Diversication and Value-at-Risk

Christophe Prignon
a
, Daniel R. Smith
b,
*
a
HEC Paris, F-78351 Jouy-en Josas, France
b
Simon Fraser University, 8888 University Drive, Burnaby BC V6P 2T8, Canada
a r t i c l e i n f o
Article history:
Received 17 May 2008
Accepted 5 July 2009
Available online 9 July 2009
JEL classication:
G21
G28
G32
Keywords:
Value-at-Risk
Diversication
Dynamic conditional correlation
Copulas
a b s t r a c t
A pervasive and puzzling feature of banks Value-at-Risk (VaR) is its abnormally high level, which leads to
excessive regulatory capital. A possible explanation for the tendency of commercial banks to overstate
their VaR is that they incompletely account for the diversication effect among broad risk categories
(e.g., equity, interest rate, commodity, credit spread, and foreign exchange). By underestimating the
diversication effect, banks proprietary VaR models produce overly prudent market risk assessments.
In this paper, we examine empirically the validity of this hypothesis using actual VaR data from major
US commercial banks. In contrast to the VaR diversication hypothesis, we nd that US banks show no
sign of systematic underestimation of the diversication effect. In particular, diversication effects used
by banks is very close to (and quite often larger than) our empirical diversication estimates. A direct
implication of this nding is that individual VaRs for each broad risk category, just like aggregate VaRs,
are biased risk assessments.
2009 Elsevier B.V. All rights reserved.
1. Introduction
Most modern commercial banks routinely publicly disclose their
aggregate, rm-level Value-at-Risk (VaR). The latter measures the
maximum trading loss that a bank can face over a given horizon
(usually one day) and under a specied condence level (usually
99%). A pervasive and puzzling feature of banks VaR is its abnor-
mally high level. Indeed, Berkowitz and OBrien (2002) show that
US banks 99% VaR systematically exceed the actual 99 percentile
of their trading revenues and that GARCH-based VaR models tted
to actual daily trading revenues lead to smaller VaR estimates. Pri-
gnon et al. (2008) nd that, because of VaR overstatement, market
risk charges of Canadian banks are much larger (ve times larger for
some banks) than it would be with unbiased VaR estimates. Out of
the 7354 trading days analyzed in their study, there are only two
exceptions, or days when the actual trading loss exceeds the VaR,
whereas the expected number of exceptions with a 99% VaR is
74.
1
The direct cost for the bank to over-report its VaR is to maintain
an excessively high regulatory capital since its level is given by a po-
sitive function of the bank (average or most recent) VaR.
2
A rst at-
tempt to quantify empirically the cost of VaR over-reporting can be
found in Prignon et al. (2008).
An increasing number of banks also report an individual VaR for
each broad risk category (e.g., equity, interest rate, commodity,
credit spread, and foreign exchange). In their international survey,
Prignon and Smith (forthcoming) nd that 68% of their sample
rms publicly disclose individual VaRs. This particular disclosure
format allows outsiders to have a good understanding of the cur-
rent exposure of banks trading portfolio with respect to the main
sources of risk in the economy. As acknowledged by the Basel Com-
mittee on Banking Supervision (1996) in the Amendment of the
Based Accord, banks have discretion to recognize empirical corre-
lations within and across broad risk categories when computing
their aggregate or diversied VaR. In practice, because the correla-
tion across the risks is less than perfectly positive, the aggregate
VaR will be less than the sum of the individual VaRs.
3
A possible explanation for the tendency of commercial banks to
overstate their aggregate VaR is that they incompletely account for
the diversication effect among broad risk categories (see Berko-
witz and OBrien, 2002; Prignon et al., 2008). We refer to this con-
jecture as the diversication hypothesis. Under this hypothesis,
even with unbiased individual VaRs, the aggregate VaR will be
too large and the number of exceptions too small. If the diver-
sication hypothesis is valid (i.e., biased correlation structure),
0378-4266/$ - see front matter 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.jbankn.2009.07.003
* Corresponding author. Tel.: +1 778 782 4675; fax: +1 778 782 4920.
E-mail addresses: perignon@hec.fr (C. Prignon), drsmith@sfu.ca (D.R. Smith).
1
See Prignon and Smith (forthcoming) for further international evidence of VaR
over-reporting.
2
This phenomenon has been labeled by Bakshi and Panayotov (2007) as the
Capital Adequacy Puzzle. Breuer et al. (2009) argue that simply adding market and
credit risk charges tends to understate aggregate capital requirements.
3
As Deutsche Bank puts it in its 2005 annual report: Simply adding the Value-at-
Risk gures of the individual risk classes to arrive at an aggregate Value-at-Risk would
imply the assumption that the losses in all risk categories occur simultaneously.
Journal of Banking & Finance 34 (2010) 5566
Contents lists available at ScienceDirect
Journal of Banking & Finance
j our nal homepage: www. el sevi er . com/ l ocat e/ j bf
individual VaRs have a chance to be informative. However, if the
diversication hypothesis is not valid (i.e., correlation structure
properly accounted for), individual VaRs are causing the Capital
Adequacy Puzzle, and as a result, are biased risk assessments.
Although there are many potential motivations for banks to over-
state VaR, at least two relate to the regulatory environment facing
US (and many other countrys) banks. One such explanation for
VaR inating comes from the asymmetric cost function faced by
the banks. If the number of exceptions is too small (i.e., large
VaR), the cost corresponds to the cost of maintaining an excessive
regulatory capital, whereas if the number of exceptions is too large
(i.e., small VaR), the cost increases exponentially with the number
of exceptions (e.g., severe regulatory penalties, reputation effects).
Furthermore, under the US regulatory environment banks are re-
quired to estimate their VaR using at least one year of daily histor-
ical data. This leads to a sluggish dynamics for the VaR especially
when it is computed through historical simulations (Pritsker,
2006). In order to prevent clusters of exceptions in case of sudden
increase of volatility, VaR is inated.
The goal of this paper is to empirically test the validity of the
VaR diversication hypothesis. Our analysis is based on individual
and diversied VaR data from several major US commercial banks
between 2001 and 2007. For each bank/quarter, we compute the
actual diversication applied by the bank as d =

