Transcript presentation

Combined functionals
as risk measures
Arcady Novosyolov
Institute of computational modeling
SB RAS, Krasnoyarsk, Russia, 660036
[email protected]
http://www.geocities.com/novosyolov/
Structure of the presentation
Risk
Risk measure
RM: Expectation
RM: Expected utility
RM: Distorted probability
RM: Combined functional
Relations among risk measures
Illustrations
Anticipated questions
A. Novosyolov
Combined functionals
2
Risk
Risk is an almost surely bounded random variable
X  X  L (, B, P)
Another interpretation: risk is a real distribution
function with bounded support
F F
Correspondence:
X  FX , FX (t )  P( X  t )
Previous
Next
Back to structure
Why bounded?
A. Novosyolov
Combined functionals
3
Example: Finite sample space
Let the sample space be finite: |  | n. Then:
Probability distribution is a vector P  ( p1 , p2 ,..., pn )
n
Random variable is a vector X  ( x1 , x2 ,..., xn )  R
Distribution function is a step function
1
FX (t )
pn
p2
p1
x1
A. Novosyolov
x2

xn
Combined functionals
t
Previous
Next
Back to structure
4
Risk treatment
Here risk is treated as gain (the more, the better).
Examples:
• Return on a financial asset
• Insurable risk
x -$1,000,000
p
0.02
• Profit/loss distribution
(in thousand dollars)
A. Novosyolov
x
0%
20%
p
0.1
0.9
$0
0.98
x
-20
p
0.01
Combined functionals
-5
10
30
0.13 0.65 0.21
Previous
Next
Back to structure
5
Risk measure
Risk measure is a real-valued functional
:XR
or
 : F  R.
Risk measures allowing both representations with
 ( X )   ( FX )
are called law invariant.
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
6
Using risk measures




Certain equivalent of a risk X ( FX )
Price of a financial asset, portfolio X ( FX )
Insurance premium for a risk X ( FX )
Goal function in decision-making
problems
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
7
RM: Expectation
 ( X )  EX

 ( F )   tdF (t )

Expectation is a very simple law invariant risk measure, describing a
risk-neutral behavior. Being almost useless itself, it is important as a
basic functional for generalizations.
Expected utility risk measure may be treated as a combination of
expectation and dollar transform.
Distorted probability risk measure may be treated as a combination
of expectation and probability transform.
Combined functional is essentially the application of both
transforms to the expectation.
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
8
RM: Expected utility

U ( F )   U (t )dF (t )
U ( X )  EU ( X )

Expected utility is a law invariant risk measure,
exhibiting risk averse behavior, when its utility function U
is concave (U''(t)<0).
Expected utility is linear with respect to mixture of
distributions, a disadvantageous feature, that leads to
effects, perceived as paradoxes.
Is EU a certain equivalent?
A. Novosyolov
Combined functionals
Previous
Next
Back to structure
9
EU as a dollar transform
X:
U (X ) :
n
Value
x1
x2
…
xn
Prob
p1
p2
…
pn
Value U(x1) U(x2)
…
U(xn)
Prob
…
pn
EX   x k p k
k 1
A. Novosyolov
p1
p2
n
U ( X )   U ( x k ) p k
k 1
Combined functionals
Previous
Next
Back to structure
10
EU is linear in probability
U (aF  (1  a)G)  aU ( F )  (1  a) U (G),
p2
a  [0,1];
Expected utility functional
is linear with respect to
mixture of distributions.
p1
p3
F, G  F
Indifference "curves" on a set of
probability distributions: parallel
straight lines
A. Novosyolov
Combined functionals
Previous
Next
Back to structure
11
EU: Rabin's paradox
Rx ( L, G) :
Value
x-L
x+G
prob
0.5
0.5
Consider equiprobable gambles implying loss L or gain G with
probability 0.5 each, with initial wealth x. Here 0<L<G. Rabin
had discovered the paradox: if an expected utility maximizer
rejects such gamble for any initial wealth x, then she would
reject similar gambles with some loss L0> L and any gain G0,
no matter how large.
Example: let L = $100, G = $125. Then expected utility
maximizer would reject any equiprobable gamble with
Previous
Next
loss L0= $600.
Back to structure
A. Novosyolov
Combined functionals
12
RM: Distorted probability
0


