreover,
VaR has to be estimated at the 99% probability level, using
daily data over a minimum length of one year (250 trading
days) , with the estimates being updated at least every
3 4 TH E EFFECT OF ASYMMETRIES ON STOCK INDEX RET UR N VALU E-AT-RISK ESTIMATES
WINTER 2003
-
8/18/2019 Eb 022959
8/16
quarter. T he rules do leave the ban k a broad degree of flex
ibility in ho w the VaR is actually calculated. Fo r exam ple,
the M C R R es timates can be updated more f requent ly
than quarterly, a longer run of data than one trading year
can be employed, and the BIS does not s tipulate which
model should be employed for the calculations. Changes
in any of these factors could potentially result in large
changes in the calculated M C R R , so it is imp ortant that
all candidate models for the calculation of VaR be thor
oughly evaluated.
The mul t ip l ica t ion fac tor , which has a minimum
value of
3 ,
dep ends on the regulator 's view of the quality
of the bank's risk management system, and more precisely
on the backtesting results of the models. Unsatisfactory
results might see an increase in the multiplication factor
of
3,
up to a maximum of 4. The regulator performs an
assessment of the soundness of the bank's procedure in
the fo l lowing way. Un derp redi c t io n of losses by VaR
models (that is, the days on which the bank's calculated
value at risk is insufficient to cover the actual realized
losses in i ts t rad ing book ) are term ed exce pt ion s .
Between zero and four exceptions over the previous 250
days places the bank in the Green Zone; between five
and nine, i t is in the Yellow Zo ne ; and wh en 10 or m ore
exceptions are noted, the bank is in the Re d Z one . W he n
the bank is in the Yellow Zone, one would almost cer
tainly expect the regulatory body to increase the multi
plication factor, while if the firm falls into th e R ed Zo ne ,
it is likely to be no longer permitted to use the internal
modeling approach. I t will instead be required to revert
back to the building bloc k approach, wh ich does not
include a reductio n in the M C R R for diversified books
and whi ch w ill almost certainly yield a mu ch h igher c ap
ital charge. It is thus important to the securities firm or
bank, as well as its regulators, that its risk measurement
procedures are sound.
The VaR for each individual index was estimated,
using the s imple 5% del ta-n orm al approach proposed
in the literature, i.e.,
VaR
=
Marked position of asset
×
sensitivity to price movement
×
Adverse price movement per day
(10)
Th e sensitivity to price mo vem ents is taken to be 1
since we s tudy equities wh ich are linear instruments; the
adverse price move per day is equal to (1.645 × 3 × a)
where σ is the estimated standard deviation of the asset
returns over the sample period. Note that we multiply
the VaRs by the regulatory scaling factor of 3 , and we use
the 5% one-sided normal cr it ical value. While the Basle
rules require the use of a 1% VaR , we use a 5% VaR since
the use of 1% together with the scaling factor results in a
VaR that is so large as to re nder the m odels virtually ind is
tinguishable from one another. σ is calculated on a length
of three years of data (based on the above seven men
tioned models) and the sample is rolled over after each
quarter (60 days). For our selected data sample, this leads
to 46 separa te subsamples and out-of -sample per iods
wh ich can be used to evaluate the adequacy of the valu e-
at-risk estimates.
Each model for VaR outlined above is estimated for
the six series under investigation and for the portfolio of
Southeast Asian equities. In the cases wh ere a p aram etric
mo del for the cond itional co variances is no t specified (the
variance, the semi-variance, all univariate models) , we
estimate the portfolio returns and determine the VaR for
the portfolio as a single series. Such an approach may use
fully be term ed the full valu ation approa ch, since i t
involves a calculation o f the return s, and therefore the full
value of the portfolio, for each time period. These results
for the portfolio calculated from a single column of port
fo lio re turns wi l l be d isplayed und er the un iva r ia te
mo del heading . For the mul t ivar ia te G A R C H m odels ,
we use a slightly different app roach, w hic h m ay be view ed
as a mod ified version of the volatilities and co rrela tions
method. This approach makes use of Modern Portfolio
Theory, whereby for an N-asset portfolio, the value at
risk can be calculated by using the following formula:
wh e r e Va R is the value at risk of the portfolio , α
i
are the
weights given to each of the assets in the portfolio, VaR.
are the values at risk of the individual series A and B, and
p
ij
is the estimated correlation between the returns to i
and
j
Since under a mul t ivar ia te GARCH approach , we
have a specification, and can therefore generate forecasts,
for the conditional covariances as well as the conditional
variances, we make use of these in Equation (11) above.
Thus the VaRs for the individual assets and the correla
tions are estimated using the forecasts obtained from the
m u l t i v a r i a t e G A R C H , E G A R C H , a n d G J R m o d e l s .
