You are on page 1of 19

Portfolio Optimization

Unit 4: Mean-Risk Multi-objective Framework


Duan LI & Xiangyu Cui
India Institute of Technology Kharagpur
May 26 - 30, 2014

2
Properties of Risk Measures

A risk measure (X) assigns a numerical value to a random wealth


X.
Coherent risk measure in Artzner et al. (1999):
(X) is monotonous, if X1 X2 implies that (X1 ) (X2 ).
(X) is subadditive, if (X1 + X2 ) (X1 ) + (X2 ).
(X) is translation invariant, if for every real number t,
(X + t) = (X) t.
(X) is positively homogeneous, if for every 0,
(X) = (X).

3
(X) is consistent with order relation , if X1 X2 implies that
(X1 ) (X2 ).
Rothschild-Stiglitz (R-S) stochastic dominance: We say X1 R
S X2 holds if and only if every risk-averse investor prefers X1 to
X2 (i.e. E[u(X1 )] E[u(X2 )] for every concave utility function
u).
First stochastic dominance (FSD): We say X1 F SD X2 holds
if and only if every non-satiable investor prefers X1 to X2 (i.e.
E[u(X1 )] E[u(X2 )] for every non-decreasing utility function
u).
Second stochastic dominance (SSD): We say X1 SSD X2 holds
if and only if every non-satiable risk-averse investor prefers X1 to
X2 (i.e. E[u(X1 )] E[u(X2 )] for every non-decreasing concave
utility function u).

4
(X) is linearizable if (X) is consistent with some stochastic order
and the optimization problem associated is linearizable.
(X) is positive if for every random variable X, (X) > 0 and (X) =
0 if and only if X is a constant almost surely.
(X) is translation invariant according to Gaivoronsky and Pflugs
(G-P) definition if for every real number t, (X + t) = (X).
(X) is convex if for every [0, 1]
(X1 + (1 )X2 ) (X1 ) + (1 )(X2 ).
Monotonous and the subadditive risk measures are convex.

5
Mean-Risk Portfolio Selection Model
There are n risky assets with random returns r = [r1 , . . . , rn ] . A riskaverse and non-satiable investor chooses a portfolio solution according to
the following optimization problem:
min (x r)
x

s.t. E(x r) m, x e = 1,
0 xi 1, i = 1, . . . , n,
for some real m [mini E(ri ), maxi E(ri )], where xi denotes the proportion of wealth invested in ith risky asset, e = [1, . . . , 1] .
Assume that we have observed T samples of r (and ri ), which are represented by r(t) (and ri,t ) for t = 1, . . . , T .

6
Safety Risk Measures

Safety Risk Measures are consistent with the first stochastic dominance
(FSD) order, and coherent risk measures are a subset of safety risk measures.
Safety-first (SF) (see Roy, 1952) The probability of portfolio return
less than a given target , i.e.,
P = P (x r ).

7
The mean-P portfolio selection model (MIPL):
min
x

T
X

vt

t=1

s.t. x E(r) m, x e = 1,
0 xi 1, i = 1, . . . , n,
vt = 0 or vt = 1, for t = 1, . . . , T,
( x r(t) ) < Lvt , t = 1, . . . , T,
where L is an opportunely large real number, r(t) = [r1,t , . . . , rn,t ] is
the scenario of returns r.

8
Value-at-risk (VaR) For given level ,
V aR = inf{z | P (x r z) > }.

9
V aR does NOT satisfy convex property.
Example: Consider two independent random wealth levels

W,
0.04
X1 =
W (1 + r), 0.96

W,
0.04
X2 =
W (1 + r), 0.96
By choosing = 0.05, we have
V aR (0.5X1 + 0.5X2 ) > 0.5V aR (X1 ) + 0.5V aR (X1 ).

10
The mean-V aR portfolio selection model:
min M[T :T ] (x r(1) , . . . , x r(T ) )
x

s.t. x E(r) m, x e = 1,
0 xi 1, i = 1, . . . , n,
where
r(t) = [r1,t , . . . , rn,t ] is the scenario of returns r,
M[k:T ] (u(1) , . . . , u(T ) ) is the kth largest among u(1) , . . . , u(T ) ,
M[T :T ] (x r(1) , . . . , x r(T ) ) is the empirical -quantile of the
scenarios x r(1) , . . . , x r(T ) .

11
MiniMax (MM) (see Young, 1998) The MiniMax risk measure of a
random variable X bounded from below is given by
M M (X) = sup{c R | P (X c) = 0}.
The mean-M M portfolio selection model (LP):
min
x

min

1tT

x r(t)

s.t. x E(r) m, x e = 1,
0 xi 1, i = 1, . . . , n,
where r(t) = [r1,t , . . . , rn,t ] is the scenario of returns r.

12
Conditional value-at-risk (CVaR) (see Bawa, 1978; Uryasev, 2000) For
given level
CV aR (X) = E[X| X V aR (X)].

13
CV aR is coherent risk measure.
The mean-CV aR portfolio selection model (LP):
T
1 X
vt
min b +
x,b
T t=1

s.t. x E(r) m, x e = 1,
0 xi 1, i = 1, . . . , n,
vt 0, t = 1, . . . , T,
vt x r(t) b, t = 1, . . . , T,
where r(t) = [r1,t , . . . , rn,t ] is the scenario of returns r.

14
Properties of Safety Risk Measures

15
Dispersion Measures

Dispersion Measures are increasing function of positive, and positively homogeneous, risk measures. Dispersion Measures are never consistent with
the FSD order.
Variance (see Markowitz, 1952) The risk measure of a random variable
X is V ar(X).
The mean-variance portfolio selection model (QP):
min V ar(x r)
x

s.t. 0 xi 1, i = 1, . . . , n,
x E(r) m, x e = 1.

16
Absolute Deviation (MAD) The risk measure of a random variable X
is
M AD(X) = E(|X E(X)|).
The mean-MAD portfolio selection model (LP):
T
1 X
min
yt
x
T t=1

s.t. yt +
yt

n
X
i=1
n
X

xi [ri,t E(ri )] 0 for t = 1, . . . , T,


xi [ri,t E(ri )] 0 for t = 1, . . . , T,

i=1

x E(r) m, x e = 1,
0 xi 1, i = 1, . . . , n.

17
Asymptotic Dispersion Approach (MQ) (see Ortobelli et al., 2002)
Assume that the returns r1 , . . . , rn are jointly -stable sub-Gaussian
distributed with > 1 (a particular elliptical distribution), whose
characteristic function is given by
r (u) = E[exp(iu r)] = exp[(u Qu)/2 + iu E(r)],
where Q = [qij ]is the dispersion matrix.
The dispersion measure of any portfolio x is given by
p
x r = x Qx.

18
Gini (see Shalit and Yitzhaki, 1984) Gini risk measure of a random
variable X is
Z Z
1
1
X =
|x y|fX (x)fX (y)dxdy = E(|X Y |),
2
2
R2
where Y is a copy i.i.d. of X.
The mean-Gini model (LP):
min
x

T X
T
X

yt,k

k=1 t>k

s.t. yt,k +
yt,k

n
X
i=1
n
X

xi (ri,t ri,k ) 0 for t > k = 1, . . . , T,


xi (ri,t ri,k ) 0 for t > k = 1, . . . , T,

i=1

x E(r) m, x e = 1,
0 xi 1, i = 1, . . . , n.

19
Properties of Dispersion Measures

You might also like