Comparing Dependent Multivariate Risks: Enkelejd Hashorva
Uploaded by
david AbotsitseComparing Dependent Multivariate Risks: Enkelejd Hashorva
Uploaded by
david AbotsitseComparing Risks
Comparing Dependent Multivariate Risks
Enkelejd Hashorva
University of Lausanne
1
Comparing Risks
Examples
Comparing Dependence
Applications
Appendix
2
Comparing Risks
Examples
Example: Gaussian/Normal RV
Let (X1 , X2 ) be jointly normal/Gaussian with marginal df's N(0, 1)
and correlation coe cient
ρ1 = E{X1 X2 } ∈ (−1, 1)
Let (Y1 , Y2 ) be another pair of jointly normal/Gaussian rv's with
marginal df's N(0, 1) and correlation coe cient
ρ2 = E{Y1 Y2 } ∈ (ρ1 , 1)
Question. Which random pair is more dependent?
In terms of correlations, since ρ(X1 , X2 ) = ρ1 < ρ2 = ρ(Y1 , Y2 ),
then the second pair is more correlated.
3
Comparing Risks
Examples
Example: Two Clayton Copulas
Let ψα (x) = (x −α − 1)/α, α > 0 be the Archimedean generator of
the Clayton copula Cα .
Consider two Clayton copulas Cα1 and Cα2 with α1 < α2 .
Question. Which copula Cαi exhibits more dependence?
If τ1 and τ2 are the respective Kendall's τ values, we have
α1 α2
τ1 = , τ2 =
α1 + 2 α2 + 2
In general, we can compare the dierent τ 's
in order to compare the dependence implied by two copulas.
4
Comparing Risks
Examples
Example: Mixture Copula
If C0 and C1 are 2-dimensional copulas, then the mixture copula
Cλ , λ ∈ (0, 1) is de ned by
Cλ (u1 , u2 ) = λC0 (u1 , u2 ) + (1 − λ)C1 (u1 , u2 ), u1 , u2 ∈ [0, 1]
How dependent is Cλ compared with C0 and C1 ?
5
Comparing Risks
Examples
Example: Geometric Copula
If C is a 2-dimensional copula, for any p = 1 − q ∈ (0, 1) de ne
pC (v1 , v2 ) ui
Cp (u1 , u2 ) = , vi = , ui ∈ [0, 1], i = 1, 2
1 − qC (v1 , v2 ) p + qui
Cp is i referred to as the geometric copula of C .
How dependent is Cp compared with C ?
6
Comparing Risks
Comparing Dependence
A formal comparison of dependence between two RV's (X1 , X2 ) and
(Y1 , Y2 ) can be done if we have a partial order:
I rst, we require that the pairs have the same marginal df's,
i.e., X1 =
d
Y1 and X2 = Y2
d
I we want to de ne a dependence order dep which is
a) re exive, b) transitive, c) antisymmetric
dep compares the strength of the underlying dependence
I we write
(X1 , X2 ) dep (Y1 , Y2 )
meaning that the dependence between Y1 and Y2 is stronger
than for the pair (X1 , X2 )
7
Comparing Risks
Comparing Dependence
Notes on the partial order
8
Comparing Risks
Comparing Dependence
Correlation Order
I Let (X1 , X2 ) and (Y1 , Y2 ) be two RV's with the same marginal
df's F1 , F2
I corr is a dependence order de ned by
(X1 , X2 ) corr (Y1 , Y2 )
if
P{X1 ≤ s, X2 ≤ t} ≤ P{Y1 ≤ s, Y2 ≤ t}, ∀(s, t) ∈ R2
9
Comparing Risks
Comparing Dependence
Example: Bivariate Gaussian/Normal RV
Let (X1 , X2 ) and (Y1 , Y2 ) be two RV's with Gausian/Normal df and
marginal df's N(0, 1).
Suppose that −1 ≤ ρ1 ≤ ρ2 ≤ 1 are the corresponding correlation
coe cients.
Then we have
(X1 , X2 ) corr (Y1 , Y2 )
10
Comparing Risks
Comparing Dependence
Archimedean Copulas: Concave ψ1(ψ2−1)
Let (X1 , X2 ) and (Y1 , Y2 ) be two RV's with df C1 and C2 which are
two bivariate Archimedean copulas with generators ψ1 and ψ2 .
If the function ψ1 (ψ2−1 ) is concave, then
(X1 , X2 ) corr (Y1 , Y2 )
Example: ψi (x) = (− ln x)αi , αi ≥ 1, i = 1, 2.
We have ψ2−1 (x) = exp(−x 1/α2 ), hence
ψ1 (ψ2−1 (x)) = (− ln(exp(−x 1/α2 )))α1 = x α1 /α2 , x ∈ (0, ∞)
which is concave for α1 ≤ α2 .
11
Comparing Risks
Comparing Dependence
Dependence Measures & Correlation Order
Let (X1 , X2 ) and (Y1 , Y2 ) be two RV's. It is not always possible to
order these RV's using the correlation order.