Then the equally weighted portfolio VaR is constructed
using (11) and labeled as M G A R C H , M GJ R, and
M E G A R C H fo r t h e m u lt iv a ri at e G A R C H , G J R , a nd
WINTER 2003
THE JOURNAL OF RISK FINANCE 35
-
8/18/2019 Eb 022959
9/16
E X H I B I T
3
Value-at-Risk Estimates: Stock Index Retu rns for the Period January 1, 1985-April 29, 1999 (3,737 Observations)
Hong Kong Japan
Singapore S. Korea
Thailand
Portfolio
S&P 500
Panel A: Symmetric Approaches
Variance
Mean
Std. Deviation
EWMA
Mean
Std. Deviation
G RCH
Mean
Std. Deviation
MG RCH
Mean
Std. Deviation
0.07411
0.01356
0.07483
0.04858
0.07317
0.03670
0.07330
0.02089
0.06019
0.01878
0.06254
0.03006
0.05534
0.02792
0.05904
0.01911
0.05165
0.01489
0.05851
0.03903
0.04908
0.02672
0.05046
0.01630
0.07002
0.02584
0.07463
0.03579
0.06793
0.04318
0.06728
0.02191
0.07486
0.01676
0.08288
0.04316
0.07152
0.04619
0.07031
0.02862
0.04005
0.00991
0.04345
0.02333
0.03208
0.04023
0.03536
0.02259
0.02010
0.00714
0.01841
0.01174
0.05273
0.03867
-
Panel B: Asymmetric Approaches
GJR
Mean
Std. Deviation
EGARCH
Mean
Std. Deviation
Semi-Variance
Mean
Std. Deviation
MGJR
Mean
Std. Deviation
MEG RCH
Mean
Std. Deviation
0.07335
0.03716
0.07399
0.05252
0.07461
0.01316
0.07397
0.02714
0.07401
.
0.04860
0.05618
0.02766
0.05923
0.02709
0.06025
0.01142
0.06004
0.02265
0.06014
0.02610
0.05062
0.02749
0.05082
0.08298
0.05166
0.01253
0.05094
0.01752
0.05100
0.02718
0.06805
0.04842
0.06953
0.05331
0.07009
0.01963
0.06818
0.02354
0.07011
0.03047
0.07289
0.04887
0.07351
0.05271
0.07500
0.01570
0.07349
0.03045
0.07414
0.03283
0.03493
0.03931
0.03823
0.07184
0.04030
0.00742
0.03696
0.02809
0.03920
0.02881
0.02025
0.00718
0.01935
0.00836
0.01800
0.00520
-
-
Note:
Model acronyms
re
as
follows:
variance denotes VaR calculated using the historical standard deviation of
returns;
EWMA denotes the VaR calculated usin
the exponentially weighted moving average
method;
MGJR and MEGARCH denote the multivariate GJR and EGARCH models respectively; semi variance denot
a VaR calculated using the scaled semi standard deviation of (9). Cell entries refer to the average and standard deviation of the VaR across the 46 out of sample wi
dows; VaR is expressed as a proportion of the initial value of the position. The portfolio is an equally weighted combination of the five Southeast Asian market returns.
EGARCH models respectively, and these results are dis
played unde r the multivariate model heading.
RESULTS OF V A R ESTIM TION
N D EV LU TION
The VaR estimates for each asset and for each model
are presented in Exhibit 3. Comparing the VaR estimates
across Southeast Asian countries, they are highest for Hong
Kong and Thailand, and lowest for the U.S., Singapore,
and the portfolio; unsurprisingly, the countries with the
highest VaR estimates were also those which had the highest
unconditional variance of returns. For most models, the
country with the least stable VaRs over time, shown by
the larger standard deviation of VaR across the 46 win
dows, is South Korea, which did not have one of the highest
average VaRs. Interestingly, the average of the VaR estimates
for the S&P is approximately half that of the portfolio of
five Southeast Asian stock indices, and about one-quarter
to one-third of that of the individual component indices.
