Hyung 2005

Journal of Financial Econometrics, 2005, Vol. 3, No.
1, 107–125
Portfolio Diversification Effects

of Downside Risk
Namwon Hyung
University of Seoul and Tinbergen Institute
Casper G. de Vries
Erasmus University Rotterdam
Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

abstract
Risk managers use portfolios to diversify away the unpriced risk of individual
securities. In this article we compare the benefits of portfolio diversification for
downside risk in case returns are normally distributed with the case of fat-tailed
distributed returns. The downside risk of a security is decomposed into a part
which is attributable to the market risk, an idiosyncratic part, and a second
independent factor. We show that the fat-tailed-based downside risk, measured
as value-at-risk (VaR), should decline more rapidly than the normal-based VaR.
This result is confirmed empirically.
keywords: diversification, portfolio decomposition, value-at-risk
securities. This topic has been well studied for global risk measures like the
variance [see, e.g., the textbook by Elton and Gruber (1995, chap.4)]. In this article
we study the benefits of portfolio diversification with respect to an extreme
downside risk measure known as the zeroth lower partial moment and its
inverse; where the inverse of the zeroth lower partial moment is better known
as the value-at-risk VaR measure. Choice theoretic considerations for this risk
measure are offered in Arzac and Bawa’s (1977) analysis of the safety-first criter-
ion. In Gourieroux, Laurent, and Scaillet (2000), the implications under the
assumption of normally distributed returns are investigated, while Jansen,
Koedijk, and de Vries (2000) implement the safety-first criterion for heavy-tailed
distributed returns. There is some concern in the literature that the VaR measure
lacks subadditivity as a global risk measure. As a measure for the downside risk,
N. Hyung benefited from a research grant from the Tinbergen Institute. We are very grateful for the
insightful comments of two referees, the editors, and Thierry Post. Address correspondence to Casper G.
de Vries, Department of Accounting and Finance, H14-25, Erasmus University Rotterdam, P.O. Box 1738,
3000 DR, Rotterdam, The Netherlands, or e-mail: [email protected].
doi:10.1093/jjfinec/nbi004
Journal of Financial Econometrics, Vol. 3, No. 1, ª Oxford University Press 2005; all rights reserved.
108 Journal of Financial Econometrics
however, the VaR exhibits subadditivity if one evaluates this criterion sufficiently
deep in the tail area.1
The portfolio diversification effects for the downside risk are evaluated in
terms of the diversification speed. The diversification speed is measured in two
different ways. Let the ‘‘VaR diversification speed’’ be the rate at which the VaR
changes as the number of assets k included in the portfolio increases. Usually the
safety-first criterion and the VaR criterion are evaluated at a fixed probability level.
It is also possible to do the converse analysis by fixing the VaR level and letting the
probability level change as the number of assets k increases. This gives what we
term the ‘‘diversification speed of the risk level.’’ We will study both concepts.

Much of the theoretical literature in finance presumes that the returns are
normally distributed. For a host of questions, this is a reasonable assumption to
make. Empirically it is well known that the return distributions have fatter tails
than the normal [see, e.g., Jansen and De Vries (1991)]. For the downside risk
measures, this data feature turns out to make a crucial difference. The diversifica-
tion speeds are shown to be quite different for the cases of the normal and the fat-
tailed distributions. The VaR diversification speed is higher for the class of (finite
variance) fat-tailed distributions in comparison to the normal distribution, but is
lower with respect to the diversification speed of the risk level. The intuition for
this result is as follows. Start with the latter result. The tails of the normal density
go down exponentially fast, while the tails of fat-tailed distributions decline at a
power rate (this is the defining characteristic of these distributions). Since an
exponential function eventually beats any power, it stands to reason that the
diversification speed of the risk level under normality is larger. The VaR diversi-
fication speed measures the speed in terms of quantiles, which are the inverse of
the probabilities. Taking the inverse reverses the diversification speed.
Consider, for example, the case of the normal versus Student’s t-distributed
returns with v degrees of freedom. It is well known that the VaR diversification
speed for the normal distribution follows the square root rule. In contrast, the
Student’s t VaR diversification speed is 1 1=v. This is greater than 1/2 if v > 2
(guaranteeing a finite variance). This intuition is made rigorous below by means
of the celebrated Feller convolution theorem for heavy-tailed (i.e., regularly
varying) distributions.
For the empirical counterpart of this analysis, we briefly review the semipara-
metric approach to estimating the (extreme) downside risk. The heavy-tail feature is
captured by a Pareto distribution-like term, of which one needs to estimate the tail
index (the equivalent of the degrees of freedom v in case of Student’s law) and a
scale coefficient. We consider estimation by means of a pooled dataset on the basis
of the assumption that the tail indexes of the different securities and risk compo-
nents are equal. We do allow for heterogeneity of the scale coefficients, though.
Most securities’ distributions display equal hyperbolic tail coefficients, but do differ
considerably in terms of their scale coefficients [see Hyung and de Vries (2002)].
1
At least this holds for the normal distribution and the class of fat-tailed distributions investigated in this
article.
Hyung & de Vries | Portfolio Diversification Effects 109
Within this framework it is possible to calculate the diversification effects beyond

the sample range and for hypothetically larger portfolios if we make some assump-
tions regarding the market model betas and scale coefficients of the orthogonal risk
factors. The diversification speeds are analyzed graphically.
We start our essay by reviewing the Feller’s convolution theorem for distri-
butions with heavy tails. Subsequently we study the diversification problem in
more detail by adding the market factor. The relevance of the theoretical results
for the downside risk portfolio diversification question is demonstrated by an
application to Standard & Poor’s (S&P) stock returns.

