Risk

Technical article
Risk is the world's leading financial risk management
magazine. Risk’s Cutting Edge articles are a showcase for the
latest thinking and research into derivatives tools and
techniques, with comprehensive coverage of the most recent
advances in areas such as option pricing and hedging, market
risk, credit risk, swaps and Monte Carlo methods.

www.risk.net/technical

Copyright statement © Risk Waters Group Ltd, 2003. All rights

This PDF is made available for personal use reserved. No parts of this article may be
only. Under the following terms you are reproduced, stored in or introduced into any
granted permission to view online, print out, retrieval system, or transmitted, in any form
or store on your computer a single copy, or by any means, electronic, mechanical,
provided this is intended for the benefit of photocopying, recording or otherwise
the individual user. Should you wish to use without the prior written permission of the
multiple copies of this article, please contact copyright owners.
reprints@riskwaters.com
 Portfolio risk management l Cutting edge

VAR: who contributes

and how much?
Understanding the risk of an asset in a portfolio is fundamental to managing and pricing risk.
Richard Martin, Kevin Thompson and Christopher Browne extend their work on saddle-point
techniques for portfolio loss distributions by analytically deriving the sensitivity of VAR to asset
allocation

For non-normal distributions, it correctly penalises large assets more

T
o begin understanding and optimising the risk of a portfolio, portfolio
and business managers need to know how each component of that strongly than mean-variance does. The allocation of risk between the as-
portfolio contributes to the overall risk. A common measure of risk sets is consistent with the investor’s view of risk implied by the chosen
contribution is the sensitivity of the risk to an infinitesimal fractional change confidence level of VAR.
in asset allocation. These have the convenient property that their sum is the
It can be reformulated and generalised using utility theory.
overall risk.1 The simplest measure of risk is the mean-variance framework,
in which the risk is equal to the mean plus a given number of standard de- Review of the saddle-point method
viations (or simply the standard deviation). This is sufficient if the assets in Before embarking on the derivation of the VAR contributions, we review
the portfolio have a joint distribution that is multivariate normal. It is easy the saddle-point methodology for deriving an analytical expression for the
to derive the sensitivities of the mean and standard deviation to asset allo- VAR of a portfolio from the loss distributions of the individual assets.
cation. But real loss distributions, and particularly those in credit, operational A portfolio loss distribution is constructed by adding the losses of the
and insurance risk, are not normal, and the standard deviation does not ad- individual assets. For such ‘additive’ problems, a useful construction is the
equately capture the risk of large losses. Value-at-risk, a quantile of the dis- cumulant generating function (KGF), which for a random variable X is de-
tribution at some specified level of confidence, better captures extreme risks. fined as K X(s) = logE[esX], for real or complex s. The tail probability of X
It has some undesirable properties, such as lack of subadditivity2, which are can be recovered from the KGF by a contour integral:
not shared by other risk measures, such as expected shortfall. However, VAR
P (X > t ) =
1
+ i ∞ exp K
X (s) − st( )
is currently an industry standard often used for setting bank capital and it is ∫
2πi − i ∞,(0+ ) s
ds (1)
the one that we shall study here. Leibowitz & Henriksson (1989) have dis-
cussed the use of VAR constraints in portfolio optimisation and Litterman in which the path of integration is up the imaginary axis and the notation
(1996) discusses risk contributions in the VAR-based framework. Most re- (0 +) indicates that the contour runs to the right of the origin to avoid the
cently, Gouriéroux, Laurent & Scaillet (2000) have given a general expres- pole there.
sion for sensitivity of VAR to portfolio allocation, but to put this into practice The KGF is a useful construction because when independent random
still requires Monte Carlo simulation (except in the Gaussian case). variables are added (Y = Σ ja jX j), their KGFs are added:
What would be most powerful, however, is a method of deriving the sen-
sitivity of VAR to asset allocation directly from a model, without any need for K Y (s) = ∑ KX j (ajs) (2)
j
simulation. In a previous article (Martin, Thompson & Browne, 2001), we
demonstrated the saddle-point method, an analytical technique that approx- For example, if a loan portfolio consists of n assets with default proba-
imates VAR with great accuracy in the tail of the loss distribution. It relies on bilities pj and losses-given-default aj, then Xj represents a default event (it is
construction of the moment-generating function, and not on Monte Carlo sim- one if the jth asset defaults and zero if not). If the recovery rate is zero, then:

∑ log (1 − p j + p j exp (aj s))

ulation. It can be applied particularly easily in the fields of credit, operational n
and insurance risk, and our discussion is slanted in that direction. Here, we K Y (s) = (3)
j =1
develop the technique to find the sensitivity of VAR to asset allocation.3 The
speed and accuracy of the saddle-point method facilitate rapid optimisation When dependent variables are added, it is no longer correct to add their
of return-on-VAR. We shall treat that problem in forthcoming articles. KGFs. In that case, the analogue of expression (3) is more complicated. In
The central assertions of this article are: Martin, Thompson & Browne (2001), we discuss a conditional indepen-

VAR, and its sensitivity to asset allocation, can be obtained by looking at dence model of correlation that allows the KGF of the sum to be calcu-
the cumulant generating function and its various derivatives at the saddle- lated easily: if (Xj) are conditionally independent on a random variable W
point, the position of which depends on the confidence level of the VAR. that takes value k with probability hk, then:

This analysis can be performed very quickly and easily for loan portfo-
( )
n
lios, and requires no Monte Carlo simulation. K Y (s) = log ∑ h k ∏ expK X j aj s k (2a)
j =1

The sum of all the saddle-point VAR contributions is the VAR. k

For normal distributions, our method is identical to risk analysis in the 1 Provided that the risk measure satisfies the simple condition of 1-homogeneity (see below)
mean-variance framework, originally studied by Markowitz (1952). 2 That is, the relationship Risk(A + B) ≤ Risk(A) + Risk(B) does not necessarily hold. For a

For non-normal distributions, it is a generalisation of mean-variance, re- fuller discussion, see, eg, Artzner et al (1999)
producing that method for quantiles close to the mean. 3 Incidentally, our methods could be extended to include expected shortfall

WWW.RISK.NET ● AUGUST 2001 RISK 99

Cutting edge l Portfolio risk management

in which K X j(s|k) ≡ logE[exp(sX j)|W = k]. For the example of a loan port- the VAR is also multiplied by β. Examples of 1-homogeneous risk mea-
folio, this reduces to: sures include the mean, the standard deviation, the expected shortfall for
some quantile and any constant multiple, or sum, difference, maximum or
( ( ))
n
K Y (s) = log ∑ h k ∏ 1 − p jk + p jk exp a j s (3a) minimum of other 1-homogeneous functions. This homogeneity property
k j =1
leads to the following well-known decomposition formula of Euler:
where pjk is the probability of the jth asset defaulting given that W = k. n ∂Tp
Referring to equation (1), the saddle-point approximation (see, eg, Mar- ∑ a j ∂a = Tp (9)
j =1 j
tin, Thompson & Browne, 2001, and references therein) consists of ap-
proximating the term in the exponential as a Taylor series around the point Hence the sum of the sensitivities of Tp to infinitesimal fractional changes
at which the term in the exponential is stationary (the saddle-point), trun- in each allocation is equal to Tp itself. Equivalently, the sum of the VAR
cating at the quadratic term and doing the resulting Gaussian integral. The contributions is just the VAR. Now K Y(s) also has a homogeneity proper-
saddle-point is ^t such that: ty, in that it depends on s and a only through their product. It obeys a
K ′Y ˆt = t () (4)
similar formula to (9):
n a ∂K ∂
∑ s ∂aY (s) = ∂ s K Y (s)
j
in which the prime denotes differentiation. As K Y is convex, there is on the (10)
j =1
real axis a unique saddle-point ^
t for each point t ‘in’ the distribution of Y. j

The lowest-order saddle-point approximation to the tail probability of Y is, The right-hand side of this expression, evaluated at the saddle-point (ie,
for t > E[Y]: at s = ^
t ), is just Tp(a). Hence:

( () (ˆt))Φ − ()
a j ∂K Y
()
n
P (Y > t ) ≈ exp K Y ˆt − t ˆt + t K ′′Y
1 ˆ2 ˆt2K′′Y ˆt  (5) ∑ t̂ = Tp (a) (11)
2  j =1 t ∂a j
ˆ

with Φ denoting the cumulative normal distribution function. The sum of the saddle-point VAR contributions is exactly the VAR, de-
The correspondence (4) is fundamental because it enables us to relate spite the fact that equation (8) is an approximation. Notice also that:
a quantile t of the distribution, or a part of the distribution local to that
quantile, to the KGF evaluated at, or in the vicinity of, the saddle-point ^
t. a j ∂K Y
(s) =
( )
a j ∫ x j exp ∑i=1 s a i xi dP X (x )
n