N
i=1
VaR
i

_
DVaR)=

N
i=1
VaR
i
, where DVaR and VaR
i
are the diversied, respec-
tively individual, 99% 1-day ahead VaRs disclosed by the bank and
N is the number of broad risk categories.
4
In our sample, there are
ve broad risk categories, i.e., equity, interest rate, commodity, credit
spread, and foreign exchange. Empirically, we nd that the d coef-
cient is typically in the 3455% range. We then compare the actual
diversication effect used by the bank to a benchmark estimated
from market indices that capture the ve broad risk categories. Spe-
cically, we aggregate the banks individual VaRs using an empirical
estimate of the correlation structure among risk category indices and
derive an empirical DVaR denoted DVaR
E
. In our tests, we model the
correlation structure using a variety of techniques: unconditional
correlation matrices, Dynamic Conditional Correlation model, BEKK
model, and a time-varying copula model.
The empirical diversication coefcient is dened as d
E
=

N
i=1
VaR
i
DVaR
E
_ __

N
i=1
VaR
i
and it is used as a benchmark for
the banks d. Supportive evidence for the VaR diversication
hypothesis would be nding d
E
d. However, our main result is
that the diversication effect used by most banks is very close to
(and quite often larger than) our empirical diversication estimate.
Overall, our results do not lend support to the hypothesis that VaR
over-reporting is primarily due to a partial account of the diversi-
cation among broad risk categories. A direct implication of this
nding is that individual VaRs, just like aggregate VaRs, are biased
risk assessments.
This study makes three contributions to the nancial risk man-
agement literature. First, we directly test the diversication
hypothesis using actual data on aggregate and individual VaRs.
Second, our empirical analysis can be viewed as the rst (indirect)
test of the accuracy of individual VaRs. Indeed, standard backtest-
ing procedures cannot be directly implemented on individual VaRs
since there is no available time-series data on prot-and-loss
disaggregated by source of risk. Third, we show how the recent
copula models with time-varying dependence parameter a la Pat-
ton (2006), Jondeau and Rockinger (2006), or Bartram et al.
(2007) can be used to compute the VaR of a given portfolio.
The rest of the paper is organized as follows. We present our
methodology in the next section and our empirical analysis in Sec-
tion 3. Robustness checks and extensions are discussed in Section
4. We summarize and conclude our study in Section 5.
2. Methodology
In this section, we present a methodology allowing us to com-
pute (1) the bank diversication coefcient and (2) the empirical
diversication coefcient from banks individual VaRs and risk
category indices. We denote the vector of asset returns by r and
its covariance matrix by H = DRD, where D is a diagonal matrix
with the standard deviation of asset i as element i on the princi-
pal diagonal: D = diag(r
i
), R is the correlation matrix, and
i = 1, . . . , N. A portfolio with weight x
i
in asset i has rate of return
r
p
= x
/
r and variance r
2
p
= x
/
Hx. If the total investment measured
in dollars is W then the dollar variance, which is required for
computing VaR, is r
2
p
W
2
= x
/
Hx where x = xW are the dollar posi-
tions. The VaR of the investment in asset i, or individual VaR, is
given by:
VaR
i
= jr
i
x
i
W = jr
i
x
i
(1)
and, as shown by Jorion (2006), the diversied VaR of the portfolio
is:
DVaR = j

x
/
Hx
_
: (2)
The scaling coefcient j depends on the particular distribution and
the coverage probability. For instance, when computing a 99% VaR
using a normal distribution, j = 2.33. Eq. (2) is most easily derived
under the assumption of multivariate normality, but applies
equally to the family of elliptical distributions (with j suitably
re-dened), which is a much broader class of distributions (see Jori-
on, 2006, p. 180 and, Appendix A for a proof). Combining individual
and diversied VaRs, we dene the bank diversication coefcient
as:
d =

N
i=1
VaR
i
DVaR

N
i=1
VaR
i
: (3)
We rewrite Eq. (2) as a function of the individual VaRs:
DVaR = j

x
/
DRDx
_
=

j
2
x
/
DRDx
_
=

V
/
RV
_
; (4)
where V is a column vector containing the individual VaRs. Eq. (4)
provides a very simple approach to computing the diversication
effect among a set of assets whose joint distribution is in the ellip-
tical family and only requires the correlation matrix. Eq. (4) also al-
lows us to see that, in case of perfect correlation between the assets,
DVaR is given by the sum of the individual VaRs. In all other cases,
DVaR will be affected by the diversication effects.
Implementing Eq. (4) requires the estimation of the correlation
matrix. A rst approach is to simply compute the sample correla-
tion matrix of the N asset returns using past observations. Alterna-
tively, the conditional correlation matrix can be computed using
the Dynamic Conditional Correlation (DCC) of Engle (2002), which
is a powerful tool for quickly estimating moderately high-dimen-
sional GARCH models. This model is based on a decomposition of
the log-likelihood function of a multivariate GARCH model. We
use quasi-maximum likelihood to estimate the parameters
describing the conditional covariance matrix H
t
= E(e
t
e
/
t
) = D
t
R
t
D
t
where e
t
= r
t
E
t1
(r
t
) denotes the unexpected returns and D
t
is a
diagonal matrix of conditional standard deviations that are mod-
eled as univariate GARCH models:
D
2
t
= diag(x
i
) diag(a
i
) (e
t1
e
/
t1
) diag(b
i
) D
2
t1
; (5)
4
To the best of our knowledge, this is the rst study to measure a diversication
coefcient from VaR data. We nd this measure appealing because of its simplicity
(i.e., percentage reduction in VaR due to diversication among sources of risk) and
because it can be computed for each bank at any point in time.
56 C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566
where denotes the Hadamard or element-by-element multiplier
and R
t
is the conditional correlation matrix. The correlation matrix
is expressed as:
R
t
= diag(Q
t
)
1=2
Q
t
diag(Q
t
)
1=2
(6)
and the matrix Q
t
is modeled as a GARCH-type function of the
lagged standardized residuals z
t
= D
1
t
e
t
that depends on only two
scalar parameters h
1
and h
2
:
Q
t
= R(1 h
1
h
2
) h
1
(z
t1
z
t1
) h
2
Q
t1
(7)
and R is the unconditional sample correlation matrix of the stan-
dardized residuals. The simplifying feature in the DCC model is that
the Gaussian log-likelihood function can be separated into two
components: one to identify the volatility parameters and the other
to identify the correlation parameters. The parameters that drive
the conditional variance dynamics are estimated by maximizing
the N univariate GARCH log-likelihoods, which is computationally
feasible, and then conditioning on the volatility dynamics estimate
the correlation parameters. Although we are maximizing the Gauss-
ian log-likelihood our estimates of the dynamics of the correlation
matrix apply even when returns are non-normal. We view the
Gaussian log-likelihood as simply an objective function that is max-
imized to produce quasi-maximum likelihood estimates that are
consistent and asymptotically normal and we thus report robust
standard errors.
Engle (2002) suggests a simple parameterization for the condi-
tional correlation in which h
1
and h
2
are scalars and R is set to