1 DIVERSIFICATION EFFECTS AND THE FELLER CONVOLUTION
THEOREM
In this section we only consider securities that are independently distributed. In
the next section this counterfactual assumption, as least as far as equities are
concerned, is relaxed by allowing for common factors. Let Ri denote the logarith-
mic return of the ith security. Suppose the fRi g are generated by a distribution
with heavy tails in the sense of regular variation at infinity. Thus, far from the
origin, the Pareto term dominates:
PrfRi xg ¼ Ai x ½1 þ oð1Þ, > 0, Ai > 0 ð1Þ

as x ! 1. The Pareto term implies that only moments up to are bounded and
hence the informal terminology of heavy tails. In contrast, the normal distribution
has all moments bounded thanks to the exponential tail shape. Distributions like
the Student’s t, Pareto, and nonnormal sum-stable distributions all have regularly
varying tails. Downside risk measures like the VaR, that is, at the desired prob-
ability level d: PrfRi VaRg ¼ d, directly pick up differences in tail behavior.
An implication of the regular variation property is the simplicity of the tail
probabilities for convoluted data. Suppose the fRi g are generated by a heavy-
tailed distribution that satisfies Equation (1). From Feller’s theorem (1971, VIII.8),
the distribution of the k sum satisfies2
( )
Xk
Pr Ri x ¼ kAx ½1 þ oð1Þ, as x ! 1:
i¼1
From this, one can derive the diversification effect for the equally weighted
P
portfolio R ¼ 1k ki¼1 Ri [see Dacorogna et al. (2001)]. The following first-order
approximation for the equally weighted portfolio diversification effect regarding
the downside risk obtains3
2
Note that in this analysis x ! 1, while k is a fixed number.
3
Note that this diversification result only holds as x ! 1. Garcia, Renault, and Tsafack (2003) show that for
symmetric stable distributions, the diversification result applies anywhere below the median. This has to
do with the fact that the sum stable distributions are self-additive throughout their support, while this
only applies in the tail region for the class of fat-tailed distributions.
( )
1X k
Pr Ri x k1 Ax : ð2Þ
k i¼1
Under the heterogeneity of the scale coefficients Ai, the equivalent of Equation (2)
reads
( ) !
1X k X
k
Pr Ri x k Ai x : ð3Þ
k i¼1 i¼1

To summarize, if at a constant VaR level x, one increases the number k of
securities included in the portfolio, this decreases the probability of loss by k1-
[see Equation (2)].
The other case is where the Ri are independent standard normally distributed
( )
1X k
1
Pr Ri x Nð0, Þ:
k i¼1 k
The following is the equivalent of Equation (1) for the normal distribution
1 1 1
PrfRi xg ¼ pffiffiffiffiffiffi expð x2 Þ½1 þ oð1Þ as x ! 1:
x 2p 2
For the equally weighted portfolio it thus holds

( )
1X k
1 1 1 1
Pr Ri x ¼ Pr pffiffiffi Ri x ’ pffiffiffi pffiffiffiffiffiffi expð kx2 Þ: ð4Þ
k i¼1 k x k 2p 2
It follows that under normality

d ln Pr 1 1
’ x2 k, ð5Þ
d ln k 2 2
while under the fat-tail model from Equation (2),
d ln Pr
’ 1 : ð6Þ
d ln k
Hence, for sufficiently high but fixed k, the normal distribution implies a higher
diversification speed of the risk level.
Next, consider holding the probability constant but letting the VaR level
change, which is the typical case considered under the safety-first criterion, to
determine the VaR diversification speed. Thus, in the case of the normal model,
we are interested in comparing VaR levels t and s such that
( )
1X k
1
PrfRi tg ¼ Pr Ri s ¼ Pr pffiffiffi Ri s : ð7Þ
k i¼1 k
Using the additivity properties of the normal distribution, or equivalently using

Equation (4) on both sides of Equation (7), it is immediate that
t
s ¼ pffiffiffi
k
so that the normal-based VaR diversification speed reads
d ln s 1
¼ : ð8Þ
d ln k 2
For the fat-tailed model, the equivalent of Equation (7) is

( ) !
1X k

X k
Ai t ¼ PrfRi tg ¼ Pr Ri s ¼ k Ai s :
k i¼1 i¼1
Solving for s gives

Pk !1=
t i¼1 Ai
s¼ :
k Ai
Furthermore, if the scale coefficients are identical, this simplifies to

t
s¼ :
k11=
So that if > 2, that is, when the variance exists,
d ln s 1 1
¼ ð1 Þ < : ð9Þ
d ln k 2
Compare Equation (9) to Equation (8). If > 2, then the VaR diversification speed
is higher for fat-tailed distributed returns than if the returns were normally
distributed.
2 DIVERSIFICATION EFFECTS IN FACTOR MODELS