( )
(12)
s ∂a j
∫ exp ∑i=1 s a i xi dP X (x )
We have found that this approximation works particularly well in the tail n

of the distribution. It is quite easy to use the saddle-point method in prac-

tice, provided that one can calculate the portfolio KGF and its first two de- in which dPX(x) is the probability density of X = (X1, ... , Xn). Accordingly,
rivatives KY(s), K′Y(s),K′′Y(s) for any value of s. the VAR contribution of the jth asset must lie between aj min X j and a j max
X j. For a loan portfolio, this means that it is positive and does not exceed
VAR contributions the jth exposure, so (8), on division by aj, lies between zero and one and
Now that we have an analytical approximation to VAR we can calculate thus has the rather nice interpretation as a risk-adjusted default probability.
the sensitivity of VAR to an infinitesimal change in asset allocation. It is worth noting that the accuracy of the saddle-point approximation

Derivation. The tail probability for a given point, t, on the distribution is not in fact fundamental in this article. Our theory rests on equations (4)
is given by (1), with X replaced by Y = Σ ja jX j. We write a = (aj)nj= 1 for and (10–12), which are exact. Our analysis below will use only those. More-
the vector of asset allocations. If we differentiate this expression with re- over, the ideas can be developed further using utility theory (see “Risk
spect to aj we obtain (note that K Y depends on a): aversion and utility theory” below).
+i ∞
∂  1 ∂ K Y ∂ t 
∂a j
P (Y > t ) =
1
∫ 
2π i −i ∞, ( 0 + )  s ∂ aj
(s) −
∂ aj 
(
 exp KY (s) − st ds ) (6)
Analysing the VAR contributions

Asset losses that follow a normal distribution. If the constituent loss
variables (X j) are multivariate normal, so that Y is also normal, then (8) is
Now we wish to find the effect of infinitesimally altering aj while keep- exact. This is because K Y(s) is then quadratic in s: K Y(s) = E[Y]s + ½V[Y]s2,
ing the tail probability constant. In other words, we wish to find how VAR, so the integrals in the numerator and denominator can be carried out ex-
denoted Tp for some fixed tail probability p, depends on the jth asset al- actly. Let us now apply this to a general portfolio of normally distributed
location. Then the left-hand side of (6) vanishes and we have: losses. The correspondence t = K′Y(^ t ) can be written t = E[Y] + V[Y] ^ t.
If our risk measure is the upper p-quantile of the distribution of Y, so:
∂Tp + i∞
∂K Y (
exp K Y (s) − sTp ) + i∞
(t − E [Y])/
∂a j
= ∫ ∂ a (s) s
ds ∫ (
exp K Y (s) − sTp ds (7)) V [Y ] = Φ −1 (1 − p)
− i ∞, (0+ ) j − i∞
then we have the following expression for ^
t:
t̂ = Φ −1 (1 − p) / V [Y ]
If we approximate the integrals in the numerator and denominator around
(13)
the saddle-point ^
t , and multiply both sides by aj, we obtain:
∂Tp a j ∂K Y If Y = Σ ja jX j, each X j is normally distributed with mean µj and standard
aj ≈ (8) deviation σj, and the correlation between Xj and Xk is ρjk, then E[Y] =
∂a j ˆt ∂ aj
s =ˆt Σjaj µj and V[Y] = Σ j, kaj ak ρjk σj σk. The jth VAR contribution is therefore:
which is our key result. The left-hand side is the sensitivity of VAR to an in-
finitesimal fractional change in asset allocation, and we call it the (exact)
a j ∂K ˆ
ˆt ∂a j ()
t = a j µ j + ˆt ∑ aj ak ρ jk σ j σ k
k
(14)
VAR contribution. The right-hand side is the saddle-point VAR contribution
relative to the point on the loss distribution determined by ^ t . Equation (8) This is the same concept as the risk contribution defined by Bürgisser
shows that we can approximate the VAR contribution if we can differenti- et al (1999), in which the variance is divided amongst the risk contribu-
ate KY with respect to asset allocation. For a loan portfolio, this can easily tors. It is easily seen that the expression (14), when summed over j, does
be done in closed form. give t (the VAR). We thereby recover the mean-variance result.