R,
the sample correlation matrix, so the nal stage optimization is
only with respect to two scalar parameters. If the volatility dynam-
ics are modeled using the standard GARCH(1, 1) model, which has
three parameters, then to estimate the N N covariance matrix H
t
,
there are N three-dimensional optimizations (to estimate each of
the N univariate GARCH model parameters) followed by a single
two-dimensional optimization. Thus quite high-dimensional prob-
lems can be handled computationally.
In the case of a bank disclosing its individual VaRs, we can esti-
mate the correlation matrix among risk category indices and gen-
erate from Eq. (4) the empirical DVaR which we denote by DVaR
E
.
From this alternative risk measure, we estimate an empirical diver-
sication coefcient:
d
E
=

N
i=1
VaR
i
DVaR
E

N
i=1
VaR
i
: (8)
The difference between d
E
and the bank diversication coefcient
is:
d
E
d =
DVaR DVaR
E

N
i=1
VaR
i
: (9)
By plotting the value of diversication differential d
E
d through
time for a given bank, one can see whether there is a systematic
underestimation of the diversication effect.
3. Data and empirical results
We collect quarterly data on individual and diversied VaRs
from four major US commercial banks between the fourth quarter
of 2001 and the rst quarter of 2007. Bank of America, JPMorgan
Chase, and Citigroup are the largest three US commercial banks
(in this order, based on total assets as of year-end 2005) and HSBC
ranks 8th. These banks are the only ones out of the 10 largest US
commercial banks to disclose the required information in their
10-Q and 10-K forms. Jorion (2002) and Hirtle (2003) show that
US banks quarterly VaRs do contain valuable information about fu-
ture risk exposures.
Although all sample banks disclose their VaR at the 99% con-
dence level, there are some differences in the exact nature of the
disclosed risk measures. Bank of America, Citigroup, and JPMorgan
Chase disclose their 1-day VaR whereas HSBC reported its 10-day
VaR level until the rst quarter of 2006 and then switched to a
1-day VaR. To allow cross-sectional comparisons, we convert all
10-day VaRs into 1-day VaRs using the square-root-of-time rule
(see Danielsson and Zigrand, 2006). Furthermore, Bank of America
is the only sample bank to disclose average VaRs instead of quar-
ter-end VaRs.
Consistent with previous evidence, we nd that all sample
banks over-report their VaR. Between 2001 and 2006, Bank of
America and JPMorgan Chase have experienced three and two
exceptions, respectively, whereas the expected number of excep-
tions is 15.
5
We formally test the null hypothesis that the actual
number of exceptions is equal to the theoretical one using an uncon-
ditional coverage test (see Jorion, 2006, p. 144):
z
UC
= (X 0:01 T)=