We relax the assumption of independence between security returns and allow for
nondiversifiable market risk. The market risk reduces the benefits from diversi-
fication to the elimination of the idiosyncratic component of the risk. First con-
sider a single index model in which all idiosyncratic risk is assumed independent
from the market risk R,
Ri ¼ bi R þ Qi , ð10Þ
and where R is the (excess) return on the market portfolio, bi is the amount of
market risk, and Qi is the idiosyncratic risk of the return on asset i. The idiosyn-
cratic risk may be diversified away fully in arbitrarily large portfolios and hence is
not priced. But the cross-sectional dependence induced by common market risk
factors has to be held in any portfolio.
We apply Feller’s theorem again for deriving the benefits from cross-
sectional portfolio diversification in this single index model. Consider an
P
equally weighted portfolio of k assets. Let ¼ 1k ki¼1 i . The case of unequally
weighted portfolios is but a minor extension left to the reader for consideration
of space. In this single index model, the Qi are cross-sectionally independent
and, moreover, are independent from the market risk factor R. Suppose, in
addition, that the Qi satisfy PrfQi xg Ai x for all i, and that
PrfR xg Ar x . The diversification benefits from the equally weighted
portfolio regarding the downside risk measure for the case of homogeneous
scale coefficients Ai = A then follow as

( )
1X k

Pr Ri x ¼ k1 Ax ½1 þ oð1Þ þ b Ar x ½1 þ oð1Þ, ð11Þ
k i¼1
as x ! 1. If the scale coefficients are heterogeneous, the equivalent of Equation

(11) reads
( ) !
1X k

Xk

Pr Ri x k Ai x þ b Ar x : ð12Þ
k i¼1 i¼1
In large portfolios one should see that almost all downside risk is driven by the
market factor, if > 1,
( )
1X k

Pr Ri x b Ar x
k i¼1
for large, but finite k.

In general, one finds the single index model does not hold exactly due to the
fact that cov½Qi ,Qj is typically nonzero for off-diagonal elements as well. Thus,
though the Qi may be independent from the market risk factor R (they are
uncorrelated with R by construction), they are typically not cross-sectionally
independent from each other. This case is usually referred to as the market
model. For example, let there be one other common factor F. This factor is assumed
P
independent from R, but the cov½Qi , F=cov½F, F ¼ i . Let ¼ 1k ki¼1 i , and
assume that PrfF xg Af x . Then, by analogy with the foregoing results,
( ) !
1X k

Xk

Pr Ri x k Ai x þ b Ar x þ Af x : ð13Þ
k i¼1 i¼1
To study the case of nonidentical a in Equation (12), one has to consider two
cases:
Case 1: r ¼ 1 ¼ . . . ¼ j < jþ1 jþ2 . . . k :
Case 2: 1 ¼ . . . ¼ j < jþ1 jþ2 . . . k and r > 1.

Here, r stands for the tail index of the market portfolio return, and the i are
the indices of the idiosyncratic parts of the security i return. Then corresponding
expressions to Equation (12) are for Case 1,
( ) j
!
1X k
r
X r
Pr Ri x k Ai xr þ b Ar xr ,
k i¼1 i¼1
and for Case 2,

( ) j
!
1X k X
Pr Ri x k1 Ai x1 :

k i¼1 i¼1
Next, consider holding the probability constant, but let the VaR level change
in Equation (12) as the number of assets k increases. From Equation (12), we have
( ) " ! #
1X k

Xk Xk
Pr Ri x k Ai þ bi Ar x :
k i¼1 i¼1 i¼1
By first-order inversion [cf. De Bruijn’s theorem in Bingham, Goldie, and Teugels

(1987)] one obtains
" ! #1=
1 Xk X
k
VaR ¼ x ¼ Ai þ bi Ar p1= ,
ð14Þ
k i¼1 i¼1
and where p is the fixed probability level. With homogeneous scale coefficients,
we may simplify this to
2 P 31=
k
1 4 i¼1 bi
VaR ¼ Aþ Ar 5 p1= :

k11= k
This should be compared with the results from the previous section on the VaR
diversification speed, where the part stemming from the market factor was
absent. In particular, we find
d ln VaR 1 A
¼ 1 þ Pk ,
d ln k bi
Aþ i¼1
k Ar
which is smaller, that is, gives a higher speed, than the simple 1 þ 1= from
before.
3 ESTIMATION BY POOLING
To investigate the relevance of the above downside risk diversification theory, we
need to estimate the various downside risk components. To explain the details of
the estimation procedure, consider again the simple setup in Equation (3). To be
able to calculate the downside risk measure, one needs estimates of the tail index
a and the scale coefficients Ai . A popular estimator for the inverse of the tail index
is Hill’s (1975) estimator. If the only sources of heterogeneity are the scale coef-
ficients, one can pool all return series. Let fR11 , . . . , R1n , . . . , Rk1 , . . . , Rkn g be the
sample of returns. Denote by Z(i) the ith descending order statistic from
fR11 , . . . , R1n , . . . , Rk1 , . . . , Rkn g. If we estimate the left tail of the distribution, it
is understood that we take the losses (reverse signs). The Hill estimator reads
X m
d¼ 1
1= ln ZðiÞ ln Zðmþ1Þ : ð15Þ
m i¼1