0
 g ( X )   [ g (1  FX (t ))  1]dt   g (1  FX (t ))dt
Distortion function g : [0,1]  [0,1], g (0)  0, g (1)  1
Distorted probability is a law invariant risk measure,
exhibiting risk averse behavior, when its distortion
function g satisfies g(v)<v, all v in [0,1].
Distorted probability is positive homogeneous, that
may lead to improper insurance premium calculation.
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
13
DP as a probability transform
X:
X:
n
EX   x k p k
k 1
Value
x1
x2
…
xn
Prob
p1
p2
…
pn
Value
x1
x2
…
xn
Prob
q1
q2
…
qn
n
 g ( X )   xk qk  EQ X
k 1
 n
  n

qk  g   pi   g   pi , k  1,..., n
 i k   i k 1 
Previous
Next
Q  (q1 , q2 ,..., qn )
Back to structure
A. Novosyolov
Combined functionals
14
DP is positively homogeneous
 g (aX )  a g ( X ),
a  0,
X X
Distorted probability is a positively homogeneous
functional, which is an undesired property
in insurance premium calculation.
Consider a portfolio containing a number of "small" risks with loss
$1,000 and a few "large" risks with loss $1,000,000 and identical
probability of loss. Then DP functional assigns 1000 times larger
premium to large risks, which seems intuitively insufficient.
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
15
RM: Combined functional
Recall expected utility and distorted probability functionals:
1
1
U ( X )   U ( F (v))dv,  g ( X )   FX1 (v)dg (1  v)
0
1
X
0
Combined functional involves both dollar and
probability transforms:
1
U , g ( X )   U ( FX1 (v))dg (1  v).
Discrete case:
0
n
U , g ( X )   U ( xk )qk  EQU ( X )
k 1
A. Novosyolov
Previous
Next
Back to structure
Combined functionals
16
CF, risk aversion
Combined functional exhibits risk aversion in a flexible
manner: if its distortion function g satisfies risk aversion
condition, then its utility function U need not be concave.
The latter may be even convex, thus resolving Rabin's
paradox. Next slides display an illustration.
Note that if distortion function g of a combined
functional does not satisfy risk aversion condition, then
the combined functional fails to exhibit risk aversion.
Concave utility function alone cannot provide "enough"
risk aversion.
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
17
CF, example parameters
U(t)
g(v)
2
1.0
1
0.5
0
0.0
-2
0
2
0.0
0.5
1.0
v
t0
 0.5 exp( t ),
U (t )  
t  0.5 exp( t ), t  0
A. Novosyolov
Combined functionals
g (v)  v 2.33
Previous
Next
Back to structure
18
CF: Rabin's paradox resolved
Given the combined functional with parameters from the previous
slide (with t measured in hundred dollars), the equiprobable
gamble with L = $100, G = $125 is rejected at any initial wealth,
and the following equiprobable gambles are acceptable at any
wealth level:
L0
$600
$1000
$2000
A. Novosyolov
G0
$2500
$4100
$8100
Combined functionals
Previous
Next
Back to structure
19
Relations among risk measures
Generalization
Partial generalization
A. Novosyolov
Combined functionals
Legend
Previous
Next
Back to structure
20
Legend for relations
EX
U (X )
 g (X )
U , g ( X )
- expectation
- expected utility
- distorted probability
- combined functional
RDEU – rank-dependent expected utility, Quiggin, 1993
Coherent risk measure – Artzner et al, 1999
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
21
Illustrations
Expected utility indifference curves
Distorted probability indifference curves
Combined functional indifference curves
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
22
EU: indifference curves
Over risks in R2
Over distributions in R3
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
23
DP: indifference curves
Over risks in R2
Over distributions in R3
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
24
CF: indifference curves
Over risks in R2
Over distributions in R3
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
25
A few anticipated questions
Why are risks assumed bounded?
Is EU a certain equivalent?
Previous
Next
Back to structure
A. Novosyolov
Combined functionals
26
Why are risks assumed bounded?
Boundedness assumption is a matter of convenience.
Unbounded random variables and distributions with
unbounded support may be considered as well, with
some additional efforts to overcome technical
difficulties.
Previous
Next
Back to structure
Back to Risk
A. Novosyolov
Combined functionals
27
Is EU a certain equivalent?
Strictly speaking, the value of expected utility
functional itself is not a certain equivalent. However,
the certain equivalent can be easily obtained by
applying the inverse utility function:
 U ( X )  U ( U ( X )), X  X
1
Previous
Next
Back to structure
Back to EU
A. Novosyolov
Combined functionals
28