0:01 0:99 T
_
; (10)
where X is the number of exceptions, T is the length of the sample
period in days, and z
UC
is a statistic that follows a standard nor-
mal distribution. We nd that the null hypothesis is rejected at
the 1% level for Bank of America and JPMorgan Chase. The other
two sample banks do not disclose the exact number of exceptions
in their nancial reports. Instead, they report the histograms of
daily trading revenues and general statements about backtesting
that both suggest that the number of exceptions is close to zero
every year.
As shown in the left panels of Fig. 1, all banks but Citigroup
break-down their VaR into ve individual components focusing
on one major source of risk. Interest rate accounts for a large frac-
tion of a banks VaR, although credit spread became the largest
contributor to the most recent VaR of Bank of America. We also
note that for most sample banks the individual VaRs do not change
much from one quarter to the next.
6
For each sample bank, we display in the right panels of Fig. 1 a
graph of the diversied VaR. The latter measure is the sum of the
four or ve individual VaRs, i.e., the so-called undiversied VaR,
minus the diversication effect. We see that for all banks/quarters,
the undiversied VaR exceeds the diversied one, which shows
that diversication among broad sources of risk plays a central role
in determining the bank disclosed VaR.
We display in Fig. 2 a measure of the actual banks diversica-
tion effect as dened in Eq. (3). We clearly see that the level of
diversication acknowledged by the sample banks in their VaR cal-
culation is not trivial. Indeed, the average diversication coefcient
is 41.2% and rm-specic values range from 33.9% for HSBC to
54.6% for Bank of America. The latter gure suggests that ignoring
diversication would lead Bank of America to double both its DVaR
and its market risk charge. At times, the diversication acknowl-
edged by HSBC is clearly smaller than those of other sample banks.
For instance, during three quarters in 2004, the HSBCs diversica-
tion coefcient was between 4% and 12%. We also note that the
strength of the diversication effect varies through time and that
all sample banks have at least one quarter with a diversication
coefcient in excess of 50%.
7
5
Numbers of exceptions per year have been obtained from rms nancial
statements.
6
The interest rate VaR of HSBC is an exception. For instance, between 2005-Q4 and
2006-Q1, the 10-day interest rate VaR jumped from 22 million to 143 million (the
corresponding 1-day VaR jumped from 7 million to 45.2 million).
7
There are two reasons for the diversication coefcient to vary through time: (1)
correlations between broad risk categories are time-varying and (2) the composition
of the trading portfolio changes from one quarter to the next.
C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566 57
For comparison purposes, we conduct a similar analysis using
data from ve major Canadian banks for the year 2005.
8
This alter-
native set of banks has been shown to be extremely conservative in
setting their VaR (see Prignon et al., 2008).
9
Interestingly, we nd a
very similar average diversication coefcient (40.0%) and the range
across Canadian rms is 23.649.5%.
While the diversication effects acknowledged by US banks ini-
tially appear to be quite large, we still need to investigate whether
they are large enough. To answer this question we construct a
benchmark diversication coefcient from individual VaRs and
risk category indices.
10
Our risk proxies are as follows: S&P500
Composite Index for equity; 1-Year Treasury Constant Maturity Rate
for interest rate; Dow Jones AIG Spot Commodity Index for commod-
ity; the difference between Moodys BBA corporate yield and
1-year Treasury yield for credit spread; and Trade-Weighted Major
JPMorgan Chase : VaRi
0
50
100
150
200
Credit spread
Commodity
Foreign exchange
Equity
Interest rate
JPMorgan Chase : DVaR
0
50
100
150
200
Citigroup : VaRi
0
50
100
150
200
250
Commodity
Foreign exchange
Equity
Interest rate
HSBC : VaRi
0
20
40
60
Credit spread
Commodity
Foreign exchange
Equity
Interest rate
Bank of America : VaRi
0
50
100
150
Credit spread
Commodity
Foreign exchange
Equity
Interest rate
Citigroup : DVaR
0
50
100
150
200
250
HSBC : DVaR
0
20
40
60
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1 01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1 01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
Bank of America : DVaR
0
50
100
150
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
Fig. 1. Individual and diversied Value-at-Risk. Note: This gure displays the quarterly individual VaRs (VaR
i
, left panels) and the quarterly diversied VaR (DVaR, right
panels) disclosed by our sample banks. All gures are in million of US dollars.
8
Canadian data on individual and diversied VaRs come from Table 1 in Prignon
et al. (2008).
9
Several institutional reasons could potentially explain the extreme cautiousness
of the Canadian banks when setting their VaR. First, this could be due to the
competitive environment of the Canadian banking industry. Indeed, while the top ve
US banks in 1999 accounted for just 21% of US deposits, the top ve banks in Canada
accounted for 76% of Canadian deposits. A second reason is the fact that US banks are
mainly exposed to US interest rate risk (see our Fig. 1) while Canadian banks clearly
must have large exposures to currency risk in their proprietary trading desks.
10
Berkowitz and OBrien (2007) also use risk category indices to study the exposure
of banks trading revenues to major sources of risk.
58 C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566
Currencies Dollar index for FX.
11
All risk indices are available at the
daily frequency over the period January 1991March 2007. We pro-
vide summary statistics for all risk indices in Table 1. We compute
the rst four moments of the log daily returns, as well as the Bera-
Jarque normality test. We display in the lower part of Table 1, the
unconditional correlation matrix of the returns of the ve risk
indices.
At the end of each quarter, we estimate the correlation matrix,
which is used to compute the empirical VaR for this quarter (see
Eq. (4)). For now we assume that all sample banks have a long
net position in each risk category.
12
The rst estimate of R is the his-
torical correlation matrix based on the past 250 daily observations
and the second is given by the DCC model. We display the parameter
estimates and the standard errors in Table 2. For each correlation
estimate, we estimate the empirical diversication coefcient d
E
as
in Eq. (8) and we plot the value of the diversication differential
d
E
d in Fig. 3 (unconditional correlation-based results are in the left
panels and DCC-based results are in the right panels). We note that
our computation of the diversication coefcient applies for the
wide multivariate elliptical density family and we have estimated
the DCC model parameters by quasi-maximum likelihood, which
JPMorgan Chase : Diversification
0%
20%
40%
60%
80%
100%
Citigroup : Diversification
0%
20%
40%
60%
80%
100%
HSBC : Diversification
0%
20%
40%
60%
80%
100%
Bank of America : Diversification
0%
20%
40%
60%
80%
100%
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
Fig. 2. Actual diversication. Note: This gure displays the actual diversication coefcient (in percentage) estimated from banks individual and diversied VaRs. For each
sample bank and each quarter, it is given by one minus the ratio of the diversied VaR and the sum of the individual VaRs.
Table 1
Summary statistics.
Interest rate Equity FX Commodity Credit spread
Mean 0.0077 0.0357 0.0013 0.0274 0.0216
Standard deviation 1.5221 0.9845 0.4032 0.8019 1.4085
Skewness 0.1236 0.0770 0.0347 0.3234 0.6434
Kurtosis 15.339 7.127 4.317 8.526 9.095
Bera-Jarque 40,402.2 8,722.5 3,199.9 12,551.9 14,484.4
p-Value (0.000) (0.000) (0.000) (0.000) (0.000)
Correlation matrix
Interest rate 1
Equity 0.1107 1
FX 0.1613 0.0701 1
Commodity 0.0343 0.0082 0.1299 1
Credit spreads 0.3417 0.1093 0.0814 0.0396 1
Notes: This table presents summary statistics for the daily log returns of the ve risk category indices: mean, standard deviation, skewness, kurtosis, Bera-Jarque normality
test, with its associated p-value. All risk indices are available at the daily frequency over the period January 1991March 2007. The total number of observations is 4120 for
each index. The lower panel of the table displays the correlation matrix among the returns of the ve risk category indices.
11
These indices were selected because they are popular indices and also because
they were available with a daily frequency over a long sample period. The sources of
the data are CRSP for equity, the Federal Reserve Economic Data (FRED) database for
interest rate, http://www.djindexes.com for commodity, http://www.globalnda-
ta.com for credit spread, and the Federal Reserves website for FX. The Dow Jones AIG
Spot Commodity Index is a weighted-average of the spot prices of 19 commodities.
The Trade-Weighted Major Currencies Dollar index is a weighted average of the
foreign exchange values of the US dollar against a subset of currencies that circulate
widely outside the country of issue. The index weights are derived from US export
shares and from US and foreign import shares.
12
To get the value of the basket of foreign currencies in US dollars, we multiply the
FX index by 1. We will relax the assumption that all positions are long in Section 4.2.
C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566 59
provides consistent estimates for a broad class of return densities. In
order to interpret these results one needs to recall that a positive
diversication differential would imply that the bank underesti-
mates the diversication effect.
The most important results in Fig. 3 are as follows. First, our two
correlation estimation strategies provide consistent results. Second
and most importantly, there is no systematic underestimation of
the diversication effect in our sample. The diversication differ-
ential is modest for most sample banks, sometimes being positive
and sometimes negative. There are two notable exceptions though.
The rst one concerns the diversication assumed by Bank of
America between 2001 and 2004. During this period, there is a
negative differential which suggests that the bank exaggerated
its diversication effect. This nding clearly contradicts the diver-
sication hypothesis. Another interesting episode covers the quar-
ters between 2004-Q1 and 2005-Q1 during which HSBC virtually
neglected the correlations among risk categories (as already noted
in Fig. 2).
We note that the gap between the actual and empirical diversi-
cation measures is smaller for Citigroup and JPMorgan Chase than
