Professional Documents
Culture Documents
1
0
(1 G
L
(u)) du
a safe insurance premium can be dened by
[L] =
1
0
(1 G
L
(u)) d(u), (1)
1
Department of Statistics and Decision Support Systems, Universitaetsstrasse 5, Uni-
versity of Vienna, 1090 Wien-Vienna, Austria; e-mail: georg.pflug@univie.ac.at
1
where is function mapping [0,1] to [0,1], such that
(u) u for u [0, 1]. (2)
The condition (2) guarantees that the premium is not smaller than the
expectation. However one usually considers more specic functions .
Denition. Distortion functions. A function is called a distortion
function, if it is is monotonic, left continuous and satises (u) u, (0) = 0,
(1) = 1. If is a distortion function,
[L] =
1
0
G
1
L
(p) d(1 p). For
a distortion function , d(1 p) is a probability distribution on [0,1].
While the expectation of L equals the expectation of G
1
L
(U) with a uniform
[0,1] U, the insurance premium
1
0
G
1
(p) dH(p).
If the random variable Y has distribution function G, we also use the notation
/
H
[Y ], if no confusion may occur. We allow the distribution G to have jumps
to cover nite sample situations.
The following formula for partial integration is well known
E[Y ] =
u dG(U) =
G(u) du +
0
(1 G(u)) du.
The next Lemma generalizes this formula. Notice that we allow H to
have jumps at values G(u), for which G(u) and G(u) = lim
vu
G(v) are
dierent. A similar formula is found in [13].
Lemma 1. Let G (
H
. Then
G
1
(p) dH(p) =
H(G(u)) du +
1
0
G
1
(p) dH(p) =
u dH(G(u)).
Notice rst that
1
0
G
1
(p) dH(p) =
G
1
(G(u)) dH(G(u)).
Now notice that G
1
(G(u)) u and G
1
(G(u)) < u only if there is a v < u
such that G(v) = G(u) and consequently H(G(v)) = H(G(u)). Thus setting
A = u : G
1
(G(u)) = u, we see that
G
1
(G(u)) dH(G(u)) =
A
G
1
(G(u)) dH(G(u))
=
A
u dH(G(u)) =
u dH(G(u))
The partial integration formula for Stieltjes integrals is
[a,b]
K
1
(u+) dK
2
(u)+
[a,b]
K
2
(u) dK
1
(u) = K
1
(b+)K
2
(b+)K
1
(a)K
2
(a).
An application of this formula gives
u dH(G(u)) =
(,0]
u dH(G(u)) +
(0,)
u dH(G(u))
=
(,0]
u dH(G(u))
(0,)
u d[H(1) H(G(u))]
=
(,0]
H(G(u)) du + 0 H(G(0)) () H(G())
+
(0,)
[H(1) H(G(u))] du
[H(1) H(G())] + 0 [H(1) H(G(0))]
=
H(G(u)) du +
1
0
G
1
(p) dp =
G(u) du +
0
(1 G(u)) du.
Setting H(p) = 1l
[,1]
(p) one gets the value-at-risk
V@R
[Y ] = G
1
().
Setting H(p) = min(p/, 1) one gets the average value-at-risk
AV@R
[Y ] =
1
0
G
1
(p) dp
=
min(G(u)/, 1) du +
0
max(1 G(u)/, 0) du.
Several other names have been proposed for this functional, such as con-
ditional value-at-risk (Rockefellar and Uryasev [18]), expected shortfall
(Acerbi and Tasche [2]) and tail value-at-risk (Artzner et al. [4]). The
name average value-at-risk is due to Follmer and Schied ([10]).
If H has a density, i.e. H(p) =
p
0
h(q) dq the pertaining class coincides
with Yaaris dual functionals, see [20].
5
3 Dual representations
If U is uniform [0,1], then G
1
(U) has the same distribution as Y . The
converse is not true: If G has jumps, then G(Y ) is not uniformly distributed,
it is in fact stochastically larger than a uniform [0,1] distribution, PG(Y )
p PU p.