Comparing the VaR estimates across models, on the
whole the average VaRs are fairly similar. The uncondi
tional variance and the semi-variance-based estimates and
the multivariate EGARCH model seem to give the largest
VaRs,
while the simple GARCH models (both univariate
and multivariate), which do not allow for asymmetries, give
3 6 THE EFFECT OF ASYMMETRIES ON STOCK INDEX RET URN VALUE-AT-RISK ESTIMATES
WINTER 2003
-
8/18/2019 Eb 022959
10/16
E X H I B I T 4
VaR for the Portfolio in 6 Ro l l ing Sa m ples as a Proportion of Initial Value,
Est imated Using GARCH
and
Semi-Variance
the lowest average VaRs across the 46 windows for all coun
tries, except for the U.S., where the univariate symmetric
G A R C H V a R s are uncharacteristically high. For example,
in the case of the equally w eighted portfolio, the univariate
GARCH(1,1) model yields an average daily VaR of aro und
3.2%, while the semi-variance-based estimator provides an
average VaR of ust above 4%. The least stable VaRs, evi
denced
by the
highest standard deviation across
the 46
w i n
dows, arise from the E G A R C H m o d el s, in bo th their
univariate and multivariate forms. By far the most stable
VaRs over time for all countries and the portfolio are those
calculated using the semi-variance. The semi-var iance-
based VaRs have a standard deviation across windows of
the order of one-half that of other mod els, except for those
based on the (symmetric) variance, which have only slightly
higher variabilities across windows. In the context of the
S&P index returns, the semi-variance gives a mean VaR of
1.8%, the
lowest of any mod el.
An
example of the stability
of the VaR based on the semi-variance com pared w ith its
competitors can be gleaned from Exhibit 4, which plots
the estimated VaRs
for the
portfolio using
the
G A R C H
model and the semi-variance. Both
the
high er average value-
at-r isk estimate under the la t ter model , and its relative
smoothness over time, are apparent. The stability of the
semi-variance VaR estimates relative to those calculated
using the variance is quite surprising. If the returns were iid
normally distributed, the semi-variance VaR should have
a larger sampling va riability since it employs approximately
half as ma ny observation s as for the variance (and the trailing
samples used for the variance and semi-variance cover the
same three-year period).
Yet the
semi-variance VaRs
are
equally stable
for
H o n g K o n g
and are
more stable
for all
other series. Clearly, however, the return distributions are
non-normal and in particular are asymm etric. It also appe ars
to
be the
case that
the
lower halves of the distributions that
are picked ou t by the sem i-variance are m ore stable over time
than the upper tails, which are included in the variance cal
culations but not in the semi-variance.
Finally, comparing the univariate models with their
corresponding multivariate counterparts ,
we
note that
in
all cases, that
is, for
returns
to the
individual indices
and
fo r the portfolio, allowing for t ime-va ry ing co -mov e
ments between
the
series leads
to
slightly higher average
VaRs which are considerably more stable over time.
O ne may suggest at first blush tha t an optimal m odel
is one that gives the lowest average and the most stable
VaR estimate over t ime. A low VaR wou ld be deemed
preferable for the obvious reason that a lower VaR implies
that the firm should be required to tie up less of its capital
in an unprofitable, liquid form, while a stable VaR would
be appealing on the grounds that a highly variable VaR
would make it difficult to assess the riskiness of the secu
rities firm over
the
long term. However,
a
more appro
priate test
of
the adequacy
of
the
VaR
models
is
under
an
assessment of how they actually perform wh en used on a
holdout sample ( backtests in the Basle Comm ittee ter