for Bank of America and HSBC. Indeed the average absolute differ-
ence between d and d
E
(computed from unconditional correlations)
is 5.7% and 5.2% for the rst two banks and 9.5% and 13.2% for the
later two. For each bank and each quarter, we compute the ratio of
Trading Assets to Total Assets from their consolidated nancial
statements (FR Y-9C form). It turns out that these two groups of
banks differ dramatically in terms of trading activity. Trading as-
sets constitute between 20% and 30% of the total assets for Citi-
group and JPMorgan Chase but only between 10% and 13% for
Bank of America and HSBC, which is close to the average level in
the US (see Hirtle, 2007). Although anecdotal, this evidence sug-
gests that banks with large trading operations are particularly
careful when handling the correlation structure of their trading
portfolio.
The main nding of our empirical analysis is that VaR over-
reporting does not seem to be consistently due to a partial diversi-
cation adjustment. An important implication of our main result is
that individual VaRs are inated or, in other words, that banks dis-
closures do not reect accurately the exposures of their trading
books with respect to the main sources of risk in the economy.
4. Robustness checks and extensions
4.1. Risk index selection
We rst check the sensitivity of out empirical results to the
choice of the risk proxies. To do so, we nd for each source of risk
an alternative daily time series. The new indices are the follow-
ing: for equity, Dow Jones Wilshire 5000 Composite (source:
Datastream); for interest rate, the 3-year treasury yield (source:
FRED); for commodity, the S&P GSCI Commodity Total Return
(source: Datastream); for credit spread, the difference between
the yield on BAA 20-year corporate bonds and the 20-year trea-
sury yield (source: FRED); and for FX, the broad Trade-Weighted
Exchange Index for the value of the USD computed by the Board
of Governors (source: Federal Reserves website). The equity,
interest rate, and commodity indices are available over the entire
sample period (January 1991March 2007), whereas the FX and
credit spread series started in January 1995 and October 1993,
respectively.
As a robustness check for our choice of risk proxies, we vary
each one-by-one and re-estimate the DCC model with the one
new risk proxy using the longest common sample period. We com-
pute the diversication coefcient for each bank and compare it to
its value computed using the original dataset. We nd that the
average absolute difference in the percentage diversication re-
mains always below 5% for all bank/index. For instance, the aver-
age absolute difference in the diversication coefcient for
JPMorgan Chase is 3.06% for interest rate, 2.18% for equity, 2.34%
for FX, 0.79% for commodity, and 1.74% for credit spread. Given
the stability of the results, we will be using the original sample
in the rest of the study.
4.2. Allowing for short positions
Thus far, we have assumed that banks have long net positions in
all asset classes. However it is conceivable that some net positions
could be short. We re-run our analysis assuming that our sample
banks have a short net position in credit spread (left panels of
Fig. 4) or in FX (right panels of Fig. 4). Note that introducing a short
position is simply done by multiplying by 1 the risk index used in
Table 2
DCC parameter estimates.
Interest rate Equity FX Commodity Credit spread
x 100 0.8554 0.5219 0.0916 0.2461 2.0170
s.e. (0.4016) (0.2704) (0.1408) (0.0973) (1.0984)
a 0.0648 0.0522 0.0350 0.0571 0.0894
s.e. (0.0186) (0.0105) (0.0127) (0.0111) (0.0268)
b 0.9315 0.9435 0.9602 0.9423 0.9056
s.e. (0.0187) (0.0115) (0.0206) (0.0109) (0.0282)
Correlation matrix (R)
Interest rate 1.0139 0.0151 0.1392 0.0512 0.4263
Equity (0.0229) 0.9927 0.0652 0.0130 0.1026
FX (0.0175) (0.0174) 0.9935 0.1228 0.0782
Commodity (0.0162) (0.0161) (0.0164) 0.9853 0.0228
Credit spreads (0.0190) (0.0193) (0.0172) (0.0157) 0.9984
Parameter s.e.
h
1
0.0162 (0.0018)
h
2
0.9789 (0.0027)
Notes: This table presents the coefcient estimates for the Dynamic Conditional Correlation. The returns are assumed to be multivariate normal with a conditional covariance
matrix H
t
= E(et e
/
t
) = D
t
R
t
D
t
, where D
t
is a diagonal matrix of conditional standard deviations that are modeled as univariate GARCH models: D
2
t
=
diag(x
i
) diag(a
i
) (r
t1
r
/
t1
) diag(b
i
) D
2
t1
, where denotes the Hadamard multiplier, and R
t
is the conditional correlation matrix. The correlation matrix is expressed as
R
t
= diag(Q
t
)
1/2
Q
t
diag(Q
t
)
1/2
and the matrix Q
t
is modeled as a GARCH-type function of the lagged standardized residuals zt = D
1
t
rt that depends on only two scalar
parameters h
1
and h
2
: Q
t
= R(1 h
1
h
2
) + h
1
(z
t1
z
t1
) + h
2
Q
t1
and R is the unconditional sample correlation matrix of the standardized residuals. The parameters that drive
the conditional variance dynamics are estimated by maximizing the N univariate GARCH log-likelihoods and then conditioning on the volatility dynamics estimate the
correlation parameters. Robust standard errors (s.e.) are displayed in parenthesis (for the correlation coefcients standard error is reported below the principal diagonal).
60 C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566
the all-long case. Interestingly, we nd that our main conclusion
is robust to these alternative weighting schemes. Indeed, the over-
all pattern of d
E
d is generally preserved for both the historical
correlation and DCC approaches. Of particular interest are the re-
sults for Bank of America since the trading portfolio of this bank
is particularly exposed to credit spread (see Fig. 1). For this bank,
d is always larger than d
E
when the credit spread position is as-
sumed to be short, which clearly contradicts the diversication
hypothesis.
4.3. BEKK model
Another robustness check is to consider an alternative econo-
metric model for the correlation structure among risk indices.
The BEKK model of Baba et al. (1990) and Engle and Kroner
(1995) was introduced as an easy way to ensure the covariance
matrices are well dened. The covariance dynamics is:
H
t
= CC
/
A(e
t1
e
/
t1
)A
/
BH
t1
B
/
; (11)
JPMorgan Chase : E -
-50%
-30%
-10%
10%
30%
50%
Citigroup : E -
-50%
-30%
-10%
10%
30%
50%
HSBC : E -
-50%
-30%
-10%
10%
30%
50%
Bank of America : E -
-50%
-30%
-10%
10%
30%
50%
JPMorgan Chase : DCC E -
-50%
-30%
-10%
10%
30%
50%
Citigroup : DCC E -
-50%
-30%
-10%
10%
30%
50%
HSBC : DCC E -
-50%
-30%
-10%
10%
30%
50%
Bank of America : DCC E -
-50%
-30%
-10%
10%
30%
50%
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
Fig. 3. Comparing actual and empirical diversication. Note: This gure displays the difference between d the actual diversication coefcient (shown in Fig. 2) and d
E
an
empirical estimate of the diversication coefcient estimated from banks individual VaRs and market indices. For each sample bank and each quarter, the empirical
diversication coefcient is given by one minus the ratio of the empirical VaR and the sum of the banks individual VaRs. In the left panels, we estimate the empirical VaR
using an historical estimate of the correlation matrix among risk category indices. In the right panels, we estimate the empirical VaR tting the Dynamic Conditional
Correlation model of Engle (2002) to the risk category indices.
C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566 61
which is guaranteed to be positive semi-denite by construction.
We assume that the matrices A and B are diagonal to make estima-
tion feasible. The parameters of the model are estimated by quasi-
maximum likelihood and reported in Table 3. For each bank/quar-
ter, we estimate the empirical diversication coefcient d
E
using
the results from the BEKK model. Then we plot in the left panels
of Fig. 5 the value of the diversication differential d
E
d for all
banks. Our main conclusion is not materially altered by the use of
this alternative covariance dynamics.
4.4. Time-varying copula model
Our empirical results to date use a very simple approach to
compute the diversication effect in Eq. (4) using only the correla-
tion matrix. Unfortunately this elegant approach is only valid if re-
turns are from the elliptical family of distributions such as the
multivariate normal or t. This assumption is limiting, as the multi-
variate elliptical family requires that the marginal standardized
densities for all assets are identical. For example, if the vector of
JPMorgan Chase : DCC E -
(Credit Spread SHORT)
-50%
-30%
-10%
10%
30%
50%
Citigroup : DCC E -
(Credit Spread SHORT)
-50%
-30%
-10%
10%
30%
50%
HSBC : DCC E -
(Credit Spread SHORT)
-50%
-30%
-10%
10%
30%
50%
Bank of America : DCC E -
(Credit Spread SHORT)
-50%
-30%
-10%
10%
30%
50%
JPMorgan Chase : DCC E -
(FX SHORT)
-50%
-30%
-10%
10%
30%
50%
Citigroup : DCC E -
(FX SHORT)
-50%
-30%
-10%
10%
30%
50%
HSBC : DCC E - (FX SHORT)
-50%
-30%
-10%
10%
30%
50%
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
Bank of America : DCC E -
(FX SHORT)
-50%
-30%
-10%
10%
30%
50%
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1 01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
Fig. 4. Comparing actual and empirical diversication with short positions. Note: This gure displays the difference between d the actual diversication coefcient (shown in
Fig. 2) and DCC d
E
an empirical estimate of the diversication coefcient estimated from banks individual VaRs and market indices. For each sample bank and each quarter,
the empirical diversication coefcient is given by one minus the ratio of the empirical VaR and the sum of the banks individual VaRs. In all panels, we estimate the empirical
VaR tting the Dynamic Conditional Correlation model of Engle (2002) to the risk category indices. In the left (respectively right) panels, we assume that the bank net
aggregate position in credit spread (FX) is short, whereas all other net aggregate positions are assumed to be long.
62 C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566
returns r is multivariate t (which is a member of the elliptical fam-
ily) then all the individual returns must have the same degrees of
freedom. Rather than restrict ourselves to the elliptical family we
can model the joint distribution of returns using the copula ap-
proach. A copula is a function that enables N different marginal
densities to be coupled together to form a multivariate density:
F(r
1
; r
2
; . . . ; r
n
) = C(F
1
(r
1
); . . . ; F
n
(r
n
)); (12)
where the function C is a cumulative density function with uniform
marginals. Similarly, and most useful to our application, a density
can be expressed as a function of the marginals and the density cop-
ula c:
f (r
1
; r
2
; . . . ; r
n
) =