The sum of the VAR contributions is the VAR. The VAR Tp(a) is a ‘1- We now consider in this and the next subsection how the VAR contribu-
homogeneous’ function of the asset allocations a = (aj)nj= 1, which means tion depends on asset allocation. It is important to understand that here we
that if all the allocations are multiplied by some positive number β then are talking about risk contributions in a static portfolio, ie, looking at the risk

100 RISK AUGUST 2001 ● WWW.RISK.NET

contributions of the existing assets and seeing how these depend on their
sizes (we call this a passive analysis). By an active analysis, we mean that A. Default models for portfolio
we are altering the sizes of one or more assets and changing the character-
istics of the portfolio. This will lead to different conclusions, which we will Uncorrelated Correlated
discuss in later work. In a passive analysis,^ t , which depends on the whole 20% 40% 30% 10%
portfolio, is fixed. In an active analysis, we must take into account its de- Rating Default probability Default probability
pendence on the asset allocations. A 0.36% 0.00% 0.10% 0.40% 2.00%
In a passive analysis, (14) is quadratically dependent on each aj. Let us B 0.90% 0.00% 0.25% 1.00% 5.00%
rewrite the right-hand side of (14) as: C 3.60% 0.00% 1.00% 4.00% 20.00%
The mean default rates are the same for the dependent model as for the
 
a j  µ j + ˆt ∑ ak ρ jk σ j σk  + ˆta 2j
independent. These numbers are for the purposes of illustration only
 k (≠ j) 

The first term is linear in aj and predominates if the jth risk is small compared 1. Loss exceedance curve for the test portfolio
with the other risks in the portfolio. The second term puts a non-proportional
penalty on large risks. This is because a large risk brings no diversification to a. Independent loss events b. Dependent loss events
the portfolio. For example, one asset of standard deviation five gives the same 1 1
Saddle-point Saddle-point
loss distribution as 25 (not five) completely independent assets of standard Monte Carlo Monte Carlo
deviation one. This observation leads to a related point on correlation. If an 0.1 0.1

Tail probability
Tail probability
asset is not correlated with the rest of the portfolio, its risk contribution will
be smaller than if it is positively correlated, for in the presence of positive 0.01 0.01
correlation there is less diversification to be had by cutting one large expo-
sure into several small ones. Indeed, in the extreme case when there is 100% 0.001 0.001
correlation, five assets of size one are no better than one of size five, and the
risk contribution for one asset of size one is one-fifth that of an asset of size 0.0001 0.0001
five. Negative correlation can cause a VAR contribution to be negative. 0 2 4 6 8 10 0 2 4 6 8 10 12

Non-normal distributions. Real loss distributions are not normal, and Loss Loss
the standard deviation does not capture the extreme risk well enough. For
example, suppose that loan A is £10 million to a counterparty with a de- tion model are given in table A. The correlation model is that condition-
fault probability of 10% and loan B is £30 million to a counterparty with a ally on a latent (or ‘mixing’) variable W, which takes four states with prob-
default probability of 1%. The mean losses are £1 million and £300,000 re- abilities given in the top row of the table, default events occur independently
spectively, and the standard deviations are both roughly £3 million. To an and with probabilities given in the table.
investor who assesses risk by variance, the loans are equally good. But a Figure 1a shows the loss exceedance curve for the portfolio, assuming
portfolio manager with £20 million capital will disagree. If loan A defaults independence (note the excellent agreement between saddle-point and
they lose half the capital, but if loan B defaults their fund is bankrupt. A Monte Carlo). Figure 2a shows the saddle-point VAR contributions calcu-
consequent criticism of mean-variance analysis is that it leads to over-con- lated for different quantiles of the loss distribution (loss = 1, 3, 6). Note
centration, and so (awkward-looking) concentration constraints are tacked first that the VAR contribution of each asset increases as we go further into
on. The saddle-point VAR contribution obviates these by measuring risk in the tail (the yellow bars are higher than the red, which are higher than the
a more appropriate way. We can also see that the portfolio manager with blue). However, the VAR contributions of the assets with the biggest ex-
£20 million capital will want to use a risk measure that incorporates a high- posures increase more rapidly. Start by looking at the VAR contributions
er degree of risk aversion than a mean-variance measure, because of their for assets 1–5 and assets 6–10 (in both of these sequences the asset allo-
concern about the fund going bankrupt. cations are in arithmetic progression). The mean total loss is 0.31. For a
If we examine the saddle-point VAR contribution in more depth we ob- quantile fairly close to this, such as loss = 1 (blue bars), the VAR contri-
serve that it inherently accounts for the extra risk that arises from asset size bution is roughly proportional to the square of the asset allocation (see
^
and which the mean-variance analysis misses. In general, K Y (t ) is not qua- figure 2a). The risk contribution of asset 5 (0.21) is in fact 36 times the VAR
^
dratic in t . Expanding it in a Taylor series around the origin we have4: contribution of asset 1 (0.0059), when the asset allocation is five times larg-
()
K Y ˆt = E [Y ]ˆt + 1 V [Y ]ˆt 2 + 1 H3 [Y]ˆt 3 + 1 H4 [Y]ˆt 4 + ...
2 3! 4! (15)
er. Further into the tail (loss = 6, yellow bars), the variation of risk contri-
bution with asset allocation is much stronger, as we predicted above (see
in which H3[Y] and H4[Y] are the third and fourth cumulants of Y (in “Non-normal distributions”). The risk contribution of asset 5 (2.45) is 225
essence the skewness and kurtosis). For^ t ≈ 0, corresponding to points on times the risk contribution of asset 1 (0.011). In other words, we find that
the distribution near the mean, we have the mean-variance result as be- the tail of the distribution is almost entirely dominated by the assets with
fore. However for larger ^ t (corresponding to points further into the tail) the highest exposures. For loss = 1 (blue bars), assets 5 and 10 contribute
the higher-order moments come into play. As higher powers of ^ t carry roughly equally, with assets 4, 9, and 11–15 also being quite significant.
with them higher powers of the exposure parameters, the VAR contribu- For loss = 6, asset 5 is by far the most significant, followed by 4 and 10,
tion is no longer a quadratic function of allocation. The mean-variance re- and the VAR contributions of the others are trivial. Note in passing that the
sult is therefore part of the more general theory described here. risk contributions of assets 11–15 are exactly equal, as they should be (this

