This document discusses an upcoming summer school presentation on modeling correlated risks using copulas. It will include a short introduction to copulas, quantifying dependence, statistical inference of copulas, and properties of aggregating risks. The presentation will define copulas and discuss their use in modeling multivariate distributions and quantifying dependence between random variables. It will also provide references on applying copulas in finance and insurance to model large correlated risks.
1. Arthur CHARPENTIER - École d'été EURIA.
mesures de risques et dépendance
Arthur Charpentier
Université de Rennes 1 & École Polytechnique
arthur.charpentier@univ-rennes1.fr
http://blogperso.univ-rennes1.fr/arthur.charpentier/index.php/
1
2. Arthur CHARPENTIER - École d'été EURIA.
3 0.9
V (rank of Y)
Y
0.4
-1
(X i,Y i)
(U i,V i)
-3 -1 1 3 0.2 0.5 0.8
X U (rank of X)
Density of the copula
Isodensity curves of the density
2
3. Arthur CHARPENTIER - École d'été EURIA.
Agenda
• General introduction
Modelling correlated risks
• A short introduction to copulas
• Quantifying dependence
• Statistical inference
• Agregation properties
3
4. Arthur CHARPENTIER - École d'été EURIA.
Agenda
• General introduction
Modelling correlated risks
• A short introduction to copulas
• Quantifying dependence
• Statistical inference
• Agregation properties
4
5. Arthur CHARPENTIER - École d'été EURIA.
Some references on large and correlated risks
Rank , J. (2006). Copulas: From Theory to Application in Finance. Risk Book ,
Nelsen , R. (1999,2006). An introduction to copulas. Springer Verlag ,
Cherubini , U., Luciano , E. & Vecchiato, W. (2004). Copula Methods in
Finance. Wiley,
Beirlant , J., Goegebeur , Y., Segers, J. & Teugels
, J. (2004). Statistics of
Extremes: Theory and Applications. Wiley,
McNeil , A. Frey , R., & Embrechts , P. (2005). Quantitative Risk
Management: Concepts, Techniques, and Tools. Princeton University Press,
5
6. Arthur CHARPENTIER - École d'été EURIA.
Agenda
• General introduction
Modelling correlated risks
• A short introduction to copulas
• Quantifying dependence
• Statistical inference
• Agregation properties
6
7. Arthur CHARPENTIER - École d'été EURIA.
Copulas, an introduction (in dimension 2)
Denition 1. A copula C is a joint distribution function on [0, 1]2 , with uniform
margins on [0, 1].
Set C(u, v) = P(U ≤ u, V ≤ v), where (U, V ) is a random pair with uniform
margins.
C is a distribution function on [0, 1]2 , and thus C(0, v) = C(u, 0) = 0, C(1, 1) = 1.
Furthermore C is increasing : since P is a positive measure, for all u1 ≤ u2 and
v1 ≤ v2 ,
Copula, positive area
1.0
P(u1 U ≤ u2 , v1 V ≤ v2 ) ≥ 0,
0.8
0.6
thus
0.4
C(u2 , v2 ) − C(u1 , v2 )
0.2
−C(u2 , v1 ) + C(u1 , v1 ) ≥ 0.
0.0
0.0 0.2 0.4 0.6 0.8 1.0
7
8. Arthur CHARPENTIER - École d'été EURIA.
C has uniform margins, and thus
C(u, 1) = P(U ≤ u, V ≤ 1) = P(U ≤ u) = u on [0, 1].
Proposition 2. C is a copula if and only if C(0, v) = C(u, 0) = 0, C(u, 1) = u
and C(1, v) = v for all u, v , with the following 2-increasingness property
C(u2 , v2 ) − C(u1 , v2 ) − C(u2 , v1 ) + C(u1 , v1 ) ≥ 0,
for any u1 ≤ u2 and v1 ≤ v2 .
8
9. Arthur CHARPENTIER - École d'été EURIA.
Borders of the copula function
!0.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
!0.2
!0.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
Figure 1: Value of the copula on the border of the unit square.
9
10. Arthur CHARPENTIER - École d'été EURIA.
Fonction de répartition à marges uniformes
Z
Y
X
Figure 2: Graphical representation of a copula.
10
11. Arthur CHARPENTIER - École d'été EURIA.
If C is twice dierentiable, one can dene its density as
∂ 2 C(u, v)
c(u, v) = .
∂u∂v
11
12. Arthur CHARPENTIER - École d'été EURIA.
Densité d’une loi à marges uniformes
z
x x
Figure 3: Density of a copula.
12
13. Arthur CHARPENTIER - École d'été EURIA.
Fonction de répartition à marges uniformes Densité d’une loi à marges uniformes
Fonction de répartition à marges uniformes Densité d’une loi à marges uniformes
Figure 4: Distribution functions and densities.
13
14. Arthur CHARPENTIER - École d'été EURIA.
Fonction de répartition à marges uniformes Densité d’une loi à marges uniformes
Fonction de répartition à marges uniformes Densité d’une loi à marges uniformes
Figure 5: Distribution functions and densities.
14
15. Arthur CHARPENTIER - École d'été EURIA.
Sklar's theorem
Theorem 3. (Sklar) Let C be a copula, and FX and FY two marginal
distributions, then F (x, y) = C(FX (x), FY (y)) is a bivariate distribution
function, with F ∈ F(FX , FY ).