To correct the distribution of G(Y ) towards a uniform distribution, let
| be the countable set of all jump points of G. For u |, let (V
u
) be a
collection of independent random variables, which are Uniform [0, G(u)
G(u)] distributed and which are independent of Y . For u / |, set V
u
= 0.
By composition, one may form the random variable V
Y
.
Lemma 2. G(Y ) V
Y
is uniformly [0,1] distributed and G
1
(G(Y )
V
Y
) = Y a.s.
Proof. Let 0 < p < 1. We have to show that PG(Y ) V
Y
p = p.
One may nd an u such that G(u) p G(u). If G(u) = G(u), then
V
u
= 0 and PG(Y ) p = p. Otherwise G(Y ) V
Y
p i Y < u or Y = u
and V
u
G(u) p, i.e. PG(Y ) V
Y
p = G(u) + [G(u) G(u)][p
G(u)]/[G(u) G(u)] = p.
To prove the second assertion, notice that conditional on Y = u, G(u)V
u
lies in the interval [G(u), G(u)] and with probability 1 in (G(u), G(u)].
However in the latter interval, G
1
equals u. 2
Lemma 3. Suppose that H(p) =
p
0
h() d. Then
/
H
[Y ] = E[Y h(G(Y ) V
Y
)].
Proof. Let U = G(Y ) V
Y
. Then by Lemma 2, U is uniform[0,1] and
G
1
(U) = Y a.s. and therefore /
H
(Y ) = E[G
1
(U)h(U)] = E[Y h(G(Y )
V
Y
)]. 2
The next result shows the dual representation of /
H
.
Proposition 1. (see also Pug [16])
/
H
[Y ] = infE[Y Z] : Z = h(U), where U is uniformly [0,1] distributed.
which is the same as
/
H
[Y ] = infE[Y Z] : Z
CXD
Z
, where Z
= h(U),
with U uniformly [0, 1] distributed .
The inmum in (4) is attained for Z = h(Y V
Y
).
Proof. By virtue of Lemma 3, /
H
[Y ] infE(Y Z) : Z = h(U),U Uniform[0,1].
To prove the other inequality, notice Hoedings Lemma rst ([11], see also
[15]):
Cov(Y, Z) =
G
Y,Z
(u, v) G
Y
(u) G
Z
(v) dudv
6
where G
Y,Z
is the joint distribution and G
Y
, G
Z
are the marginal distributions
of the vector (Y, Z). If we x the marginals, the covariance Cov(Y, Z) (and
equivalently E(Y Z)) is minimized for the joint distribution, which is the
lower Frechet bound G
Y,Z
(u, v) = max(G
Y
(u) + G
Z
(v) 1, 0), i.e. if Y and
Z are antimonotone coupled.
Since h is nonincreasing and G is nondecreasing, it is easy to see that
the variables Y and h(G(Y ) V
Y
) are indeed antimonotone coupled and
therefore (4) is the lower bound of E(Y Z) : Z = h(U).
For the second representation, we use a result by Dentcheva and Ruszczyn-
ski ([9]):
conv Z : Z
L
= V = Z : Z
CXD
V .
Here Z
CXD
V i E[f(Z)] E[f(V )] for all convex functions f. 2
Since the functions H(p) = min(p/, 1) are the extremal elements in
the set of all concave, monotonic probability distribution functions on [0,1],
there is a Choquet representation for distortion functionals (This particular
representation is not due to Choquet [6], but in his spirit).
Proposition 2. Choquet representation. (Acerbi [1], [13]). Suppose
that H concave, i.e. has a representation as H(p) =
p
0
h() d with a
nonincreasing h(p). Then
/
H
G =
1
0
AV@R
G dK() (4)
for a monotonically increasing K, which satises K(0) = 0, K(1) = H(1).
Thus for concave H, the distortion functionals are mixtures of AV@Rs.