Example. We consider a portfolio of 15 assets (loans) with the follow- is another advantage of analytical techniques over simulation-based ones).
ing ‘ratings’ and exposures: Figures 1b and 2b are analogous to 1a and 2a, but with correlation in-

Assets 1–5: rating A, exposures 1.0, 2.0, 3.0, 4.0, 5.0. troduced. The most important difference is that the smaller assets now con-

Assets 6–10: rating B, exposures 0.7, 1.4, 2.1, 2.8, 3.5. tribute more to the risk than they did for a model of independent defaults,

Assets 11–15: rating C, exposures 0.9, 0.9, 0.9, 0.9, 0.9. 4 K Y is analytic at the origin, and hence has a Taylor expansion (though the radius of
We shall analyse this portfolio, assuming first independence and then de- convergence is generally not infinite). As the expansion is for the purposes of illustration
pendence between default events, and calculate the VAR contributions only, we do not go into this further – but note that some divergent series can be summed
using the methods described here. The default probabilities and correla- in any case (Euler’s method, Cesaro summation, Padé approximants, etc)

WWW.RISK.NET ● AUGUST 2001 RISK 101

Cutting edge l Portfolio risk management

The generalisation of our theory lies in the following observations (for

2. Saddle-point VAR contributions brevity we write γ for GU(λ, Y)):

∂γ E  YU′ (γ − λ Y ) E[(∂γ/∂λ − λ Y ) U′′ (γ − λ Y )]

2
a. Independent loss events ∂ 2γ
2.5 = , =− (17)
Loss = 6 (~99.9% confidence)
∂λ E U′ (γ − λY ) ∂λ 2
E [ U′ (γ − λ Y )]
2.0 Loss = 3 (~98% confidence)
The first expression shows that ∂GU(λ, Y)/∂λ is a weighted average of val-
VAR contribution

Loss = 1 (~75% confidence)

1.5 ues of Y, and hence it must correspond to a quantile of Y (ie, if Y lies be-
tween Ymin and Ymax with probability one, then ∂GU(λ, Y)/∂λ also lies in
1.0 this range, for any λ). It is also apparent that ∂GU(λ, Y)/∂λ = E[Y] when
λ = 0. The second expression shows that, if U is concave (so that U′′ <
0.5 0; this will be so for a risk-averse participant), then ∂GU(λ, Y)/∂λ is an in-
creasing function of λ. So for a concave utility function we have that ∂GU(λ,
0 Y)/∂λ is a point in the distribution of Y, which corresponds to the mean of
1 2 3 4 57 8 9 10 11 12 13 14 15
6
Asset number Y when λ = 0 and increases as λ increases.
b. Dependent loss events Now suppose, as we have done in the first part of the article, that Y is
2.5
Loss = 8 (~99.9% confidence) a weighted sum of other random variables: Y = Σ j a j X j. As GU(λ, Y) depends
2.0 Loss = 4 (~98% confidence) on λ and Y only through their product, it obeys the homogeneity relation:
VAR contribution