If (X1 , X2 ) corr (Y1 , Y2 ), then
τ (X1 , X2 ) ≤ τ (Y1 , Y2 ) (1)
ρS (X1 , X2 ) ≤ ρS (Y1 , Y2 ) (2)
and when correlations are de ned, then we have
ρ(X1 , X2 ) ≤ ρ(Y1 , Y2 ) (3)
12
Comparing Risks
Applications
Comparison of Stop-loss Premium - Aggregated Risks
Let (X1 , X2 ) ∼ F and (Y1 , Y2 ) ∼ G be two RV's with positive
components modelling some insurance risks. Suppose that marginal
df's of F and G are the same.
The sum of risks is X1 + X2 in model A, and Y1 + Y2 in model B.
Question. How can we compare stop-loss premiums for A and B?
If (X1 , X2 ) corr (Y1 , Y2 ), then
E{(X1 + X2 − d)+ } ≤ E{(Y1 + Y2 − d)+ }, ∀d ∈ (0, ∞) (4)
13
Comparing Risks
Applications
Comparison of Stop-loss Premium - Compound Risks
Let (N1 , N2 ) and (N1∗ , N2∗ ) be two RV's with integer-valued
components. De ne
Nj N∗
j
j = 1, 2
X X
Sj = Xij , Sj∗ = Xij∗ ,
i=1 i=1
the corresponding compound sums with Xij , Xij∗ , i ≥ 1, j = 1, 2
positive iid random variables modelling the claim sizes.
If we have (N1 , N2 ) corr (N1∗ , N2∗ ), then we can order the
compound sums as
(S1 , S2 ) corr (S1∗ , S2∗ ) (5)
14
Comparing Risks
Appendix
Archimedean Copula: Non-decreasing ψ1/ψ2
Let (X1 , X2 ) and (Y1 , Y2 ) be two RV's with df C1 and C2 which are
two bivariate Archimedean copulas with generators ψ1 and ψ2 .
If the function ψ1 /ψ2 is non-decreasing on (0, 1), then
(X1 , X2 ) corr (Y1 , Y2 )
Example: If ψi (x) = (1 − x)αi , αi > 0 for i = 1, 2, then we have
ψ1 (x)
= (1 − x)α1 −α2 , x ∈ (0, 1)
ψ2 (x)
which is non-decreasing for α1 ≤ α2 .
15
Comparing Risks
Appendix
Archimedean Copula: Non-decreasing ψ10 /ψ20
Let (X1 , X2 ) and (Y1 , Y2 ) be two RV's with df C1 and C2 as in the
previous slide. Assume that both ψ10 , ψ20 exist.
If the function ψ10 /ψ20 is non-decreasing on (0, 1), then
(X1 , X2 ) corr (Y1 , Y2 )
Example: If ψi (x) = exp(αi /x) − exp(αi ), αi > 0 for i = 1, 2, then
we have
ψ10 (x) α1 α − α
1 2
0 = exp , x ∈ (0, 1)
ψ2 (x) α2 x
which is non-decreasing for α1 ≤ α2 .
16
Comparing Risks
Appendix
Supermodular Order
I φ : R2 → R is supermodular if its corresponding set function
de nes a nite positive measure in the sense that
φ[a , b ] = φ(b1 , b2 ) − φ(b1 , a2 ) − φ(a1 , b2 ) + φ(a1 , a2 ) ≥ 0
for a ≤ b in R2
I When φ is twice dierentiable such that
∂2
φ(1,1) (x1 , x2 ) := φ(x1 , x2 ) ≥ 0
∂x1 ∂x2
then φ is called regular.
17
Comparing Risks
Appendix
Supermodular Order
I Given X and Y two bivariate RV's, then X SN Y if
E{φ(X )} ≤ E{φ(Y )}
for any regular supermodular φ such that the expectations
above exist
I SN is a dependence order
I X SN Y implies
ρ(X1 , X2 ) ≤ ρ(Y1 , Y2 )
ρS (X1 , X2 ) ≤ ρS (Y1 , Y2 )
τ (X1 , X2 ) ≤ τ (Y1 , Y2 )
18
Comparing Risks
Appendix
Supermodular Order & Correlation Order
I Let φ be a regular supermodular function with φ(1,1) ≥ 0
I When X and Y have identical marginal df's and non-negative
components. Integration by parts implies
E{φ(Y )} − E{φ(X )}
Z ∞Z ∞
= [FY (x ) − FX (x )]φ(1,1) (x ) dx1 dx2
0 0
I Consequently, SN and corr are equivalent
19
Comparing Risks
Appendix
Supermodular Order: Higher Dimensions
I φ : Rd → R is supermodular if it is pairwise (i, j) supermodular
I If φ it is pairwise dierentiable with φ(i,j1,1) ≥ 0, then φ is called
regular
I Similar de nition of X SN Y as in the bivariate setup
I Supermodular order does not reduce to correlation order in
higher dimensions
20