This estimator requires a choice of the number of the highest-order statistics m to
be included, that is, one needs to locate the start of the tail area. We implemented
the subsample bootstrap method proposed by Danielsson et al. (2000) to deter-
mine m. The estimator for the scale A when Ai ¼ A for all i is
^ ¼ m ðZðmþ1Þ Þ^ :
A
kn
Note that m=nk is the empirical probability associated with Zðmþ1Þ , and the
estimator Â follows intuitively from Equation (1). Under the heterogeneity of Ai
one takes
^ i ¼ mi ðZðmþ1Þ Þ^ ,
A
n
where mi is such that
Rið1Þ . . . Riðmi Þ Zðmþ1Þ Riðmi þ1Þ . . . RiðnÞ :

P
Note that ki¼1 mi ¼ m. This implies that by the pooling method we obtain exactly
the same portfolio probabilities whether or not one assumes (counterfactually
incorrect) identical or heterogeneous scale coefficients, since
! !
X
k Xk
mi ^

k
^ ^
^
Ai x ¼ k
^
ðZðmþ1Þ Þ x^
i¼1 i¼1
n
P
k
i¼1 mi
¼ k^ ðZðmþ1Þ Þ^ x^
n
^ ^ :
¼ k1^ Ax
We can adapt this pooling method to the market model with little modifica-
tion. Pooling the series fRg,fQ1 g, . . . ,fQk g, one can use the same procedure as in
the case of cross-independence.4 For the estimation of the tail index one again
uses Equation (15), where in this case fZg ¼ fRr1 , . . . , Rrn , Q11 , . . . , Q1n , . . . ,
Qk1 , . . . , Qkn g. Estimators for the scales are
^ i ¼ mi ðZðmþ1Þ Þ^ ,i ¼ 1; . . . ; k and r,

A
n
where mi is such that
Xið1Þ . . . Xiðmi Þ Zðmþ1Þ Xiðmi þ1Þ . . . XiðnÞ ,
where Xi can be R or Qi.

In case the tail indexes differ across securities and risk factors, the above can
be easily adapted to estimation on individual series. There is, however, consider-
able evidence that the tail indexes are comparable for equities from the S&P 500
index [see, e.g., Jansen and De Vries (1991) and Hyung and De Vries (2002)].
Therefore we decided to proceed on the basis of the assumption that the tail
indexes are equal.
4 EMPIRICAL ANALYSIS OF THE DIVERSIFICATION SPEED

We now apply our theoretical results to the daily returns of a set of stocks. In order to
estimate the parameters of the market model we choose the Standard & Poor’s 500
index as a representation of the market factor. This is certainly not the market port-
folio as in the capital asset pricing model (CAPM); nevertheless, the S&P 500 index
represents about 80% of the total market capitalization. To see the effects of portfolio
diversification, we choose 15 stocks arbitrarily from the S&P 100 index in March 2001.
We use the daily returns (close-to-close data), including cash dividends. The data
were obtained from Datastream. The data span runs from January 2, 1980, through
March 6, 2001, giving a sample size of n = 5,526. Thus more than 20 years of daily data
are considered, including the short-lived 1987 crash. All results are in terms of the
excess returns above the risk-free interest rate (three-month U.S. Treasury bills).
The summary statistics for each stock return series and the market factor are
given in Table 1. On an annual basis, the excess returns hover around 7.5% and
have comparable second moments. The excess returns all exhibit considerably
higher than normal kurtosis. This latter feature is also captured by the estimates
of the tail index a in Table 2. In this table we report tail index and scale estimates
using the individual series, counter to the pooling method outlined above. This is
done in order to show that the tail indexes are indeed rather similar, while there is
considerable variation in the scales. This motivates the single-tail index, hetero-
geneous scale model implemented in the other tables. Table 2 also gives the beta
estimates for the market model.
4
The determination of the parameters bi and the residuals Qi entering in the definition of the market model
is done by regressing the stock returns on the market return. The coefficient bi is thus given by the
ordinary least squares estimator, which is consistent as long as the residuals are white noise and have
zero mean and finite variance. The idiosyncratic noise Qi is obtained by subtracting bi times the market
return to the stock return.
Table 1 Selected stocks and summary statistics of excess returns.

Series Name m1 m2 m3 m4
m S&P 500 Index .0747 2.52 2.31 55.49

1 ALCOA .0707 4.84 0.26 13.39
2 AT&T .0392 4.33 0.35 16.41
3 Black & Decker .0168 5.61 0.32 10.57
4 Campbell Soup .0897 4.37 0.28 9.06
5 Disney (WALT) .0981 4.86 1.30 29.82
6 Entergy .0454 4.06 0.97 23.66

7 General Dynamics .0764 4.53 0.26 10.24
8 Heinz HJ .0968 3.99 0.11 6.35
9 Johnson & Johnson .1053 4.08 0.32 9.45
10 Merck .1212 3.96 0.03 6.31
11 Pepsico .1170 4.43 0.04 7.82
12 Ralston Purina .1077 4.08 0.70 15.41
13 Sears Roebuck .0542 4.91 0.24 16.83
14 United Technologies .0851 4.19 0.10 6.83
15 Xerox .0423 5.48 1.78 33.74
Observations cover January 1, 1980–March 6, 2001, giving 5,526 daily observations. The m1, m2, m3, and m4
denote the sample mean, standard error, skewness, and kurtosis of annualized excess returns, respectively.
The estimates are reported in terms of the excess returns above the risk-free interest rate (U.S. three-month
Treasury bill).
Table 2 Left-tail parameter estimates.