Proof. We may assume that the nonincreasing function h dened on
[0, 1] is continuous from the left and has the representation
h(p) =
(0,1]
1
1l
[0,]
(p) dK(). (5)
To show (5), let h
n
be the largest nonincreasing left continuous stepfunction,
which is dominated by h and which jumps only at dyadic rational points
k/2
n
. Clearly, one may write h
n
(p) =
(0,1]
1
1l
[0,]
(p) dK
n
() with K
n
(0) = 0,
K
n
(1) =
1
0
h
n
(p) dp H(1). The sequence dK
n
is a sequence of bounded
measures on [0,1], which has a weak limit dK with K(0) = 0, K(1) =
7
lim
n
1
0
h
n
(p) dp = H(1). Now
/
H
G =
1
0
G
1
(p) dH(p) =
1
0
G
1
(p)h(p) dp
=
1
0
G
1
(p)
(0,1]
1
1l
[0,]
(p) dK() dp
=
(0,1]
1
1
0
G
1
(p)1l
[0,]
(p) dp dK()
=
(0,1]
0
G
1
(p) dp
dK() =
(0,1]
AV@R
(G) dK().
2
Proposition 3. Properties of distortion functionals. (see also De
Giorgi [7])
Distortion functionals /
H
[Y ] =
1
0
G
1
(u) dH(u) are
(i) version independent, i.e. they depend on the random variable Y only
through its distribution function G
(ii) linear in the quantile function G
1
and hence comonotone additive,
(iii) positively homogeneous, i.e. /
H
[cY ] = c/
H
[Y ] for c > 0
(iv) isotonic w.r.t. rst order stochastic dominance (
FSD
).
If dH is a probability measure (i.e. H(1) = 1) then /
H
is
(iv) translation equivariant, i.e. /
H
[Y + a] = /
H
[Y ] + a.
If H is concave (i.e. H(p) =
p
0
h(p) dp) for a nonincreasing h, then
(vi) concave, i.e. /[Y
1
+(1)Y
2
] /
H
[Y
1
]+(1)/[Y
2
] for 0 1,
(vii) isotonic w.r.t. second order stochastic dominance (
SSD
).
(viii) If H is concave and h L
q
[0, 1], then Y /
H
(Y ) is continuous in the
L
p
sense, where 1/p + 1/q = 1.
Proof. (i) is obvious. (ii). Recall that a functional T is called comonotone
additive if for two comonotone random variables Y
1
and Y
2
, T[Y
1
+ Y
2
] =
T[Y
1
] + T[Y
2
]. Two random variables are comonotone, if the joint distribu-
tion G
1,2
(u) is related to the marginal distributions G
1
, G
2
by G
1,2
(u, v) =
min(G
1
(u), G
2
(v)). For comonotone variables, the quantile function of Y
1
+Y
2
8
is the sum of the marginal quantile functions: G
1
1
(u) + G
1
2
(u). (iii). No-
tice that cY has quantile function cG
1
(u) for positive c. (iv) Y
1
is dom-
inated in the rst order sense if G
1
(u) G
2
(u), which is equivalent to
G
1
1
(u) G
1
2
(u) for all u. (iv) The variable a + Y has quantile function
a + G
1
(u).
(vi). If H is concave then /
H
[Y ] has a dual representation as given in
Proposition 1. Thus /
H
is the inmum of linear (in Y ) functions and hence
concave. Finally, for concave H, /
H
(Y ) has by Proposition 2 a representation
as a mixture of AV@Rs. The assertion (vii) is proved, if we show it holds
true for all AV@R
[Y ] = maxa
1
E([Y a]
) : a R.