Loss = 1 (~75% confidence)

∂GU (λ, Y ) a j ∂G U (λ, Y )
1.5
∂λ
= ∑λ ∂aj
(18)
j

1.0 We identify the jth summand in the right-hand side as a risk contribution
of the jth asset, which we shall call the U-risk contribution. The sum of the
0.5 U-risk contributions is the Y-quantile determined by λ (and U). Equation
(12) and the ensuing remarks also generalise, for the jth U-risk contribution
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 is E[aj X j U′(γ – λY)]/E[U′(γ – λY)]. All this is the same exposition as before
Asset number but in a more general setting. The parameter λ plays the same role as the
The VAR contributions of each of the 15 assets in the portfolio, at three different saddle-point s (or ^ t ) did before. In this context, λ has a natural interpreta-
points on the distribution. The bars are superimposed rather than stacked tion as a risk-aversion parameter. When λ = 0 we are risk-neutral, and when
λ > 0 we are risk averse, with increasing risk aversion as λ increases. ■

as we discussed earlier. Richard Martin and Kevin Thompson are members of the fixed-income

General representation of the VAR contribution. A nice argument by capital management team at BNP Paribas in London. We thank
Gouriéroux, Laurent & Scaillet (2000) is that the VAR contribution of the jth professor Olivier Scaillet for helpful discussions. A longer version of this
asset in a portfolio is the expected loss from that asset conditional on the article is available on request. E-mail: richard.martin@bnpparibas.com
portfolio loss being equal to the VAR, ie, E[a j X j|Y = t]. From this it is clear and kevin.thompson@bnpparibas.com
that the VAR contributions total E[Y|Y = t], ie, t, and also that the jth VAR Comments on this article may be posted on the technical discussion forum on the Risk
website at http://www.risk.net
contribution lies in [aj min X j, aj max X j]. It also shows why small uncorre-
lated assets contribute so little to credit VAR. For the portfolio loss to be large,
many small assets need to default, and this is very unlikely. But when they REFERENCES
are correlated, it is more likely, and then their risk contributions increase.
Artzner P, F Delbaen, J-M Eber and D Heath, 1999
Coherent measures of risk
Risk aversion and utility theory Mathematical Finance 9, pages 203–208
We have mentioned risk aversion and now we show that equation (9) can Bühlmann H, 1970
Mathematical methods of risk theory
be generalised using utility theory (for an introduction, see, eg, Bühlmann, Springer-Verlag
1970, or Eatwell, Milgate & Newman, 1990). The theory as described so Bürgisser P, A Kurth, A Wagner and M Wolf, 1999
far follows if we use an exponential utility function. Let U be a utility func- Integrating correlations
tion, by which we mean that it is strictly increasing and continuous. (It is Risk July, pages 57–60
Eatwell J, M Milgate and P Newman, 1990
commonplace, but not necessary, to have U(0) = 0). We can think of U(x) The new Palgrave: utility and probability
as being how satisfied we are with receiving an amount x. If U(x) = x for Macmillan
all x then we are risk-neutral, which means that we are indifferent to play- Gouriéroux C, J-P Laurent and O Scaillet, 2000
Sensitivity analysis of values at risk
ing any game with zero expected gain (or loss). If U is concave then play- Journal of Empirical Finance 7, pages 225–245
ing such a game is unattractive because the down-side outweighs the Leibowitz M and R Henriksson, 1989
up-side. In that case, we require a premium to play the game, which we Portfolio optimization with shortfall constraints: a confidence-limit approach to
modelling downside risk
now quantify using the following construction. For a random variable Y Financial Analysts Journal, March-April, pages 34–41
describing the loss of some financial instrument (Y > 0 means that we Litterman R, 1996
lose) let us define the following generating function: Hot spots and hedges
Journal of Portfolio Management 22 (special), pages 52–75
GU (λ, Y ) = γ s.t. E U (γ − λ Y ) = U(0) (16) Markowitz H, 1952
Portfolio selection
The assumptions made about U ensure that GU is well defined. When U(y) Journal of Finance 7, pages 77–91
= 1 – e–y we just have the KGF, and GU(λ, Y) = lnE[eλY]. The quantity Martin R, K Thompson and C Browne, 2001
Taking to the saddle
GU(λ, Y)/λ has a natural interpretation as the premium required to make Risk June, pages 91–94
us indifferent between these two strategies: (i) accept the risk Y, and re-
ceive the premium; (ii) do nothing.

102 RISK AUGUST 2001 ● WWW.RISK.NET