Conversely, if F ∈ F(FX , FY ), there exists C such that
F (x, y) = C(FX (x), FY (y)). Further, if FX and FY are continuous, then C is
unique, and given by
C(u, v) = F (FX (u), FY (v)) for all (u, v) ∈ [0, 1] × [0, 1]
−1 −1
We will then dene the copula of F , or the copula of (X, Y ).
In that case, the copula of (X, Y ) is the distribution function of (FX (X), FY (Y )).
Proposition 4. If (X, Y ) has copula C , the copula of (g(X), h(Y )) is also C for
any increasing functions g and h.
15
16. Arthur CHARPENTIER - École d'été EURIA.
Copulas, an introduction (in dimension d ≥ 2)
Denition 5. A copula C is a joint distribution function on [0, 1]d , with
uniform margins on [0, 1].
Let U = (U1 , ..., Ud ) denote a random pair with uniform margins.
C is a distribution function on [0, 1]d , and thus C(u) = 0 if ui = 0 for some
i ∈ {1, . . . , d}, and C(1) = 1.
Furthermore C satises some increasing property since P is a positive measure
(for all 0 ≤ u ≤ v ≤ 1, P(u U ≤ v) ≥ 0), thus
sign(z)C(z) ≥ 0,
z
where the sum is taken over all vertices of [u × v], and where sign(z) is +1 if
zi = ui for an even number of i (and −1 otherwise, see Figure 6). And nally C
has uniform margins, and thus
C(1, . . . , 1, ui , 1, . . . , 1) = ui on [0, 1].
16
17. Arthur CHARPENTIER - École d'été EURIA.
Increasing functions in dimension 3
Figure 6: The notion of 3-increasing functions.
17
18. Arthur CHARPENTIER - École d'été EURIA.
Theorem 6. (Sklar) Let C be a copula, and F1 , . . . , Fd be d marginal
distributions, then F (x) = C(F1 (x1 ), . . . , Fd (xd )) is a distribution function, with
F ∈ F(F1 , . . . , Fd ).
Conversely, if F ∈ F(F1 , . . . , Fd ), there exists C such that
F (x) = C(F1 (x1 ), . . . , Fd (xd )). Further, if the Fi 's are continuous, then C is
unique, and given by
C(u) = F (F1 (u1 ), . . . , Fd (ud )) for all (ui ) ∈ [0, 1]
−1 −1
We will then dene the copula of F , or the copula of X .
In that case, the copula of (X = (X1 , . . . , Xd ) is the distribution function of
U = (F1 (X1 ), . . . , Fd (Yd )).
Again, if C is dierentiable, one can dene its density,
∂ d C(u1 , . . . , ud )
c(u1 , . . . , ud ) = .
∂u1 . . . ∂ud
18
19. Arthur CHARPENTIER - École d'été EURIA.
Copulas in high dimension, a dicult problem
It is usually dicult to represent dependence in dimension d 2, and it is
usually studied by pairs.
In dimension d = 2, one can dene the following Fréchet class F(FX , FY , FZ )
dened by its marginal distributions. But it can also be interested to study
F(FXY , FXZ , FY Z ) dened by it paired distributions.
One of the problem that arises is the compatibility of marginals: one has to
verify that
CXY (x, y) = CX|Z (x|z)CY |Z (y|z)dz,
for instance.
19
21. Arthur CHARPENTIER - École d'été EURIA.
Copulas and ranks
The copula of X = (X1 , . . . , Xd ) is the distribution function of
U = (F1 (X1 ), . . . , Fd (Yd )).
In practice, since marginal distributions are unknown, the idea is to substitute
empirical distribution function,
n
#{observations Xi,j 's lower than xi } 1
Fi (xi ) = = 1(Xi,j ≤ xi ).
#{observations } n j=1
Note that
n
#{observations Xi,j 's lower than Xi,j0 } 1 Ri,j0
Fi (Xi,j0 ) = = 1(Xi,j ≤ Xi,j0 ) = ,
#{observations } n j=1
n
where Ri,j0 denotes the rank of Xi,j0 within {Xi,1 , ..., Xi,n }.
On a statistical point of view, studying the copula means studying ranks.
21
22. Arthur CHARPENTIER - École d'été EURIA.
Scatterplot of (X,Y) Scatterplot of the ranks of (X,Y)
9.0
20
8.5
8.0
15
Ranks of the Yi’s
7.5
Y (raw data)
10
7.0
6.5
5
6.0
5.5
2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 5 10 15 20
X (raw data) Ranks of the Xi’s
Scatterplot of the ranks of (X,Y), divided by n Scatterplot o+ ,-,/0, t1e copula!t3pe tran+orm o+ ,6,70
1.0
1.0
0.8
0.8
Vi=Ranks of the Yi’s/n+1
Ranks of the Yi’s/n
0.6
0.6
0.4
0.4
0.2
0.2
0.0
0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Ranks of the Xi’s/n Ui=Ranks of the Xi’s/n+1
Figure 8: Copulas, ranks and parametric inference, from (Xi , Yi ) to (Ui , Vi ).
22
23. Arthur CHARPENTIER - École d'été EURIA.
Some very classical copulas
• The independent copula C(u, v) = uv = C ⊥ (u, v).