n
i=1
f
i
(r
i
)
_ _
c(F
1
(r
1
); . . . ; F
n
(r
n
)): (13)
The multivariate Gaussian Copula is dened as:
c(u
1
; . . . ; u
n
; R) =
/
n
(U
1
(u
1
); . . . ; U
1
(u
n
); R)

n
i=1
/(U
1
(u
i
))
; (14)
where /
n
(z
1
; . . . ; z
n
; R) denotes the probability function of an n-
dimensional multivariate normal random vector with correlation
matrix R and z
it
= U
1
(u
it
) = U
1
(F
i
(r
it
)) is the normal quantile of
the probability integral transform of model i, and can be thought
of as a pseudo-standardized residual. If the conditional density is
correctly specied the probability integral F
i
(r
it
) will be indepen-
dent uniform random variables, and then z
it
will be independent
standard normal random variables. The key benet here is that
the marginal densities f
i
(and hence marginal distributions F
i
) can
be specied independent of each other. For example, we can allow
for symmetric Student t distributions with different degrees of free-
dom or even allow, as we do in the sequel, for asymmetry (see Pat-
ton, 2006; Jondeau and Rockinger, 2006; Bali et al., 2008; Bartram
et al., 2007). As in Section 2, we model the dynamics of R
t
using
the correlation dynamics used in the DCC model of Engle
(2002).
13
The difculty we now face is that the distribution of a port-
folio is not known in closed form, which complicates computing the
DVaR. Recall that assuming zero expected returns we can express the
VaR as in Eq. (1) as VaR
i
= j
i
r
i
x
i
though in our case the scaling coef-
cient j
i
will vary across risk indices. We can recover x
i
from the
conditional variance and shape parameters of the conditional distri-
bution along with the VaR
i
, giving:
x
i
=
VaR
i
r
i
j
i
; (15)
where r
i
j
i
is the VaR of a $1 investment in index i. We can then
compute that implied DVaR by simulating the returns of a portfolio
with investment weights x
i
for i = 1, . . . , N. To do this, we take the
conditional correlation coefcient R
t
describing the shape of the
conditionally normal copula and generate the N-vector of returns:
y
(k)
t
= R
1=2
t
z
(k)
(16)
for i.i.d. standard normal z, and R
1=2
t
denotes the Cholesky factoriza-
tion of the correlation matrix. We then compute the percentiles
u
(k)
it
= U(y
(k)
it
) and simulate the portfolio returns by computing the
conditional percentiles based on the empirical model for the mar-
ginal distributions r
(k)
it
= F
1
it1
(y
(k)
it
), where F
1
it1
denotes the inverse
cumulative density function or quantile function conditional on
information available at time t-1 for risk index i. We then compute
the hypothetical return on the portfolio as:
r
(k)
pt
= x
/
t
r
(k)
t
; (17)
where r
(k)
t
= r
(k)
it
; i = 1; . . . ; n is the vector of index returns. The VaR
is computed as the empirical quantile across the simulated hypo-
thetical portfolio returns.
We model the marginal distribution of each factor return as a
GARCH(1, 1) process where the standardized distributions are ta-
ken from the Skewed-t distribution of Hansen (1994):
h
it
= x
i
a
i
e
2
i;t1
b
i
h
i;t1
; (18)
where z
it
= e
it
=

h
it
_
- Skewt(k; g) and 1 6 k 6 1 is the asymme-
try parameter and 2 < g denotes the degrees of freedom. The prob-
ability density function of the Skewed-t random variable is given
by:
f (z; k; g) =
bc 1
1
g2
bza
1k
_ _
2
_ _
(g1)=2
; z < a=b;
bc 1
1
g2
bza
1k
_ _
2
_ _
(g1)=2
; z _ a=b;
_