Series A m
Rm 2.963 2.522 298

1 3.789 110.117 113
2 2.785 7.953 289
3 3.220 58.601 136
4 3.505 48.766 68
5 2.549 6.211 496
6 1.981 1.339 682
7 3.218 27.687 140
8 3.404 25.811 197
9 3.377 23.663 292
10 4.035 104.724 62
11 3.789 103.171 71
12 3.136 14.106 190
13 3.166 28.244 256
14 4.335 288.036 66
15 2.098 2.999 537
The values in columns a, A, and m are, respectively, the tail index, the scale parameter, and the
estimated optimal number of order statistics.
Table 3 Left-tail parameter estimates.

Excess returns Residuals
Series A m b A m
T 23.0 1609 – 19.6 1021

Rm – – 1 4.3 15
1 26.2 122 0.877 24.7 86
2 19.5 91 0.929 15.2 53
3 46.4 216 0.938 42.2 147

4 22.7 106 0.719 19.5 68
5 24.0 112 1.012 22.1 77
6 14.4 67 0.475 14.9 52
7 25.3 118 0.710 25.0 87
8 16.3 76 0.640 14.9 52
9 13.9 65 0.927 10.6 37
10 15.7 73 0.854 11.5 40
11 24.2 113 0.867 18.7 65
12 15.0 70 0.669 16.4 57
13 29.0 135 1.074 17.5 61
14 20.2 94 0.895 13.2 46
15 32.4 151 0.949 26.7 93
The values in row T give estimates from the pooled series imposing scale homogeneity. The values in rows
Rm, 1, 2, . . . , 15 give estimates for the market returns and the individual stock series for the total excess
returns and the residual parts. The values in columns A and m are the scale parameter and the estimated
optimal number of order statistics imposing identical tail indexes. The values in column b are the market
model beta.
In Table 3, computations proceed by using the pooling method, assuming

identical tail indexes for all risk components. We report the estimates of the scale
parameter A and the optimal number of order statistics m. Both are calculated for
the series of excess returns and for the (constructed) orthogonal residuals from the
market model (using the betas). The tail index estimate using all excess returns is
3.163, while when we use all the residuals the tail index is 3.246. The scale parameter
estimates, however, differ considerably as these range between 14.4 and 46.4 for the
excess returns, and are between 4.3 and 42.2 for market returns and residuals. We
note that the scale estimates for the excess returns using the pooling method are
more homogeneous than when using the individual series approach from Table 2.
The effects of portfolio diversification are reported in Table 4. The downside
risk measure is the probability of a loss in excess of the VaR level s; we report at
four different loss levels (s = 7.10, 11.69, 13.33, and 15.97).5 Four different levels of
portfolio aggregation are considered: 1 stock, 5 stocks, 10 stocks, and 15 stocks. The
numbers in row EMP are the probabilities from the empirical distribution function
5
We chose this particular set of VaR values from the 5.0%, 1.0%, 0.5%, and 0.25% quantiles of the market
returns.
Table 4 Lower tail probabilities in percentages.

s 7.10 11.69
k 1 5 10 15 1 5 10 15
EMP 4.995 1.195 0.633 0.579 0.995 0.253 0.145 0.145

NOR 7.325 0.934 0.225 0.198 0.817 0.005 0.000 0.000
FAT 6.551 1.181 0.741 0.706 0.988 0.265 0.185 0.171
CDp – 0.633 0.392 0.423 – 0.125 0.078 0.084
s 13.33 15.97
k 1 5 10 15 1 5 10 15

EMP 0.489 0.163 0.109 0.127 0.235 0.109 0.090 0.090
NOR 0.309 0.000 0.000 0.000 0.051 0.000 0.000 0.000
FAT 0.603 0.179 0.129 0.118 0.304 0.104 0.078 0.071
CDp – 0.082 0.051 0.055 – 0.046 0.028 0.030
The entries in row EMP are the probabilities from the empirical distribution. The rows NOR row and FAT
report the probabilities calculated directly from the parameters of the averaged series itself, where in the
former case one uses the presumption of normality and in the latter case regular variation is imposed. The
numbers in rows CDp are the probabilities estimated using the pooled series. The k denotes the number
of individual stocks included in the averaged series, and s is the loss quantile. Probabilities are written
in Percentage format.
of the total return series. The normal law is often used as the workhorse distribu-
tion model in finance, even though it does not capture the characteristic tail feature
of the data. Therefore, in the rows labeled NOR, we give the probabilities from the
normal model-based formula, using the mean and variance estimates from the
averaged series. The estimated values in rows FAT were obtained by the heavy-
P
tailed model using the averaged total excess returns ki¼1 Ri =k. The rows CDp give
the probability estimates from the pooled series on the basis of Equation (12),
assuming the heterogeneous scale model. One notes that the normal model does
well in the center, but performs poorly as one moves into the tail part. In contrast,
the averaged series in rows FAT is always quite close to the empirical distribution
function in the tail area. This shows that the heavy-tailed model is much better at
capturing the tail properties. If we turn to the last rows, one notes that the model in
Equation (12) does capture a considerable part of the tail risk of the portfolio, but
that there is a gap between the tail risk that is explained by the model and which is
left unexplained. This is further interpreted below.
To judge these results and to study the speed of diversification, a graphical
exposition is insightful. In Figures 1 and 2 we show the diversification speed of
the risk level by plotting the probability of loss for two different VaR levels
against the number of securities that are included in the portfolio.6 Figure 1 is
for the 7.10 VaR level and Figure 2 is for the 15.97 VaR level. The top line gives the
total amount of tail risk by means of the empirical distribution function. The gray
6
The order by which the securities are included corresponds to the numbering in Table 1.
0.05
0.04
0.03
Probability