The copula is standardly denoted Π, P or C ⊥, and an independent version of
L
(X, Y ) will be denoted (X ⊥ , Y ⊥ ). It is a random vector such that X⊥ = X and
L
Y⊥ =Y, with copula C ⊥.
In higher dimension, C ⊥ (u1 , . . . , ud ) = u1 × . . . × ud is the independent copula.
• The comonotonic copula C(u, v) = min{u, v} = C + (u, v).
The copula is standardly denoted M, or C +, and an comonotone version of
L
(X, Y ) will be denoted (X + , Y + ). It is a random vector such that X+ = X and
L
Y+ =Y, with copula C +.
(X, Y ) has copula C+ if and only if there exists a strictly increasing function h
L−1 −1
such that Y = h(X), or equivalently (X, Y ) = (FX (U ), FY (U )) where U is
U([0, 1]).
23
24. Arthur CHARPENTIER - École d'été EURIA.
Note that for any u, v
P(U ≤ u, V ≤ v) = P({U ∈ [0, u]} ∩ {V ∈ [0, v]})
≤ min{P(U ∈ [0, u]), P(V ∈ [0, v])}
thus, C(u, v) ≤ min{u, v} = C + (u, v). Thus, C+ is an upper bound for the set of
copulas.
In higher dimension, C + (u1 , . . . , ud ) = min{u1 , . . . , ud } is the comonotonic
copula.
• The contercomotonic copula C(u, v) = max{u + v − 1, 0} = C − (u, v).
The copula is standardly denoted W, or C −, and an contercomontone version of
L
(X, Y ) will be denoted (X − , Y − ). It is a random vector such that X− = X and
L
Y− =Y, with copula C −.
(X, Y ) has copula C− if and only if there exists a strictly decreasing function h
L−1 −1
such that Y = h(X), or equivalently (X, Y ) = (FX (1 − U ), FY (U )) where U is
U([0, 1]).
24
25. Arthur CHARPENTIER - École d'été EURIA.
Note that for any u, v ,
P(U ≤ u, V ≤ v) = P({U ∈ [0, u]} ∩ {V ∈ [0, v]})
= P(U ∈ [0, u]) + P(V ∈ [0, v]) − P({U ∈ [0, u]} ∪ {V ∈ [0, v]})
thus, C(u, v) ≥ u + v − 1 since P({U ∈ [0, u]} ∪ {V ∈ [0, v]}) ≤ 1, and since
C(u, v) ≥ 0, C(u, v) ≥ max{u + v − 1, 0} = C − (u, v). Thus, C − is a lower bound
for the set of copulas.
In higher dimension, C − (u1 , . . . , ud ) = max{u1 + . . . + ud − (d − 1), 0} is not a
copula: if (X, Y ) and (X, Z) are countercomonotonic, (Y, Z) is necessarily
comonotonic - it is not possible to have all component highly negatively
correlated.
Anyway, it is still the best pointwise lower bound.
25
27. Arthur CHARPENTIER - École d'été EURIA.
Pitfalls on independence and comonotonicity
The following proposition is false,
Uncorrect Proposition 7. If X and Y are independent, if Y and Z are
independent, then X and Z are independent.
If
(X, Y, Z) = (1, 1, 1) with probability 1/4,
(1, 2, 1) with probability 1/4,
(3, 2, 3) with probability 1/4,
(3, 1, 3) with probability 1/4,
then X and Y are independent, and Y and Z are independent, but X = Z.
27
28. Arthur CHARPENTIER - École d'été EURIA.
X and Y independent Y and Z independent X and Z comonotonic
4
4
4
3
3
3
Component Y
Component Z
Component Z
2
2
2
1
1
1
0
0
0
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
Component X Component Y Component X
Figure 10: Mixing independence and comonotonicity.
28
29. Arthur CHARPENTIER - École d'été EURIA.
Pitfalls on independence and comonotonicity
The following proposition is false,
Uncorrect Proposition 8. If X and Y are comonotonic, if Y and Z are
comonotonic, then X and Z are comonotonic.
If
(X, Y, Z) = (1, 1, 1) with probability 1/4,
(1, 2, 3) with probability 1/4,
(3, 2, 1) with probability 1/4,
(3, 3, 3) with probability 1/4,
then X and Y are comonotonic, and Y and Z are comonotonic, but X and Z are
independent.
29
30. Arthur CHARPENTIER - École d'été EURIA.
X and Y comonotonic Y and Z comonotonic X and Z independent
4
4
4
3
3
3
Component Y
Component Z
Component Z
2
2
2
1
1
1
0
0
0
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
Component X Component Y Component X
Figure 11: Mixing independence and comonotonicity.
30
31. Arthur CHARPENTIER - École d'été EURIA.
Pitfalls on independence and comonotonicity
The following proposition is false,
Uncorrect Proposition 9. If X and Y are comonotonic, if Y and Z are
independent, then X and Z are independent.
If
(X, Y, Z) = (1, 1, 3) with probability 1/4,
(2, 1, 1) with probability 1/4,
(2, 3, 3) with probability 1/4,
(3, 3, 1) with probability 1/4,
then X and Y are comonotonic, and Y and Z are independent, but X and Z are
anticomonotonic.