For xed a, the function Y a
1
E([Y a]
[ya]
is monotonic
and concave. Thus also the maximum is monotonic w.r.t. second order
stochastic dominance and proves (vii). Finally, notice that because of (4) /
H
is upper semicontinuous. If /
H
is not continuous, one may nd a sequence
Y
n
converging to Y in L
p
sense, such that
/
H
(Y ) limsup
n
/
H
(Y
n
) +
for some 0. Represent /
H
[Y
n
] = E[Y
n
h(U
n
)] with U
n
a Uniform[0,1]
variable. Notice that |h(U
n
)|
q
=
h (say) for all n. If |Y Y
n
|
n
< /
h, then
/
H
(Y ) E[Y h(U
n
)] E[Y
n
h(U
n
)] +
h|Y
n
Y |
n
< /
H
(Y
n
) + ,
a contradiction. 2
Notice that (iii) and (vi) imply that /
H
is superadditive
/
H
[Y +
Y ] /
H
[Y ] +/
H
[
Y ].
Remark. Kusuoka [14] has shown that any version independent func-
tional T, which is positively homogeneous, translation equivariant and coin-
cides with its concave bidual can be represented as
T[Y ] = inf
1
0
AV@R
(Y ) dM() : M /,
where / is a set of probability measures on [0,1]. If T is comonotone
additive, then /contains only one probability measure and T has a Choquet
representation of the form (4). In Proposition 2 we have derived this result in
9
a direct manner by relating the mixture measure M to the distortion function
H.
Proposition 3 has an inverse given by the following result.
Proposition 4. Let /G be version independent functional, which is
nite for bounded distributions. If / is positively homogeneous, monotonic
w.r.t. FSD and comonotone additive, then it is of the form /G =
1
0
G
1
(u) dH(u).
Proof. W.l.o.g we may assume that /(1l) = 1 and /(0) = 0. By positive
homogeneity /(c) = c for c 0. By /(c) + /(c) = /(0) = 0, also
/(c) = c for c < 0. Let G
p
(u) be the distribution which has point masses at
0 (with probability p) and 1 (with probability 1p). Let H(p) = 1/(G
p
).
Then H is monotonically nondecreasing, satises H(0) = 0, H(1) = 1. Let
/
(G) =
1
0
G
1
(p) dH(p). Notice that any discrete distribution can be
represented as a comonotone sum of discrete distribution with just two point
masses. This implies that / coincides with /
(G
2
) /
(G
1
) . Since also /(G
1
) /(G) /(G
2
) one sees
that in fact /(G) = /
(G). 2
4 Portfolio optimization
In this section, we consider a one-period portfolio optimization problem,
where the objective is to maximize the expected return under the constraint
that a distortion acceptability functional does not fall below a prespecied
level. Since the negative distortion risk functional has the interpretation as
required risk capital (see [3]), one may equivalently say that the portfolio
optimization problem seeks for maximizing the return for a given maximal
risk capital.
Let = (
(1)
, . . . ,
(M)
) a row vector of random portfolio returns dened
on some probability space (, B, P) and x = (x
1
, . . . , x
M
)
1l = 1
x 0
(6)
Here 1l is the column vector of length M with entries 1. (6) is a linear
optimization problem under a convex constraint. To derive the necessary
conditions for optimality, a characterization of the supergradient set
/
H
[Y ] = Z L
q
: /
H
[
Y ] /
H
[Y ] +E[(
Y Y ) Z] for all
Y L
p
of /
H
at Y is needed. Assume that Y L
p
(), p 1 and q = p/(p 1) for
p > 1 and q = for p = 1. Assume further that H(u) =
u
0
h(v) dv with
h L
q
[0, 1].
Lemma 4. The supergradient of /
H
is /
H
[Y ] = h(G(Y ) V
Y
),
where V
Y
runs through all random variables on , for which the conditional
distribution of V
Y
given Y = u is uniform [0, G(u) G(u)]. The supergra-
dient set is a singleton if and only if G is continuous, given that is rich
enough to carry a uniform [0,1] distribution.
Proof. Let Z = Z = h(U), where U is uniformly [0,1] distributed.