minology). Regulators impose severe penalties
on
firms
whose models generate more than an acceptable num ber
of exceedances or except ions (see section 4 above),
and for this reason as well as to minimize the p ossibility of
financial distress, it is in the firm's interests to ensure that
its VaR mo dels perf orm satisfactorily in backtests. Back-
WINTER 2003
THE J OUR NAL
OF RISK
FINANCE
37
-
8/18/2019 Eb 022959
11/16
E X H I B I T 5
Backtesting (i.e. Out-of-Sample Tests) of the VaR for Stock Index Returns January 1985-April 29, 1999
(3,737 Ob servations)
Hong Kong
J a p a n S i n g a p o r e
S . Korea Tha i l and Por t fo l i o
S&P 500
P a n e l A : S y m m e t r i c A p p r o a c h e s
V a r i a n c e
Mean
Std. Deviat ion
Green
Yellow
Red
E W M A
M e a n
Std. Deviat ion
Green
Yellow
Red
G A R C H
M e a n
Std. Deviat ion
Green
Yellow
Red
M G A R C H
M e a n
Std. Deviat ion
Green
Yellow
Red
1.10870
1.43338
4 3
3
0
1.65217
2 . 0 8 3 1
4 2
4
0
1.84783
2 . 2 4 0 7 1
3 9
6
1
1.43478
1.77203
4 1
5
0
0 . 3 9 1 3 0
0 . 8 8 1 3 7
4 6
0
0
1.78261
5 . 9 0 6 3 5
4 3
1
2
1.04348
2 . 4 6 7 1 8
4 1
4
1
0 . 5 2 1 7 4
1.41011
4 4
2
0
0 . 7 8 2 6 1
1.28085
4 4
2
0
1.34783
2 . 0 7 8 6 5
4 2
4
0
1.43478
2 . 2 8 6 7 0
4 1
3
2
1.13043
1.52911
4 3
3
0
0 . 6 0 8 7 0
1.02151
4 6
0
0
0 . 7 6 0 8 1
2 . 6 2 6 3
4 4
1
1
1.00070
1.98889
4 3
3
0
0 . 5 4 3 4 8
1.14904
4 4
2
0
1.36957
1.27120
4 5
1
0
1.43478
2 . 1 6 6 9 5
4 2
3
1
1.69565
3 . 8 2 8 9 5
4 2
2
2
1.30435
1.77476
4 2
4
0
1.06522
1.10357
4 6
0
0
1.69565
2 . 4 5 7 3 7
4 2
3
1
1.00000
1.57762
4 3
3
0
0 . 1 0 8 7 0
0 . 7 3 7 2 1
4 5
1
0
0 . 3 2 6 0 9
0 . 7 0 0 9 3
4 6
0
0
0 . 6 3 0 4 3
1.10270
4 6
0
0
0 . 5 6 5 2 2
0 . 7 7 8 9 5
4 6
0
0
-
P a n e l B : A s y m m e t r i c A p p r o a c h e s
GJR
M e a n
Std. Deviat ion
Green
Yellow
Red
E G A R C H
Mean
Std. Deviat ion
Green
Yellow
Red
S e m i - V a r i a n c e
M e a n
Std. Deviat ion
Green
Yellow
Red
M G JR
M e a n
Std. Deviat ion
Green
Yellow
Red
M E G A R C H
M e a n
Std. Deviat ion
Green
Yellow
Red
1.73913
2 . 1 2 3 2 6
4 1
5
0
0 . 8 6 9 5 7
1.80873
4 0
6
0
1.00000
1.07497
4 6
0
0
1.36957
1.70436
4 2
4
0
0 . 5 6 5 2 2
1.40874
4 4
2
0
0 . 8 2 6 0 9
2 . 0 1 4 4 4
4 3
3
0
0 . 8 4 7 8 3
2 . 2 3 0 7 7
4 1
5
0
0 . 2 6 0 8 7
0 . 4 9 1 4 7
4 6
0
0
0 . 6 0 8 7 0
1.86760
4 3
3
0
0 . 1 3 0 4 3
0 . 8 8 4 6 5
4 5
1
0
1.17391
1.62350
4 1
5
0
0 . 9 3 4 7 8
1.59725
4 2
4
0
0 . 9 5 6 5 2
1.26415
4 6
0
0
1.06522
1.43608
4 3
3
0
0 . 1 9 5 6 5
1.18546
4 5
1
0
1.00000
1.98886
4 3
3
0
0 . 9 2 1 7 4
1.89626
4 4
2
0
0 . 5 2 1 7 4
0 . 9 6 0 0 7
4 6
0
0
0 . 1 7 3 9 1
1.03932
4 5
1
0
0 . 1 3 0 4 3
0 . 7 4 8 5 9
4 5
1
0
1.21739
1.56224
4 4
2
0
0 . 3 2 6 0 9
0 . 8 7 0 6 2
4 5
1
0
1.17391
1.23476
4 6
0
0
1.21739
1.93118
- 4 2
4
0
0 . 3 6 9 5 7
1.08236
4 5
1
0
0 . 9 9 8 6 4
1.21100
4 4
2
0
0 . 3 2 6 0 9
1.24819
4 4
2
0
0 . 9 5 6 5 2
1.09456
4 6
0
0
0 . 1 7 3 9 1
0 . 7 6 8 9 6
4 5
1
0
0 . 3 0 4 3 5
1.05134
4 5
1
0
0 . 3 6 9 5 7
0 . 7 4 1 1 3
4 6
0
0
0 . 5 2 1 7 4
0 . 8 8 7 9 2
4 6
0
0
0 . 6 3 0 4 1
1.10270
4 6
0
0
-
-
Note: Model
acronyms
are as ollows:
variance denotes
V aR
calculated
using the
historical standard deviation
of returns; EWMA
denotes
the VaR
calculated
using
the exponentially
weighted
moving
average
method; MCJR and MEGARC H
denote
the multivariate GJR and EGARC H
models respectively; semi variance
denotes a VaR
calculated
using the
scaled
semi-standard deviation of
(9).
Cell
entries refer
to the
average
and standard deviation of the number of
exceedances
of
the VaR
across
the 46 out-of-sample windows, and the number of times that
such
a number of
exceedances
would have
placed
the irm in the Green, Yellow, and
Red Zones; VaR is
expressed
as a
proportion
of
the
initial value of
the
position. The portfolio is an equally weighted
combination
of
the
five Southeast Asian
market returns.