31
32. Arthur CHARPENTIER - École d'été EURIA.
If
(X, Y, Z) = (1, 1, 1) with probability 1/4,
(2, 1, 3) with probability 1/4,
(2, 3, 1) with probability 1/4,
(3, 3, 3) with probability 1/4,
then X and Y are comonotonic, and Y and Z are independent, but X and Z are
comonotonic.
32
33. Arthur CHARPENTIER - École d'été EURIA.
X and Y comonotonic Y and Z independent X and Z comonotonic
4
4
4
3
3
3
Component Y
Component Z
Component Z
2
2
2
1
1
1
0
0
0
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
Component X Component Y Component X
Figure 12: Mixing independence and comonotonicity.
33
34. Arthur CHARPENTIER - École d'été EURIA.
Elliptical (Gaussian and t) copulas
The idea is to extend the multivariate probit model, Y = (Y1 , . . . , Yd ) with
marginal B(pi ) distributions, modeled as Yi = 1(Xi ≤ ui ), where X ∼ N (I, Σ).
• The Gaussian copula, with parameter α ∈ (−1, 1),
Φ−1 (u) Φ−1 (v)
1 −(x2 − 2αxy + y 2 )
C(u, v) = √ exp dxdy.
2π 1 − α2 −∞ −∞ 2(1 − α2 )
Analogously the t-copula is the distribution of (T (X), T (Y )) where T is the t-cdf,
and where (X, Y ) has a joint t-distribution.
• The Student t-copula with parameter α ∈ (−1, 1) and ν ≥ 2,
t−1 (u) t−1 (v) −((ν+2)/2)
1 ν ν x2 − 2αxy + y 2
C(u, v) = √ 1+ dxdy.
2π 1 − α2 −∞ −∞ 2(1 − α2 )
34
35. Arthur CHARPENTIER - École d'été EURIA.
Archimedean copulas
Denition of Archimedean copulas
• Archimedian copulas C(u, v) = φ−1 (φ(u) + φ(v)), where φ is decreasing
convex (0, 1), with φ(0) = ∞ and φ(1) = 0.
Example 10. If φ(t) = [− log t]α , then C is Gumbel's copula, and if
φ(t) = t−α − 1, C is Clayton's. Note that C ⊥ is obtained when φ(t) = − log t.
How Archimedean copulas were introduced ?
1. The frailty approach ( Oakes (1989)).
Assume that X and Y are conditionally independent, given the value of an
heterogeneous component Θ. Assume further that
P(X ≤ x|Θ = θ) = (GX (x))θ and P(Y ≤ y|Θ = θ) = (GY (y))θ
for some baseline distribution functions GX and GY .
Then
F (x, y) = P(X ≤ x, Y ≤ y) = E(P(X ≤ x, Y ≤ y|Θ = θ))
35
36. Arthur CHARPENTIER - École d'été EURIA.
thus, since X and Y are conditionally independent,
F (x, y) = E(P(X ≤ x|Θ = θ) × P(Y ≤ y|Θ = θ))
and therefore
F (x, y) = E (GX (x))Θ × (GY (y))Θ = ψ(− log GX (x) − log GY (y))
where ψ denotes the Laplace transform of Θ, i.e. ψ(t) = E(e−tΘ ). Since
FX (x) = ψ(− log GX (x)) and FY (y) = ψ(− log GY (y))
and thus, the joint distribution of (X, Y ) satises
F (x, y) = ψ(ψ −1 (FX (x)) + ψ −1 (FY (y))).
Example 11. If Θ is Gamma distributed, the associated copula is Clayton's. If
Θ has a stable distribution, the associated copula is Gumbel's.
36
37. Arthur CHARPENTIER - École d'été EURIA.
Consider two risks, X and Y, such that
X|Θ = θG ∼ E(θG ) and Y |Θ = θG ∼ E(θG ) are independent,
X|Θ = θB ∼ E(θB ) and Y |Θ = θB ∼ E(θB ) are independent,
(unobservable good (G) and bad (B ) risks).
The following gures start from 2 classes of risks, then 3, and then a continuous
risk factor θ ∈ (0, ∞).
37
38. Arthur CHARPENTIER - École d'été EURIA.
Conditional independence, two classes Conditional independence, two classes
20
3
2
15
1
10
0
!1
5
!2
!3
0
0 5 10 15 !3 !2 !1 0 1 2 3
Figure 13: Two classes of risks, (Xi , Yi ) and (Φ−1 (FX (Xi )), Φ−1 (FY (Yi ))).
38
39. Arthur CHARPENTIER - École d'été EURIA.
Conditional independence, three classes Conditional independence, three classes
3
40
2
30
1
0
20
!1
10
!2
!3
0
0 5 10 15 20 25 30 !3 !2 !1 0 1 2 3
Figure 14: Three classes of risks, (Xi , Yi ) and (Φ−1 (FX (Xi )), Φ−1 (FY (Yi ))).
39
41. Arthur CHARPENTIER - École d'été EURIA.
2. The survival approach: assume that there is a convex survival function S,
with S(0) = 1, such that
P(X x, Y y) = S(x + y),
then the joint survival copula of (X, Y ) is
S(S −1 (u) + S −1 (v)).
Example 12. If S is the Pareto survival distribution, the associated copula is
Clayton's. If S is the Weibull survival distribution, the associated copula is
Gumbel's.
41
42. Arthur CHARPENTIER - École d'été EURIA.
3. The use of Kendall's distribution function K(t) = P(C(U, V ) ≤ t) where
(U, V ) is a random pair with distribution function C .