By (4) /
H
[Y ] = E[Y h(G(Y )V
Y
)] = infE[Y Z] : Z Z. Since /
H
[
Y ]
E[(
Y h(G(Y )V
Y
)], one sees that h(G(Y )V
Y
) /
H
[Y ] for all versions of
V
Y
. Conversely, let
Z /
H
[Y ]. If
Z is not in Z, then there is an > 0 and
a
Y such that E[
Y
Z] + infE[
Y Z] : Z Z = /
H
[
Y ]. Consequently
/
H
[Y +
Y ] /
H
[Y ] +E[
Y
Z] /
H
[Y ] +/
H
[
Y ]
/
H
[Y +
Y ] ,
a contradiction. Thus the rst assertion has been proved. If G is continuous,
then V
Y
= 0 and thus the supergradient is unique. If however G has jumps,
then h(G(Y ) V
Y
) may be realized in dierent ways and is thus not unique.
2
Lemma 5. If H has a nonnegative density h L
q
[0, 1] and G is con-
tinuous, then
Y /
H
[
Y ] is strongly dierentiable in L
p
sense at Y . The
derivative is
/
H
[Y ] = h(G(Y )).
Consequently also the mapping x /
H
[ x] is dierentiable at Y
x
=
x
with derivative
x
/[ x] = E[ h(G( x)]].
11
Proof. Let Z = /
H
(Y ). If /
H
is not dierentiable at Y , then there is
a sequence Y
n
converging to Y in L
p
sense such that
/
H
(Y
n
) /
H
(Y ) E[(Y
n
Y )Z] |Y
n
Y |
p
. (7)
for some > 0 and all n. Let Z
n
/
H
(Y
n
). Then
/
H
(Y ) /
H
(Y
n
) E[(Y Y
n
)Z
n
] 0 (8)
(7) and (8) together give
|Y
n
Y |
p
|Z
n
Z|
q
E[(Y
n
Y )(Z Z
n
)] |Y
n
Y |
p
. (9)
By uniform integrability, the sequence (Y
n
) must have a weak cluster point,
say Z
. By (9), Z , = Z
Y ) /(Y
n
) +E[(
Y Y
n
)Z
n
],
for all
Y , using the L
p
-continuity of /
H
(see Proposition 3 (viii)) it follows
that letting n
/
H
(
Y ) /(Y ) +E[(
Y Y )Z
.
Thus Z
Maximize E[] x
subject to
/
H
[ x] = q
x
1l = 1
x 0
Under the assumption of dierentiability, the necessary Karush-Kuhn-
Tucker (KKT-) conditions for this problem are:
E() +
x
/
H
[ x] + 1l + = 0
(/
H
[ x] q) = 0
(x
1l 1) = 0
m
x
m
= 0 for m = 1, . . . , M
0
Suppose that the optimum lies in the interior of X, i.e. asset not present
in the portfolio are neglected. Interpreting /
H
as the necessary risk capital,
the KKT conditions can be formulated as
E[
(m)
]E[x] is proportional to
x
m
e
m
x
/
H
[ x]
M
m=1
x
m
e
m
x
/
H
[ x]
/
H
[x]/
H
[x],
12
where e
m
is the m-th unit vector. Introducing the quantity
x
m
e
m
x
/
H
[ x]
M
m=1
x
m
e
m
x
/
H
[ x]
/
H
[ x]
as the local risk capital contribution, the following condition holds at opti-
mality: For each asset in the portfolio, the return contribution minus total
return is proportional to the local risk capital contribution minus the total
risk capital.
5 Numerical portfolio optimization
A numerical procedure to solve (6) may be based on the dual representation:
Using /[Y ] = infE[Y Z] : Z Z this problem may be solved by the
following dual iterative procedure:
1. Set
Z = .
2. Outer problem. Solve
1
, . . .
S
. Let us form the probability vector
p =
P(
1
)
.
.
.
P(
S
)
s,m
=
(m)
(
s
).
The set Z is a set of [1 S] vectors. The outer problem reads
Maximize p
x
such that
z
s
i=1
p
i
and
z
s
=
p
s
p
s1
h(u) du s = 1, . . . S.