3 8 TH E EFFECT OF ASYMMETRIES ON STOCK INDEX RET UR N VALU E-AT-RISK ESTIMATES
WINTER 2003
-
8/18/2019 Eb 022959
12/16
E X H I B I T
6
N u m b e r
of
Exceedances
for the
Portfolio
in 6
Ro l l ing Sa m ples
of VaR
Calculated
Using GARCH
and
Semi-Variance
test results
for
each model
and for
each asset series
are
p re
sented in Exhibit 5. The mean and standard deviation of
the percentage of days
for
which there
was an
exceedance
in each
of the 46
rolling samples
of
length
250
observa
tions are given, together w ith the number of t imes where
the firm's
VaR was
insufficient
on
enough days
in the
sample
to
place
it in the
Yellow
or Red
Zone. Again, first
considering the results across coun trie s, in general the esti
mated VaR models for Hong Kong and Singapore seem to
be
the
least adequa te, while those
for the
U.S., South Korea,
an d
for the
portfolio seem
to
yield
the
fewest exceptions.
On ly the univariate GARCH model would have ever led
a securities firm with positions
in
these equities into
the
Red Zone, with consequent disallowance
of
the internal
mode l and increased capital charges. Of the 46 sample
periods,
the
models
for
four
of
the five countries would
have been
in the Red
Z o n e
at
least once— with
the
worst
being Singapore and Thailand, wh ere the firm wo uld have
been
in the Red
Z o n e
for two
periods.
C ompar ing
the
G A R C H
and
mul tivar ia te G A R CH
models with their asymmetric counterparts,
the
latter seem
to
do
somewhat bet ter .
The
average propor t ion
of
exceedances is lower for the GJR and E G A R C H m o d el s,
especially
the
latter, than those conditional variance models
that
do not
allow
for
asymmetries ,
for all
countries .
For
example, the average percentage of exceedances in the case
of Thailand is approximately
1.7
using
the
GA R C H m odel ,
but only
1.2 and 0.3 for the GJR and
E G A R C H m o de ls
respectively. Also, the multivariate models tend to fare
better than their univariate counterparts , irrespective
of
whethe r
the
models
are
symm etr ic
or
asymmetric. This
is
particularly tru e in the portfolio context, w here the model
estimation approach uses
the
time-v arying forecasts of the
conditional covariances
as
well
as the
volatilities.
For
example, the univar ia te G A R C H m odel has an average of
exactly
1%
exceedances, placing
the
firm
in the
Yellow
Z o n e
on
three occasions;
the
multivariate GA R C H model,
on the other hand, has an average percentage of exceedances
of only 0 .1% , placing the firm in the Yellow Z on e only
once.
In the
case of the
U.S.
stock re turns,
all
models
are
deemed safe, with
no
ventures into either
the
Yellow
or
Red Zones . This appears to arise from a fall in S&P
volatility d urin g
the
out-of-sample periods compared with
the start
of
the sample.
For
example, splitting
the
whole
sample exactly in half, the S&P variance in the second half
is approximately 50%
of
that
in the
first
half.
In terms
of
the Basle Committee criteria,
the
clear
winne r is the semi-variance-based estimator, whic h, if it
had been employed
by a
securities firm with long posi
tions
in any
of these marke t indices
or
a portfolio of them ,
would never have been
out
of the Green Z one . W hile this
improved performance would
not
come costlessly
to the
firm, in
the
sense that
the VaR
from
the
semi-variance-
based model
is
almost always
the
highest,
the
improvement
in perform ance is considerable relative to the models which
genera ted
the
next highest
VaR
levels
(the
asymmetr ic
multivariate models).
The
usefulness of the measure based
on the semi-variance is also shown in Exhib i t 6, which
presents
the
number
of
exceedances
for the
portfolio
of
stock returns, using this method
and
also
the
univariate
WINTER 2003
T H E J O U R N A L OF RISK FINAN CE 39
-
8/18/2019 Eb 022959
13/16
G A R C H m o d e l . As can be seen, the semi-variance pro
duces, for most w indow s, a smaller num ber o f exceedances
in the 250-day holdou t samples than the G A R C H m o d e l.
Moreover, comparing Exhibits 4 and 6, the numbers of
G A R C H m o d e l e x c e e d a n c e s are greater when the
GARCH model VaR calcu la t ions are considerably below
those of the sem i-variance e stimator. This seems to sug
gest that the G A R C H m o de l at times underestimates the
VaR relative to the semi-variance m odel and relative to
the actual out- turn.