0.02
Market component Idiosyncratic component Total risk
0.01
0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Figure 1 Downside risk decomposition at s ¼ 7.10 (fat-tailed case).
0.0040
0.0035
0.0030
0.0025
Probability
0.0020
Market component Idiosyncratic component Total risk
0.0015
0.0010
0.0005
0.0000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Figure 2 Downside risk decomposition at s ¼ 15.97 (fat-tailed case).
area constitutes the market risk component, while the black area contains the
idiosyncratic risk from Equation (12). Note that the idiosyncratic risk is basically
eliminated once the portfolio includes about seven stocks. To put this result into
perspective, we also provide a graph for the speed of diversification concerning
the variance (see Figure 3). This is a global risk measure, and under independence
25.0
20.0
Market component Idiosyncratic component Total variance
15.0
Variance

10.0
5.0
0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Figure 3 Variance decomposition.
the variance of the idiosyncratic part should decline linearly in k. As can be seen
from this latter figure, it takes approximately twice the number of securities to
eliminate the variance part contributed by the idiosyncratic risk [cf. Elton and
Gruber (1995)]. Note that this corroborates the rate given in Equation (6) and the
value of ’ 3 as in Table 2 (while the variance declines at speed 1). Interestingly,
as noted at the end of the previous paragraph, another remarkable difference
between the last figure and the first two figures is the size of the residual risk
driven by the factors other than the market factor. While this component is rela-
tively minor for the variance risk measure, it is even larger than the market risk
component for the downside risk measure. This points to the presence of another
factor, F, uncorrelated with R as in Equation (13). This other factor induces a small
correlation between the residuals (see Figure 3). This small correlation not with-
standing, the other factor appears important with respect to the downside risk. In
future research we hope to relate this factor to economic variables.
Next we compare the VaR diversification speed under the normal model with
the fat-tailed model. To plot the VaR diversification speed, we now look in the
VaR k space. From Equation (14) it is clear one cannot separate the market part
from the idiosyncratic part due to the power 1=. Nevertheless, one can first plot
the VaR level doing as if only the market factor were relevant (e.g., this would be
the case if the idiosyncratic risks have a higher tail index compared to the market
index). The market factor is from Equation (14):
1X k
x¼ð b Þ½Ar 1=
p1= : ð16Þ
k i¼1 i
8.0
7.0
Market component Idiosyncratic component VaR of Portfolio
6.0
5.0
VaR
4.0

3.0
2.0
1.0
0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Figure 4 VaR decomposition at p = 0.05 (fat-tailed case).
The next line plots the combined effect, market factor, and idiosyncratic compo-
nents, which simply is Equation (14). Third, one plots the empirical quantile
function as more assets are added. Similarly one can proceed in this fashion
under the assumption that the returns follow the normal distribution.
Figures 4–7 show the decreasing level of VaR for the given probability. Figure 4
is for the 0.05 probability level, and Figure 5 is for the 0.0025 probability level for
a fat-tailed distribution. The top line gives the total amount of VaR by means of
the empirical distribution function. The gray area constitutes the VaR level from
the market risk component, as in Equation (16), while the black area plus the
gray area displays Equation (14). Figure 6 is for the 0.05 probability level, and
Figure 7 is for the 0.0025 probability level for a normal distribution. These figures
clearly display the theoretical prediction of Equation (9), that the VaR diversifi-
cation speed for the idiosyncratic risk is lower for the normal model than for the
fat-tailed model.
5 OUT OF SAMPLE, OUT OF PORTFOLIO

The semiparametric approach we followed to construct the downside risk mea-
sure can also be used to go beyond the sample. We consider two possible
applications of this technique which might be of use to risk managers. The first
application asks the question how much extra diversification benefits could be
derived from adding more securities, without having observations on these
securities. By making an assumption regarding the value of the average beta
and the average scale of the residual risk factors in the enlarged portfolio, one
18.0
16.0

14.0
12.0
10.0
VaR

8.0
6.0
4.0
2.0
0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Figure 5 VaR decomposition at p = 0.0025 (fat-tailed case).
8.0
7.0
6.0
5.0
VaR
4.0
3.0
2.0
1.0
0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Figure 6 VaR decomposition at p = 0.05 (normal case).
can use Equation (12) to extrapolate to larger than sample size portfolios. A
second application is to increase the loss levels at which one wants to evaluate
the downside risk level beyond the worst case in the sample. Moreover, even at
18.0
16.0
14.0
12.0
10.0
VaR