Then, for Archimedean copulas,
φ (t)
K(t) = t − = t − λ(t),
φ(t)
which can be inverted easily in
1
1
φ(t) = φ(t0 ) exp dt ,
t0 λ(t)
for some 0 t0 1 and 0 ≤ u ≤ 1.
42
44. Arthur CHARPENTIER - École d'été EURIA.
Some characterizations of Archimedean copula
L
• Frank copula is the only Archimedean such that (U, V ) = (1 − U, 1 − V )
(stability by symmetry),
• Clayton copula is the only Archimedean such that (U, V ) has the same
copula as (U, V ) given (U ≤ u, V ≤ v) (stability by truncature),
• Gumbel copula is the only Archimedean such that (U, V ) has the same
copula as (max{U1 , ..., Un }, max{V1 , ..., Vn }) for all n ≥ 1 (max-stability),
44
45. Arthur CHARPENTIER - École d'été EURIA.
Extreme value copulas
• Extreme value copulas
log u
C(u, v) = exp (log u + log v) A ,
log u + log v
where A is a dependence function, convex on [0, 1] with A(0) = A(1) = 1, et
max{1 − ω, ω} ≤ A (ω) ≤ 1 for all ω ∈ [0, 1] .
An alternative denition is the following: C is an extreme value copula if for all
z 0,
1/z 1/z
C(u1 , . . . , ud ) = C(u1 , . . . , ud )z .
Those copula are then called max-stable: dene the maximum componentwise of
a sample X 1 , . . . , Xn , i.e. Mi = max{Xi,1 , . . . , Xi,n }.
45
46. Arthur CHARPENTIER - École d'été EURIA.
The joint distribution of M is
P(M ≤ x) = C(F1 (x1 , . . . , Fd (xd ))n ,
where C is the copula of the X i 's. Since P(Mi ≤ xi ) = Fi (xi )n , it can be written
P(M ≤ x) = C(P(M1 ≤ x1 )1/n , . . . , P(Md ≤ xd )1/n )n .
1/n 1/n
Thus, C(u1 , . . . , ud )n is the copula of the n maximum componentwise from a
sample with copula C .
Example 13. : If A is constant (1 on [0, 1]), then X and Y are independent,
and if A(ω) = max {ω, 1 − ω}, X and Y are comonotonic. Gumbel's copula is
obtained if
(
A(ω) = ((1 − ω)α + ω α + 1) 1/α),
for all 0 ≤ ω ≤ 1 and α ≥ 1.
46
47. Arthur CHARPENTIER - École d'été EURIA.
Pickands dependence function A
1.0
0.9
0.8
0.7
0.6
0.5
0.0 0.2 0.4 0.6 0.8 1.0
Figure 16: Shape of Gumbel's dependence function A(ω).
47
48. Arthur CHARPENTIER - École d'été EURIA.
How to construct much more copulas ?
Using geometric transformations
From a given copula C, cdf of random pair (U, V ), dene
• the copula of (U, 1 − V ),
C(U,1−V ) (u, v) = u − C(u, 1 − v)
• the copula of (1 − U, V ),
C(1−U,V ) (u, v) = v − C(1 − u, v)
• the copula of (1 − U, 1 − V ), the rotated or survival copula,
C(1−U,1−V ) (u, v) = C ∗ (u, v) = u + v − 1 + C(1 − u, 1 − v)
Note that if P(X ≤ x, Y ≤ y) = C(P(X ≤ x), P(Y ≤ y)), then
P(X x, Y y) = C ∗ (P(X x), P(Y y)).
48
49. Arthur CHARPENTIER - École d'été EURIA.
0.8
0.8
0.4
0.4
0.0
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Figure 17: Using geometric transformation to generate new copulas.
49
50. Arthur CHARPENTIER - École d'été EURIA.
0.8
0.8
0.4
0.4
0.0
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Figure 18: Using geometric transformation to generate new copulas.
50
51. Arthur CHARPENTIER - École d'été EURIA.
0.8
0.8
0.4
0.4
0.0
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Figure 19: Using geometric transformation to generate new copulas.
51
52. Arthur CHARPENTIER - École d'été EURIA.
0.8
0.8
0.4
0.4
0.0
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Figure 20: Using geometric transformation to generate new copulas.
52
53. Arthur CHARPENTIER - École d'été EURIA.
Using mixture of copulas
Lemma 14. The set of copulas is convex, i.e. if {Cθ , θ ∈ Ω} is a collection of
copulas,
C(u, v) = Cθ (u, v)dΠ(θ)
R
is a copula, where Π is a distribution on Ω
Thus C = αC1 + (1 − α)C2 denes a copula for all α ∈ [0, 1].
Example 15. Fréchet (1951) suggested a mixture of the lower and the upper
bound,
C(u, v) = αC − (u, v) + (1 − α)C + (u, v), for some α ∈ [0, 1].
Example 16. Mardia (1970) suggested a mixture of the lower, the upper
bound, and the independent copula
α2 − 2 ⊥ α2 +
C(u, v) = C (u, v) + (1 − α )C (u, v) + C (u, v), α ∈ [0, 1].
2 2
53
54. Arthur CHARPENTIER - École d'été EURIA.
Using distortion functions
Denition 17. A distortion function is a function h : [0, 1] → [0, 1] strictly
increasing such that h(0) = 0 and h(1) = 1.