C O N C L U S I O N S
This study sought to consider the effect of two classes
of asymmetry on the size and adequacy of market-based
value-at-risk estimates in the context of
five
Southeast Asian
stock market indices, and an equally weighted portfolio
comprising these indices, with S&P 500 index returns exam
ined for comparison. We found significant statistical ev i
dence that both uncond itional skewness and a conditionally
asymmetric response of volatility to positive and negative
returns were present in the data. We then examined a number
of symmetric and asymmetric models for the determina
tion of a securities firm's position risk requirement. Our pri
mary f inding was that the semi-variance mod el , which
explicitly allows for asymmetry, leads to more stable VaRs
which would be deem ed m ore accurate under the Basle
Com mittee rules than models which do not allow for such
asymmetries. In particular, VaR estimates based upon a simple
modification to the usual semi-variance estimator were the
only ones whic h w ould have left
the
firm w ith
a
margin
of
safety during every time period and for every asset.
Put another way, models that do not allow for asym
metries either
in the
uncon ditional retu rn dis tr ibution
or
in the response of volatility to the sign of returns, lead
to inappropriately small VaRs. Altho ugh the cost of
c a p
ital is on ave rage l% -20% h ighe r for the as ymmet r i c
models , we conjecture that the additional m argin of safety
is necessary and makes the addit ional cost w or thw hi le .
Firms w hich fall into the Yel low Z on e are likely to have
their multiplication factor raised from 3 to 4 or 5, an
increase of at least
3 3%
in the required VaR. Firms w hich
trip into
the Red
Z o n e w o u l d
be
forbidden from using
an internal model, and wou ld be required to revert to the
building block approach which specif ies a flat charge of
8% for equi t ies . This would represent an approximate
doubl ing of the required capital for the por tfo l ios co m
pared with that reported for the asymm etr ic models in
Exhib i t 3.
Skewness is a hitherto vir tually neglected feature of
financial asset return series, and it seems plausible that
better value-at-r isk estimates should arise from method
ologies which are able to capture all of the stylized fea
tures that are undeniably present in the data. We propose
that future research may seek to form a mode l wh ich can
capture the unco ndition al skewness in the data as well as
volatility clustering, leverage effects, and uncondi t ional
kurtosis ; one such mode l is the autoregressive cond itional
skewness formulation, recently proposed by Harvey and
Sidd ique [1999 ] , wh ich , due to its infancy, is as yet
untested
in the
risk man agem ent arena.
E N D N O T E S
1
The models employed here are widely used and hence
only brief model descriptions are given, although see Brails-
ford and Faff [1996] or Brooks [1998] for thorough discussions
of alternative models
for
prediction of volatility and their rel
ative forecasting performances under standard statistical evalu
ation metrics.
2
The method
of
maximum likelihood using
a
Gaussian
density
and
employing
the
BFGS algorithm
is
employed
for
estimation of the parameters of all models from the G A R C H
family, including the multivariate and asymmetric models.
3
The model coefficient estimates for each mo del are not
presented due to space constraints, although they are available
upon request from the authors.
4
In fact, this assumption can be made witho ut loss of gen
erality, for a zero mean series
x
t
:
t
=
1, . . ., T can be constructed
by subtracting the mean of the series from each observation t.
5
In (9), this is scaled by T_/ T_ - 1) 1 as T_ ∞ o
obtain a consistent (asymptotically unbiased) es timator.
RE FE RE NCE S
Alexander, C.O ., and C.T . Leigh. On the Covariance Models
Used in Value at Risk Models. Th e
Journal
of
Derivatives
4
(1997),
pp. 50-62.
Beltratti, A., and
C.
Morana. Com puting Value at Risk with
High Frequency Data.
Journal
of
Empirical
Finance
6
(1999),
pp .
431-455.
Berkowitz, J. Tes ting Density Forecasts, with Applications
to Risk Management. Journal of Business and Economic Statis-
tics
19 (2001), pp. 465-474.
Black, F. Studies in Price Volatility Changes. In
Proceedings
of the
976 Meeting
o f
the Business and Economics Statistics
Section
American Statistical Association
(1976), pp. 177-181.
40 THE EFFECT OF ASYMM ETRIES ON STOCK INDEX RETU RN V ALUE-AT -RISK ESTIMATES
W I N T E R 2003
-
8/18/2019 Eb 022959
14/16
Dollers lev, T. , R.F . Engle, and J .M . Wo oldr idge . A Capital
Asset Pr icing Model with Time-Varying Covar iances . Journal
of Political Economy, 96 (1988) , pp. 116-131 .
Brailsford, T.J. , and R. W .
Faff.
An Evaluat ion of Volat i li ty
For ecas t ing Techniques . Journal of Banking and Finance, 2 0
(1996) , pp. 419-438.
Brooks, C. Forecast ing Stock Retu rn Volat il i ty: Does Vo lume
Help? Journal of Forecasting,
17 (1998) , pp. 59-80.
Brook s, C. , A.D. Clare , and G. Persand. A W or d of Cau tion
on Calculat ing Market-Based Minimum Capital Risk Require
m e n t s .
Journal of Banking and Finance,
14 (2000) , pp. 1557-
1574.