Let nally z
s
= z
(s)
. Then the minimal value of (12) is
S
s=1
y
[s:S]
z
s
=
S
s=1
y
s
z
s
14
and the minimizer takes the value z
s
for the scenario s. Form the vector
z = (z
1
, . . . , z
S
)
x q
where
u
= z
diag (p) .
Proposition 7. For a nite probability space, the dual iterative proce-
dure stops after nitely many steps at a solution.
Proof. The set Z is a polyhedral set of vectors in R
s
. This set has only
nitely many extremal points. The inner problem generates at each step a
new extremal point. Assume that the procedure does not stop at step n.
Then infE[Y
x
n
z] : z
Z < infE[Y
x
n
z] : z z
1
, . . . z
n
and therefore
z
n+1
, the minimizer of infE[Y
x
n
z] : z
Z, cannot be contained in the
convex hull of z
1
, . . . , z
n1
. Thus the procedure must stop in nitely many
steps. 2
References
[1] Acerbi, C.: Spectral measures of risk: a coherent representation of sub-
jective risk aversion. Journal of Banking and Finance 26, 15051518
(2002)
[2] Acerbi, C., Tasche, D.: Expected shortfall: a natural coherent alterna-
tive to value-at-risk. Economic notes 31(2), 379388 (2002)
[3] Artzner, P.: Application of coherent risk measures to capital require-
ments in insurance. North American Acturial Journal 3, 1125 (1999)
[4] Artzner, P., Delbaen, F., Eber, J.M., Heath, D.: Coherent measures of
risk. Mathematical Finance 9, 203228 (1999)
[5] Artzner, P., Delbaen, F., Eber, J.M., Heath, D.: Risk management
and capital allocation with coherent measures of risk (2000). Working
Paper,http://www.math.ethz.ch/nance
[6] Choquet, G.: Theory of capacities. Annales de lInstitut Fourier 5,
131295 (1954)
[7] De Giorgi, E.: Reward-risk portfolio selection and stochastic dominance.
Journal of Banking & Finance 29(4), 895926 (2005)
15
[8] Denneberg, D.: Distorted probabilities and insurance premiums. In:
Proceedings of the 14th SOR, Ulm. Athen aum, Frankfurt (1989)
[9] Dentcheva, D., Ruszczy nski, A.: Convexication of stochastic ordering.
C. R. Acad. Bulgare Sci. 57(4), 1116 (2004)
[10] Follmer, H., Schied, A.: Stochastic Finance: An Introduction in Discrete
Time. Walter DeGruyter, Berlin (2002)
[11] Hoeding, W.: Mastabinvariante korrelationstheorie. Schriften Math.
Inst. Univ. Berlin 5, 181233 (1940)
[12] H urlimann, W.: Distortion risk measures and economic capital. N. Am.
Actuar. J. 8(1), 8695 (2004)
[13] Jones, B.L., Zitikis, R.: Empirical estimation of risk measures and re-
lated quantities. N. Am. Actuar. J. 7(4), 4454 (2003)
[14] Kusuoka, S.: On law invariant risk measures. Advances In Mathematical
Economics 3, 8395 (2001)
[15] Lehmann, E.: Some concepts of dependence. Ann. Math. Stat. 37,
11371153 (1966)
[16] Pug, G.: Subdierential representations of risk measures. Mathemati-
cal Programming (to appear) (2005)
[17] Rockafellar, R., Uryasev, S.: Optimization of conditional value-at-risk.
Journal of Risk 2, 2141 (2000)
[18] Rockafellar, R., Uryasev, S.: Conditional value-at-risk for general loss
distributions. Journal of banking and Finance 26(7), 14431471 (2002)
[19] Wang, S.S.: A class of distortion operators for nancial and insurance
risks,. Journal of risk and insurance 67, 1536 (2000)
[20] Yaari, M.: The dual theory of choice under risk. Econometrica 55,
95115 (1987)
16