The set of distortion function will be denoted H.
Note that h∈H if and only if h−1 ∈ H. Given a copula C, dene
Ch (u, v) = h−1 (C(h(u), h(v))).
If h is convex, then Ch is a copula, called distorted copula.
Example 18. if h(x) = x1/n , the distorted copula is
Ch (u, v) = C n (u1/n , v 1/n ), for all n ∈ N, (u, v) ∈ [0, 1]2 .
if the survival copula of the (Xi , Yi )'s is C , then the survival copula of
(Xn:n , Yn:n ) = (max{X1 , ..., Xn }, max{Y1 , ..., Yn }) is Ch .
54
55. Arthur CHARPENTIER - École d'été EURIA.
Example 19. if C(u, v) = uv = C ⊥ (u, v) (the independent copula), and
φ(·) = log h(·), then
Ch (u, v) = h−1 (h(u)h(v)) = φ−1 (φ(u) + φ(v)).
Example 20. if h(x) = [1 − e−αx ]/[1 − e−α ] (an exponential distortion), and if
C = C ⊥ , then
1 (e−αu − 1)(e−αv − 1)
Ch (u, v) = − log 1 + ,
α e−α − 1
which is Frank copula.
55
56. Arthur CHARPENTIER - École d'été EURIA.
Distorted Frank copula, h(x) = x Distorted Frank copula, h(x) = x(1 2)
Distorted Frank copula, h(x) = x(1 3)
Distorted Frank copula, h(x) = x(1 4)
Figure 21: Distorted copula, from Frank copula.
56
57. Arthur CHARPENTIER - École d'été EURIA.
Monte Carlo and copulas
Generation of independent variables can be done using a Random function.
Denition 21. Function Random should satisfy the following properties (i) for
all 0 ≤ a ≤ b ≤ 1,
P (Random ∈ ]a, b]) = b − a.
(ii) successive calls of function Random should generate independent draws, i.e.
0 ≤ a ≤ b ≤ 1, 0 ≤ c ≤ d ≤ 1
P (Random1 ∈ ]a, b] , Random2 ∈ ]c, d]) = (b − a) (d − c) ,
or more generally, dene k-uniformity for all 0 ≤ ai ≤ bi ≤ 1, i = 1, ..., k,
k
P (Random1 ∈ ]a1 , b1 ] , ..., Randomk ∈ ]ak , bk ]) = (bi − ai ) .
i=1
Thus, one can generate easily random vectors U = (U1 , ..., Ud ) with independent
component.
57
58. Arthur CHARPENTIER - École d'été EURIA.
The idea to generate correlated vectors U = (U1 , ..., Ud ), the idea is to use rst
P(U1 ≤ u1 , . . . , Ud ≤ ud ) = P(Ud ≤ ud |U1 ≤ u1 , . . . , Ud−1 ≤ ud−1 )
×P(Ud−1 ≤ ud−1 |U1 ≤ u1 , . . . , Ud−2 ≤ ud−2 )
×...
×P(U3 ≤ u3 |U1 ≤ u1 , U2 ≤ u2 )
×P(U2 ≤ u2 |U1 ≤ u1 ) × P(U1 ≤ u1 ).
58
59. Arthur CHARPENTIER - École d'été EURIA.
Starting from the end, P(U1 ≤ u1 ) = u1 since U1 is uniform, while
P(U2 ≤ u2 |U1 = u1 )
= P(U2 ≤ u2 , U3 ≤ 1, . . . Ud ≤ 1|U1 = u1 )
= lim P(U2 ≤ u2 , U3 ≤ 1, . . . Ud ≤ 1|U1 ∈ [u1 , u1 + h])
h→0
P(u1 ≤ U1 ≤ u1 + h, U2 ≤ u2 , U3 ≤ 1, . . . Ud ≤ 1)
= lim
h→0 P(U1 ∈ [u1 , u1 + h])
P(U1 ≤ u1 + h, U2 ≤ u2 , U3 ≤ 1, . . . Ud ≤ 1) − P(U1 ≤ u1 , U2 ≤ u2 , U3 ≤ 1, . . .
= lim
h→0 P(U1 ∈ [u1 , u1 + h])
C(u1 + h, u2 , 1, . . . , 1) − C(u1 , u2 , 1, . . . , 1) ∂C
= lim = C(u1 , u2 , 1, . . . , 1).
h→0 h ∂u1
and more generally,
∂ k−1
P(Uk ≤ uk |U1 = u1 , . . . , Uk−1 = uk−1 ) = C(u1 , . . . , uk , 1, . . . , 1).
∂u1 . . . ∂uk−1
59
60. Arthur CHARPENTIER - École d'été EURIA.
Thus, U = (U1 , .., Un ) with copula C could be simulated using the following
algorithm,
• simulate U1 uniformly on [0, 1],
u1 ← Random1 ,
• simulate U2 from the conditional distribution ∂1 C(·|u1 ),
u2 ← [∂1 C(·|u1 )]−1 (Random2 ),
• simulate Uk from the conditional distribution ∂1,...,k−1 C(·|u1 , ..., uk−1 ),
uk ← [∂1,...,k−1 C(·|u1 , ..., uk−1 )]−1 (Randomk ),
...etc, where the Randomi 's are independent calls of a Random function.