——. A n Ext r eme V alue A ppr oach to Ca lcu lat ing M inim um
C a p i t a l R i s k R e q u i r e m e n t s . The Journal of Risk Finance, 3
(2002) ,
p p . 22-33 .
Br ook s , C . , and G . Per sand . V a lue a t Risk and M ar ke t
Crashes. Journal of Risk, 2 (2000a) , pp. 5-26.
——. T he Pitfalls of Va R Estimates . Risk, 13 (2000b) , pp.
63- 66 .
Campbe l l , J . , and L . H entsche l . N o N e w s I s G o od N e w s : A n
A symmetr ic Mod e l of Changing V ola ti li ty in S tock Re tur n s .
Journal of Financial
Economics, 31 (1992) , pp. 281-318.
Cho obin eh , F . , and D . Br ant ing . A Simple A ppr oximat ion
for Semivar iance.
European Journal of Operational Research,
27
(1986) , pp. 364-370.
Chr i s t i e , A . Th e S tochast ic Behav iour of C om m on S tock
Variance: Value, Leverage and Interest Rate Effects. Journal
of Financial Economics, 10 (1982) , pp. 407-432.
Dim son, E. , and P . Marsh. Capital Requ irem ents for Secu
rities Firms. Journal of Finance, 50 (1995) , pp.
821-851 .
—— . Stress Tests of Capital
Requirements. Journal of Banking
an d Finance, 21 (1997) , pp. 1515-1546.
D o w d , K . Beyond Value at
Risk:
The New Science of Risk Man
agement.
Chich ester , UK : Wiley, 1 998.
Engle, R. F. , and K. Krone r . Mu lt ivar iate Simultaneous G en
e ra li sed A R C H . Econometric Theory, 11 (1995) , pp. 122-150.
Engle, R. F. , and V. Ng . Me asur ing and Test in g the Impact
of N ew s on
Volatility. Journal of Finance,
48 ( 1993) , pp . 174 9-
1778.
Glosten, L.R. , R . Jagannathan, and D. Ru nk le. O n the Re la
t ion Be tw een the Expec ted V a lue and the V ola t i l i ty of the
Nom inal Excess Re turn on Stocks . Journal
of Finance,
48 (1993),
p p .
1779- 1801.
Grootv eld, H. , and W . Hallerbach. Var iance vs . Do wn side
Risk: I s Th ere Really Th at M uc h Dif ference?
European Journal
of
Operational
Research,
114 (1999) , pp. 304 -319 .
H ar low , W.V . A sse t A l locat ion in a D ow ns ide - Risk F r ame
w o r k .
Financial Analysts Journal,
S e p t e m b e r - O c t o b e r 1 9 9 1 ,
pp . 28- 40 .
Harvey , C.R . , and A. Siddique . Autoregress ive Condit io nal
S k e w n e s s .
Journal of Financial and Quantitative Analysis,
3 4
(1999) , pp. 465-487.
H e n r y , O . T . M o d e l l i n g t h e A s y m m e t r y o f S t o c k M a r k e t
Volat i l i ty. Applied Financial Economics, 8 (1998) , pp. 145-153.
H op pe , R . V A R and the U nr ea l Wo r ld . Risk, 11 (1998), pp.
45- 50 .
Huism an, R. , K.G. Koedijk, and R.A.J . Pownall . Va R- x: Fat
Tails in Financial Risk M ana gem ent.
Journal of
Risk, 1 (1998),
pp . 47 -61 .
Hull , J . , and A. W hit e . Value at Risk W he n Daily Changes
in Market Variables Are Not Normally Distributed.
Th e
Journal
of Derivatives, 5 (1998) , pp. 9-19 .
Jackson, P . , D.J . M aud e, and W. Perraudin. Test i ng Value at
Risk Approaches to Capital Adequacy.
Bank of England Quar
terly Bulletin, 38 (1998) , pp. 256-266.
Jor ion , P . B ig Bets Gone Bad: Derivatives a nd Bankruptcy in Orange
County.
San Diego: Academic Press , 1995.
— — . Value at Risk: The New Benchmark for Controlling Market
Risk. Chicago: I rwin, 1996.
Joseph y, N.H . , and A.D . Aczel . A Stat is t ically Optim al Est i
mator o f Semivar iance. European
Journal
of Operational Research,
67 (1993) , pp. 267-271 .
J .P .
Morgan . Riskmetr ics Technical Do cum ent. 4th ed. 1996.
K on, S.J . Mod e ls of S tock Re tu r ns— A Com par i son .
Journal
of Finance, 39 (1984) , pp. 1 47-1 65.
Kup iec, P . Tec hniq ues for Ver ifying the Accuracy of Risk
M e a s u r e m e n t M o d e l s .