8.0
6.0
4.0
2.0
0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Figure 7 VaR decomposition at p = 0.0025 (normal case).
the border of the sample, our approach has real benefits. By its very nature, the
empirical distribution is bounded by the worst case and hence has its limitations,
since the worst case is a bad estimator of the quantile at the 1/n probability level
(and vice versa). Thus increasing the loss level x in Equation (12) beyond the
worst case gives an idea about the risk of observing even higher losses.
In Table 5, the block denoted as Case I summarizes some information from
Table 4. The Case III block addresses the first application by increasing the
number of securities k beyond the sample value of 15. We assumed the following
average beta values: b ¼ 0:7, 0.83, and 0.9. The Case II block increases the loss
return level. In Table 4 we used 15.97 as the highest loss level. Above this level,
many securities have no observations. There is one equity with much higher loss
returns and we used this one to provide the ‘‘out of sample’’ loss levels of 22.03,
25.21, 33.69, and 40.45, respectively. To interpret Case III, note that the inclusion of
more stocks that have a close correlation with the market component increases the
loss probability for a given VaR level. For example, consider a portfolio of k ¼ 30
stocks, at the 15.97 quantile, when b ¼ 0:7 the probability is 0.0169, but when
b ¼ 0:9 the probability increases to 0.0381.
6 CONCLUSION
securities. In this article we study the benefits of portfolio diversification with
respect to extreme downside risk, or the VaR risk measure. The risk of a security
is decomposed into a part that is attributable to the market risk and an independent
Table 5 Lower tail probabilities: Beyond the sample and market.

s 7.10 11.69 13.33 15.97 22.03 25.21 33.69 40.45
% 5.0 1.0 0.5 0.25 0.090 0.054 0.018 0.009
k CASE I CASE II
EMP 1.1946 .2534 .1629 .1086 .0362 .0362 .0181 .0181

5 FAT 1.1900 .2660 .1798 .1045 .0397 .0265 .0111 .0064
CDp .6093 .1205 .0789 .0439 .0154 .0100 .0039 .0021
EMP .6335 .1448 .1086 .0905 .0181 .0181 .0181 .0181
10 FAT .6800 .1490 .1001 .0578 .0217 .0144 .0060 .0034

CDp .3914 .0774 .0507 .0282 .0099 .0064 .0025 .0014
EMP .5792 .1448 .1267 .0905 .0181 .0181 .0181 .0181
15 FAT .7087 .1722 .1189 .0712 .0286 .0195 .0086 .0051
CDp .4227 .0836 .0547 .0304 .0107 .0069 .0027 .0015
CASE III
CDp1 .2375 .0470 .0307 .0171

20 CDp2 .4190 .0829 .0543 .0302
CDp3 .5318 .1052 .0689 .0383
CDp1 .2359 .0467 .0305 .0170
25 CDp2 .4175 .0826 .0541 .0301
CDp3 .5302 .1049 .0687 .0382
CDp1 .2350 .0465 .0304 .0169
30 CDp2 .4166 .0824 .0540 .0300
CDp3 .5294 .1047 .0686 .0381
The entries in rows EMP are the probabilities from the empirical distribution. The numbers in rows FAT
are the probabilities calculated directly from the parameters of the average series itself. The numbers in
row CDp are the probabilities from the fat-tail market model of Equation (12). The numbers in rows CDp1,
2, and 3 are calculated by imposing b = 0.7, 0.8358, and 0.9, respectively. The k denotes the number of
individual stocks included in the averaged series, and s gives the loss quantile. Probabilities are written in
percentage format.
risk factor. The independent part consists of an idiosyncratic part and a second
common factor. Two different measures for diversification effects are studied. The
VaR diversification speed measure holds the probability level constant and gives
the rate of change by which the VaR declines as more securities are added to the
portfolio, while the diversification speed of the risk level holds the VaR level
constant and measures the decline in the probability level. For the VaR diversifica-
tion speed measure, we argued fat-tailed distributed idiosyncratic risk factors
should go down at a higher speed than normal distributed idiosyncratic risk
factors. This theoretical prediction was also found empirically to be the case.
Furthermore, we provide predictions for the downside risk diversification benefits
beyond the range of the empirical distribution function.
This research can be extended in several directions. Given the large gaps in
Figures 1 and 3 between the total downside risk and the market factor downside
risk contribution, it is of interest to see whether one can identify the remaining
risk factors F, as in Equation (13). Moreover, one would like to explain why these
remaining risk factors are relatively unimportant for the global risk measure such
as the variance. Moreover, the above analysis may explain why many investors
seem to hold not-so-well-diversified portfolios if a global risk measure like the
variance is used as the yardstick.
Received March 1, 2004; revised October 8, 2004; accepted October 19, 2004.