_
(19)
where
a = 4kc
g 2
g 1
_ _
; b
2
= 1 3k
2
a
2
; c =
C
g1
2
_ _

p(g 2)
_
C
g
2
_ _ :
The distribution can generate both skewness and kurtosis and
the symmetric Student t distribution when k = 0 and nests the nor-
mal distribution when g
1
= 0. Because we have used the Gaussian
copula we nest the DCC model when g
1
= 0.
Table 3
BEKK parameter estimates.
Interest rate Equity FX Commodity Credit spread
A 0.2270 0.1690 0.1668 0.2428 0.2303
s.e. (0.0308) (0.0609) (0.0204) (0.0255) (0.0358)
B 0.9718 0.9841 0.9696 0.9478 0.9709
s.e. (0.0075) (0.0121) (0.0093) (0.0101) (0.0090)
(CC
/
) * 100
Interest rate 1.0174
(0.4405)
Equity 0.0589 0.3299
(0.0779) (0.4015)
FX 0.1339 0.0575 0.5206
(0.0568) (0.0380) (0.2058)
Commodity 0.1929 0.0572 0.1027 2.8271
(0.1042) (0.0699) (0.0495) (0.8303)
Credit spreads 0.7722 0.0801 0.0451 0.1524 1.2988
(0.3558) (0.0778) (0.0401) (0.1091) (0.6324)
Notes: This table presents the coefcient estimates for the BEKK model. The covariance dynamics is Ht = CC
/
A(e
t1
e
/
t1
)A
/
BH
t1
B
/
, where A and B are diagonal matrices
and CC
/
is a symmetric matrix. The model is estimated by quasi-maximum likelihood. The robust standard errors (s.e.) are displayed in parenthesis.
13
As a robustness check, we have constructed the correlation matrix on the basis of
Pearsons correlation, pair-wise Spearmans correlation, and Kendalls tau. The results
were not materially affected.
C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566 63
Table 4 reports the parameter estimates. The dynamics of both
the conditional volatility and the correlation matrix of the pseudo-
standardized residuals are quite similar in both the Copula and the
DCC models (though volatility in the Copula model responds
slightly less to innovations in returns than in the DCC model). This
suggests that accounting for skewness and excess kurtosis does not
materially affect the volatility dynamics. We also note that there is
clear variation in degrees of freedom across the risk indices. Then,
we plot in the right panels of Fig. 5 the value of the diversication
differential d
E
d. The diversication effect obtained using Copula
model is very similar to the two multivariate GARCH-based esti-
mates, and provides no support for the claim that banks systemat-
ically understate the diversication effect.
5. Conclusion
A reason often put forward in the literature to explain banks
VaR overstatement is a systematic underestimation of the diversi-
cation among broad risk categories (see Berkowitz and OBrien,
2002; Prignon et al., 2008). An implication of the diversication
hypothesis is that, although aggregate VaRs are inated risk mea-
sures, individual VaRs may be unbiased and as a result, are useful
JPMorgan Chase : BEKK E -
-50%
-30%
-10%
10%
30%
50%
Citigroup : BEKK E -
-50%
-30%
-10%
10%
30%
50%
HSBC : BEKK E -
-50%
-30%
-10%
10%
30%
50%
Bank of America : BEKK E -
-50%
-30%
-10%
10%
30%
50%
JPMorgan Chase : Copula E -
-50%
-30%
-10%
10%
30%
50%
Citigroup : Copula E -
-50%
-30%
-10%
10%
30%
50%
HSBC : Copula E -
-50%
-30%
-10%
10%
30%
50%
Bank of America : Copula E -
-50%
-30%
-10%
10%
30%
50%
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
01Q4 02Q3 03Q2 04Q1 04Q4 05Q3 06Q2 07Q1
Fig. 5. Comparing actual and empirical diversication (BEKK and Copulas). Note: This gure displays the difference between d the actual diversication coefcient (shown in
Fig. 2) and d
E
an empirical estimate of the diversication coefcient estimated from banks individual VaRs and market indices. For each sample bank and each quarter, the
empirical diversication coefcient is given by one minus the ratio of the empirical VaR and the sum of the banks individual VaRs. In the left panels, we estimate the
empirical VaR tting a BEKK model of Engle and Kroner (1995) to the risk category indices. In the right panels, we estimate the empirical VaR using a copula-based model.
64 C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566
to market participants. Contrary to the diversication hypothesis,
we nd that commercial banks make signicant adjustments to
their VaR to reect diversication effects, i.e., the average diversi-
cation effect is around 40%. Furthermore, we show that our sam-
ple banks handle their correlation structure quite well. This
conclusion is robust to the use of alternative risk proxies, correla-
tion models, and distributional assumptions. A consequence of our
nding is that individual VaRs, just like aggregate VaRs, are biased
risk assessments.
Although we study the aggregation process of risk-category
individual VaRs into a single rm-level VaR, an alternative strategy
would be to break-down VaR by trading line and to study the
diversication effect among them. Berkowitz et al. (forthcoming)
propose a promising attempt along this line.
Acknowledgements
This paper was originally submitted to Professor Giorgio Szego
on September 20, 2007 and was revised once prior to submission
through EES. We thank two anonymous referees for their com-
ments and Jiawen Lin and Yannan Qin for their help in constructing
the dataset. Part of this paper was written while Daniel Smith was
visiting the Queensland University of Technology. We are grateful
to the Social Sciences and Humanities Research Council of Canada
for their nancial support.
Appendix A. Portfolio VaR with elliptical distributions
The density of an elliptical distribution, if it exists, can be writ-
ten as a scalar function of the quadratic form (X l)
/
W
1
(X l),
and in particular:
f (X) = K[W[
1=2
g((X l)
/
W
1
(X l)); (A1)
where K is a normalizing constant and the scalar function
g : R