Th e Journal of Derivatives, 2
(1995) , pp.
73- 84 .
WINTER 2003
TH E JOURNAL O F
RISK FINANCE
4 1
http://www.emeraldinsight.com/action/showLinks?system=10.1108%2Feb043485http://www.emeraldinsight.com/action/showLinks?system=10.1108%2Feb043485
-
8/18/2019 Eb 022959
15/16
Lon gin , F .M. F r o m V alue a t Risk to S tr ess Tes t ing : T he
Ext r eme V alue A ppr oach .
Journal of Banking and
Finance, 2 4
(2000) , pp. 1097-1130.
Lopez, J .A. , and C.A. Walte r . Evalua t ing Cov ar iance M atr ix
Forecasts in a Value at Risk Framework. Journal of
Risk,
3
(2001) , pp. 69-98.
Ma dan, D.B. , and E. Seneta . Th e Var iance Gam ma (V.G.)
Model for Share Market Returns .
o u r n a l of Business,
63 (1990),
p p . 511- 524.
M cN eil , A.J ., and R . Frey. Est imation of Tail-R elated R isk
Measures for Heteroscedastic Financial Tim e Series: An Ex treme
V a l ue A p p r o a c h . Journal of Empirical Finance, 1 (2000) , pp.
271- 300.
Neftci, S.N. Value at Risk Calculation s, Ex trem e Events and
Tail Est imation.
The Journal of D erivatives,
7 (2000) , pp . 23-37 .
Nelso n, D.B . Co nditi ona l Heteroskedasticity in Asset Re turn s:
A N e w A p p r o a c h . Econometrica, 59 (1991) , pp. 347-370.
Pagan, A.R . , and G.W. Schwer t . Alternat ive Models for C on
dit ional Stock V olat i l i ty. Journal o f Econom etrics, 45 (1990) , pp.
267- 290.
Press , S .J . A Co m po un d Events M od el for Secur i ty Pr ices .
Journal of Business,
40 ( 1967) , pp . 3 17- 3 35.
Vlaar, P .J .G. Value at Risk Mo dels for Du tch Bo nd P or tfo
l ios . Journal of Banking and Fina nce, 24 (2000) , pp. 1131-1154.
To order reprints of this article, please contact Ajani Malik at
[email protected] or 212-224-3205.
4 2 TH E EFFECT OF ASYMMETRIES ON STOCK INDEX RET UR N VALU E-AT-RISK ESTIMATES
WINTER. 2003
mailto:[email protected]:[email protected]:[email protected]
-
8/18/2019 Eb 022959
16/16
This article has been cited by:
1. Sung Chil Yeo, Zhaojing Li. 2015. Performance Analysis of Volatility Models for Estimating Portfolio Value at Risk. Korean Journal of Applied Statistics 28, 541-559. [CrossRef ]
2. Alberto Burchi. 2013. Capital requirements for market risks. Journal of Financial Regulation and Compliance 21:3, 284-304.[ Abstract] [Full Text] [PDF]
3. References 479-520. [CrossRef ]
4. Timotheos Angelidis, Stavros Degiannakis. 2008. Forecasting one‐day‐ahead VaR and intra‐day realized volatility in the Athens Stock Exchange Market. Managerial Finance 34:7, 489-497. [ Abstract] [Full Text] [PDF]
5. Christos Floros. 2007. The use of GARCH models for the calculation of minimum capital risk requirements. International Journal of Managerial Finance 3:4, 360-371. [ Abstract] [Full Text] [PDF]
6. Timotheos Angelidis, Stavros Degiannakis. 2005. Modeling risk for long and short trading positions. The Journal of RiskFinance 6:3, 226-238. [ Abstract] [Full Text] [PDF]
http://www.emeraldinsight.com/doi/pdfplus/10.1108/15265940510599838http://www.emeraldinsight.com/doi/full/10.1108/15265940510599838http://dx.doi.org/10.1108/15265940510599838http://www.emeraldinsight.com/doi/pdfplus/10.1108/17439130710824361http://www.emeraldinsight.com/doi/full/10.1108/17439130710824361http://dx.doi.org/10.1108/17439130710824361http://www.emeraldinsight.com/doi/pdfplus/10.1108/03074350810874442http://www.emeraldinsight.com/doi/full/10.1108/03074350810874442http://dx.doi.org/10.1108/03074350810874442http://dx.doi.org/10.1002/9780470688014.refshttp://www.emeraldinsight.com/doi/pdfplus/10.1108/JFRC-10-2012-0042http://www.emeraldinsight.com/doi/full/10.1108/JFRC-10-2012-0042http://dx.doi.org/10.1108/JFRC-10-2012-0042http://dx.doi.org/10.5351/KJAS.2015.28.3.541