REFERENCES
Arzac, E., and V. Bawa. (1977). ‘‘Portfolio Choice and Equilibrium in Capital Markets
with Safety First Investors.’’ Journal of Financial Economics 4, 277–288.
Bingham, N. H., C. M. Goldie, and J. L. Teugels. (1987). Regular Variation. Cambridge:
Cambridge University Press.
Dacorogna, M. M., U. A. Müller, O. V. Pictet, and C. G. de Vries. (2001). ‘‘Extremal
Forex Returns in Extremely Large Data Sets.’’ Extremes 4, 105–127.
Danielsson, J., L. de Haan, L. Peng, and C. G. de Vries. (2000). ‘‘Using a Bootstrap
Method to Choose the Sample Fraction in Tail Index Estimation.’’ Journal of Multi-
variate Analysis 76, 226–248.
Elton, E. J., and M. J. Gruber. (1995). Modern Portfolio Theory and Investment Analysis,
5th ed. New York: John Wiley & Sons.
Feller, W. (1971). An Introduction to Probability Theory and Its Applications, vol. II. New
York: John Wiley & Sons.
Garcia, R., E. Renault, and G. Tsafack. (2003). ‘‘Proper Conditioning for Coherent VaR
in Portfolio Management.’’ Paper presented at the CFS workshop, November
2003.
Gourieroux, C., J. P. Laurent, and O. Scaillet. (2000). ‘‘Sensitivity Analysis of Values at
Risk.’’ Journal of Empirical Finance 7, 225–246.
Hill, B. M. (1975). ‘‘A Simple General Approach to Inference About the Tail of a
Distribution.’’ Annals of Statistics 3, 1163–1173.
Hyung, N., and C. G. de Vries. (2002). ‘‘Portfolio Diversification Effects and Regular
Variation in Financial Data.’’ Allgemeines Statistisches Archiv / Journal of the German
Statistical Society 86, 69–82.
Jansen, D., and C. G. de Vries. (1991). ‘‘On the Frequency of Large Stock Returns:
Putting Booms and Busts into Perspective.’’ Review of Economics and Statistics 73,
18–24.
Jansen, D., K. G. Koedijk, and C. G. de Vries. (2000). ‘‘Portfolio Selection with Limited
Downside Risk.’’ Journal of Empirical Finance 7, 247–269.

Hyung 2005

Uploaded by

Copyright:

Available Formats

Hyung 2005

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Hyung 2005

Uploaded by

Copyright:

Available Formats

Journal of Financial Econometrics, 2005, Vol. 3, No.

Portfolio Diversification Effects

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

keywords: diversification, portfolio decomposition, value-at-risk

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

Within this framework it is possible to calculate the diversification effects beyond

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

PrfRi xg ¼ Ai x ½1 þ oð1Þ, > 0, Ai > 0 ð1Þ

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

For the equally weighted portfolio it thus holds

It follows that under normality

Using the additivity properties of the normal distribution, or equivalently using

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

Solving for s gives

Furthermore, if the scale coefficients are identical, this simplifies to

2 DIVERSIFICATION EFFECTS IN FACTOR MODELS

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

as x ! 1. If the scale coefficients are heterogeneous, the equivalent of Equation

for large, but finite k.

Case 1: r ¼ 1 ¼ . . . ¼ j < jþ1 jþ2 . . . k :

Case 2: 1 ¼ . . . ¼ j < jþ1 jþ2 . . . k and r > 1.

and for Case 2,

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

By first-order inversion [cf. De Bruijn’s theorem in Bingham, Goldie, and Teugels

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

where mi is such that

Rið1Þ  . . .  Riðmi Þ  Zðmþ1Þ  Riðmi þ1Þ  . . .  RiðnÞ :

^ i ¼ mi ðZðmþ1Þ Þ^ ,i ¼ 1; . . . ; k and r,

Xið1Þ  . . .  Xiðmi Þ  Zðmþ1Þ  Xiðmi þ1Þ  . . .  XiðnÞ ,

where Xi can be R or Qi.

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

4 EMPIRICAL ANALYSIS OF THE DIVERSIFICATION SPEED

Table 1 Selected stocks and summary statistics of excess returns.

m S&P 500 Index .0747 2.52 2.31 55.49

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

Table 2 Left-tail parameter estimates.

Rm 2.963 2.522 298

Table 3 Left-tail parameter estimates.

T 23.0 1609 – 19.6 1021

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

In Table 3, computations proceed by using the pooling method, assuming

Table 4 Lower tail probabilities in percentages.

EMP 4.995 1.195 0.633 0.579 0.995 0.253 0.145 0.145

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

Market component Idiosyncratic component Total risk

Figure 1 Downside risk decomposition at s ¼ 7.10 (fat-tailed case).

Market component Idiosyncratic component Total risk

Figure 2 Downside risk decomposition at s ¼ 15.97 (fat-tailed case).

Market component Idiosyncratic component Total variance

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

Figure 3 Variance decomposition.

Market component Idiosyncratic component VaR of Portfolio

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

Figure 4 VaR decomposition at p = 0.05 (fat-tailed case).

5 OUT OF SAMPLE, OUT OF PORTFOLIO

Market component Idiosyncratic component VaR of Portfolio

Downloaded from http://jfec.oxfordjournals.org/ at University of Chicago on July 10, 2015

Figure 5 VaR decomposition at p = 0.0025 (fat-tailed case).

Rið1Þ . . . Riðmi Þ Zðmþ1Þ Riðmi þ1Þ . . . RiðnÞ :

Xið1Þ . . . Xiðmi Þ Zðmþ1Þ Xiðmi þ1Þ . . . XiðnÞ ,