#R

. As the name suggests, the equi-density contours of

these density functions are ellipsoids. Prominent examples of this
class of distributions are the multivariate normal and t distributions.
In the elliptical class, the mean and covariance matrix (if they
exist) are given by E(X) = l and V(x) = aW for some distribution
specic constant scalar a. For instance, for the multivariate normal
distribution we have a = 1 and for the multivariate t distribution
with g degrees of freedom we have a = (g 2)/g. If X is elliptical
then both linear transformations and the marginals are afne
too: if X - e(l; W) then Y = WX - e(Wl; WWW
/
), where W is an
mn matrix of rank m6 n. One useful implication of this result
is that each of the marginal distributions has the same distribution
as the marginal (e.g., if X is multivariate t with g degrees of free-
dom then all of the marginals are univariate t with g degrees of
freedom), which can be seen by setting W = e
i
and e
i
is the ith col-
umn of an n-dimensional identity matrix. This class of distributions
is also a member of the location-scale family, so the individual
VaRs can be expressed (assuming zero expected trading revenues)
as afne functions of the square root of the diagonal elements of W.
Given that W= cV, where c = 1/a is a density-dependent constant,
then we can write it as W= cDRD and the entire analysis leading
to Eq. (2) goes through.
We can go one step further and point out that as long as the
conditional covariance matrix and mean are correctly specied,
then maximizing the Gaussian log-likelihood as we do above will
render consistent estimates regardless of the true distribution.
Assuming the banks compute their VaR using the implied model
then we can account for the diversication effect using the Gauss-
ian-based conditional correlation coefcient and this will be
appropriate for any of the elliptical class of densities. This is partic-
ularly important given the popularity of the multivariate t density
for asset pricing tests and portfolio selection (see Owen and Rabi-
novitch, 1983).
References
Baba, Y., Engle, R.F., Kraft, D., Kroner, F., 1990. Multivariate Simultaneous
Generalized ARCH. Working Paper, University of California, San Diego.
Bakshi, G., Panayotov, G., 2007. The Capital Adequacy Puzzle. Working Paper,
University of Maryland.
Bali, T.G., Mo, H., Tang, Y., 2008. The role of autoregressive conditional skewness and
kurtosis in the estimation of conditional VaR. Journal of Banking and Finance 32,
269282.
Bartram, S., Taylor, S.J., Wang, Y., 2007. The Euro and European nancial market
dependence. Journal of Banking and Finance 31, 14611481.
Basel Committee on Banking Supervision, 1996. Amendment to the Capital Accord
to Incorporate Market Risks. Bank for International Settlements.
Berkowitz, J., Christoffersen, P.F., Pelletier, D., forthcoming. Evaluating Value-at-Risk
models with desk-level data. Management Science.
Table 4
Copula parameter estimates.
Interest rate Equity FX Commodity Credit spread
x 100 0.3696 0.3066 0.0525 0.2380 1.7603
s.e. (0.1680) (0.1394) (0.0471) (0.0920) (0.7730)
a 0.0397 0.0481 0.0381 0.0522 0.0783
s.e. (0.0070) (0.0089) (0.0088) (0.0112) (0.0213)
b 0.9575 0.9499 0.9602 0.9465 0.9167
s.e. (0.0074) (0.0094) (0.0100) (0.0109) (0.0223)
~
k
a
0.0312 0.0747 0.0794 0.0715 0.0777
s.e. (0.0335) (0.0381) (0.0415) (0.0439) (0.0416)
g
1
0.2163 0.1386 0.1309 0.1048 0.1735
s.e. (0.0126) (0.0155) (0.0144) (0.0148) (0.0148)
Correlation matrix
Interest rate 1.0279 0.0291 0.1356 0.0567 0.3961
Equity (0.0183) 0.9961 0.0704 0.0113 0.0902
FX (0.0170) (0.0166) 0.9927 0.1205 0.0763
Commodity (0.0161) (0.0159) (0.0161) 0.9848 0.022
Credit spreads (0.0174) (0.0169) (0.0165) (0.0157) 0.9967
Parameter s.e.
h
1
0.0160 (0.0015)
h
2
0.9801 (0.0021)
Notes: This table presents the coefcient estimates for the copula-based model.
a
To ensure 1 < k < 1, we estimate a monotonic transformation that is dened over the whole real line,
~
k = log[(1 k)=(1 k)[, where
~
k is dened as the extended logistic
transformation of k. The robust standard errors (s.e.) are displayed in parenthesis. For the correlation coefcients, standard errors are reported below the principal diagonal.
C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566 65
Berkowitz, J., OBrien, J., 2002. How accurate are value-at-risk models at commercial
banks? Journal of Finance 57, 10931111.
Berkowitz, J., OBrien, J., 2007. Estimating bank trading risk: A factor model
approach. In: Carey, M., Stulz, R.M. (Eds.), The Risk of Financial Institutions.
University of Chicago Press, Chicago.
Breuer, T., Jandacka, M., Rheinberger, K., Summer, M., forthcoming. Does adding up
of economic capital for market- and credit risk amount to conservative risk
assessment? Journal of Banking and Finance.
Danielsson, J., Zigrand, J., 2006. On time-scaling of risk and the square-root-of-time
rule. Journal of Banking and Finance 30, 27012713.
Engle, R.F., Kroner, K.F., 1995. Multivariate simultaneous generalized ARCH.
Econometric Theory 11, 122150.
Engle, R.F., 2002. Dynamic conditional correlation a simple class of
multivariate GARCH models. Journal of Business and Economic Statistics
20, 339350.
Hansen, B.E., 1994. Autoregressive conditional density estimation. International
Economic Review 35, 705730.
Hirtle, B., 2003. What market risk capital reporting tells us about bank risk. FRBNY
Economic Policy Review (September), 3754.
Hirtle, B., 2007. Public Disclosure, Risk, and Performance at Bank Holding
Companies. Working Paper, Federal Reserve Bank of New York.
Jondeau, E., Rockinger, M., 2006. The copula-GARCH model of conditional
dependencies: An international stock market application. Journal of
International Money and Finance 25, 827853.
Jorion, P., 2002. How informative are value-at-risk disclosures? Accounting Review
77, 911931.
Jorion, P., 2006. Value at Risk: The New Benchmark for Managing Financial Risk,
third ed. McGraw-Hill.
Owen, J., Rabinovitch, R., 1983. On the class of elliptical distribution and their
applications to the theory of portfolio choice. Journal of Finance 38, 745752.
Patton, A.J., 2006. Modelling asymmetric exchange rate dependence. International
Economic Review 47, 527556.
Prignon, C., Deng, Z.Y., Wang, Z.Y., 2008. Do banks overstate their value-at-risk?
Journal of Banking and Finance 32, 783794.
Prignon, C., Smith, D.R., forthcoming. The level and quality of Value-at-Risk
disclosure by commercial banks. Journal of Banking and Finance.
Pritsker, M., 2006. The hidden dangers of historical simulation. Journal of Banking
and Finance 30, 561582.
66 C. Prignon, D.R. Smith/ Journal of Banking & Finance 34 (2010) 5566