This is the underlying idea when using Cholesky decomposition.
60
61. Arthur CHARPENTIER - École d'été EURIA.
Example: for Clayton's copula, C(u, v) = (u−α + v −α − 1)−1/α , (U, V ) has joint
distribution C if and only if U is uniform on on [0, 1] and V |U = u has
conditional distribution
P(V ≤ v|U = u) = ∂2 C(v|u) = (1 + uα [v −α − 1])−1−1/α .
The algorithm to generate Clayton's copula is the
• simulate U1 uniformly on [0, 1],
u1 ← Random1 ,
• simulate U2 from the conditional distribution ∂2 C(·|u),
u2 ← [∂1 C(·|u1 )]−1 (Random2 ),
i.e.
u2 ← [(Random2 )−α/(1+α − 1]u−α + 1−1/α .
1
61
62. Arthur CHARPENTIER - École d'été EURIA.
1.5
0.0 0.5 1.0 1.5 2.0
1.0
0.5
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Distribution of v given u=0.3 Distribution of v given u=0.5
Generation of Clayton’s copula
0.8
0.0 0.5 1.0 1.5
0.4
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Distribution of v given u=0.8
Figure 22: Simulation of Clayton's copula.
62
68. Arthur CHARPENTIER - École d'été EURIA.
400
400
y
y
q
q
200
200
0
0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Distribution de U Distribution de V
4
0.8
2
0
0.4
!4 !2
0.0
0.0 0.2 0.4 0.6 0.8 1.0 !4 !2 0 2 4
Uniform margins Standard Gaussian margins
Figure 28: Simulation of Clayton's survival copula.
68
69. Arthur CHARPENTIER - École d'été EURIA.
400
400
y
y
q
q
200
200
0
0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Distribution de U Distribution de V
4
0.8
2
0
0.4
!4 !2
0.0
0.0 0.2 0.4 0.6 0.8 1.0 !4 !2 0 2
Uniform margins Standard Gaussian margins
Figure 29: Simulation of a copula mixture.
69
70. Arthur CHARPENTIER - École d'été EURIA.
Copulas in nance: options on multiple assets
Remark 22. Recall that Breeden Litzenberger (1978) proved that the risk
neutral probability can be obtrained from option prices: consider the price of a call
∞
C(T, K) = e−rT EQ ((ST − K)+ ). Since (ST − K)+ = K 1(ST x)dx, one gets
∞
−rT
C(T, K) = e Q(ST x)dx,
K
hence
∂C ∂P
Q(ST ≤ x) = −e−rT (T, x), or Q(ST ≤ x) = −erT (T, x)
∂K ∂K
where P denotes the price of a put option.
1 2
Consider an option on 2 assets, with payo h(ST , ST ). The price at time 0 is
e−rT EQ (h(ST , ST )).
1 2
70
71. Arthur CHARPENTIER - École d'été EURIA.
Copulas in nance: call on maximum
1 2 1 2
Here the payo is h(ST , ST ) = (max{ST , ST } − K)+ . The price is then
C(T, K) = e−rT EQ ((max{ST , ST } − K)+ )
1 2
∞
= e−rT EQ 1 2
1 − 1(max{ST , ST } ≤ x)dx
K
∞
−rT 1 2
= e 1 − Q(max{ST , ST } ≤ x) dx,
K
1 2
Q(ST ≤x,ST ≤x)
1 2
hence, if (ST , ST ) has copula C (under Q), then
∞
∂P 1 ∂P 2
C(T, K) = e−rT 1 − C erT (T, x), erT (T, x) dx.
K ∂K ∂K
71
72. Arthur CHARPENTIER - École d'été EURIA.
Copulas in nance: call on spreads
1 2 1 2
Here the payo is h(ST , ST ) = ([ST − ST ] − K)+ . The price is then
∞
−rT 1 2 −rT 2 1
C(T, K) = e EQ ((ST − ST − K)+ ) = e EQ 1(ST + K ≤ x ≤ ST )dx
−∞
∞
= e−rT 2 2 1
Q(K + ST ≤ x) − Q(ST + K ≤ x, ST ≤ x} ≤ x) dx,
−∞
1 2
Q(ST ≤x,ST ≤x+K)
1 2
hence, if (ST , ST ) has copula C (under Q), then
∞
−rT rT ∂P 2 rT ∂P
1
rT ∂P
2
C(T, K) = e e (T, x−K)−C e (T, x), e (T, x − K) dx.
−∞ ∂K ∂K ∂K
72
73. Arthur CHARPENTIER - École d'été EURIA.
Copulas in nance: bonds on option prices
Using Tchen's inequality, it is possible to derive bounds for options when the
payo is supermodular.
73
74. Arthur CHARPENTIER - École d'été EURIA.
Agenda
• General introduction
Modelling correlated risks
• A short introduction to copulas
• Quantifying dependence
• Statistical inference
• Agregation properties
74
75. Arthur CHARPENTIER - École d'été EURIA.
Natural properties for dependence measures
Denition 23. κ is measure of concordance if and only if κ satises
1. κ is dened for every pair (X, Y ) of continuous random variables,
2. −1 ≤ κ (X, Y ) ≤ +1, κ (X, X) = +1 and κ (X, −X) = −1,
3. κ (X, Y ) = κ (Y, X),
4. if X and Y are independent, then κ (X, Y ) = 0,
5. κ (−X, Y ) = κ (X, −Y ) = −κ (X, Y ),
6. if (X1 , Y1 ) P QD (X2 , Y2 ), then κ (X1 , Y1 ) ≤ κ (X2 , Y2 ),
7. if (X1 , Y1 ) , (X2 , Y2 ) , ... is a sequence of continuous random vectors that
converge to a pair (X, Y ) then κ (Xn , Yn ) → κ (X, Y ) as n → ∞.
75
76. Arthur CHARPENTIER - École d'été EURIA.
As pointed out in Scarsini (1984), most of the axioms are self-evident .
Ifκ is measure of concordance, then, if f and g are both strictly increasing, then
κ(f (X), g(Y )) = κ(X, Y ). Further, κ(X, Y ) = 1 if Y = f (X) with f almost
surely strictly increasing, and analogously κ(X, Y ) = −1 if Y = f (X) with f
almost surely strictly decreasing (see Scarsini (1984)).
76
77. Arthur CHARPENTIER - École d'été EURIA.
Association measures: Kendall's τ and Spearman's ρ
Rank correlations can be considered, i.e. Spearman's ρ dened as
1 1
ρ(X, Y ) = corr(FX (X), FY (Y )) = 12 C(u, v)dudv − 3
0 0
and Kendall's τ dened as
1 1
τ (X, Y ) = 4 C(u, v)dC(u, v) − 1.
0 0
Historical version of those coecients
Spearman's rho was introduced in Spearman (1904) as
ρ(X, Y ) = 3[P((X1 − X2 )(Y1 − Y3 ) 0) − P((X1 − X2 )(Y1 − Y3 ) 0)],
where (X1 , Y1 ), (X2 , Y2 ) and (X3 , Y3 ) denote three independent versions of
(X, Y ) (see Nelsen (1999)).
77
78. Arthur CHARPENTIER - École d'été EURIA.
Similarly Kendall's tau was not dened using copulae, but as the probability of
concordance, minus the probability of discordance, i.e.
τ (X, Y ) = 3[P((X1 − X2 )(Y1 − Y2 ) 0) − P((X1 − X2 )(Y1 − Y2 ) 0)],
where (X1 , Y1 ) and (X2 , Y2 ) denote two independent versions of (X, Y ) (see
Nelsen (1999)).
4Q
Equivalently, τ (X, Y ) = 1 − 2 − 1)
where Q is the number of inversions
n(n
between the rankings of X and Y (number of discordance).
78
79. Arthur CHARPENTIER - École d'été EURIA.
1.5 Concordant pairs Discordant pairs
1.5
1.0
1.0
0.5
0.5
Y
Y
0.0
0.0
!0.5
!0.5
!2.0 !1.5 !1.0 !0.5 0.0 0.5 1.0 !2.0 !1.5 !1.0 !0.5 0.0 0.5 1.0
X X
Figure 30: Concordance versus discordance.
79
80. Arthur CHARPENTIER - École d'été EURIA.
The case of the Gaussian random vector
If (X, Y ) is a Gaussian random vector with correlation r, Kruskal
then ( (1958))
6 r 2
ρ(X, Y ) = arcsin and τ (X, Y ) = arcsin (r) .
π 2 π
80
81. Arthur CHARPENTIER - École d'été EURIA.
Link between Kendall's tau and Spearman's rho
Note that Kendall's tau and Spearman's are linked: it is impossible to have at
the same time τ ≥ 0.4 and ρ = 0.
Hence ρ and τ satisfy
3τ − 1 1 + 2τ − τ 2
≤ρ≤ if τ ≥0
2 2
τ 2 + 2τ − 1 1 + 3τ
≤ρ≤ if τ ≤ 0.
2 2
which yield the area given below.
81
82. Arthur CHARPENTIER - École d'été EURIA.
1.0
0.5
Rho de Spearman
0.0
-0.5
-1.0
-1.0 -0.5 0.0 0.5 1.0
Tau de Kendall
Figure 31: Admissible region of ρ and τ.
82
85. Arthur CHARPENTIER - École d'été EURIA.
Alternative expressions of those coecients
Note that those coecients can also be expressed as follows
[0,1]×[0,1]
C(u, v) − C ⊥ (u, v)dudv
ρ(X, Y ) = (1)
[0,1]×[0,1]
C + (u, v) − C ⊥ (u, v)dudv
(the normalized average distance between C and C ⊥ ), for instance.
A dependence measure in higher dimension ?
From equations 1 and ??, it is possible to obtain a natural mutlidimensional
extention (see Wolf (1980), Joe (1990) or Nelsen (1996)),
[0,1]d
C(u) − C ⊥ (u)du d+1
ρ(X) = = 2d C(u)du − 1
[0,1]×[0,1]
C + (u) − C ⊥ (u)du 2d − (d + 1) [0,1]d
(2)
and similarly
1
τ (X) == d−1 2d C(u)dCu − 1 (3)
2 − 1) [0,1]d
85
86. Arthur CHARPENTIER - École d'été EURIA.
Note that a lower bound for τ is then −1/(2d−1 − 1), while it is
(2d − (d + 1)!)/(d!(2d − (d + 1))).
In dimension 3, Kendall's τ is the average of the three 2-dimensional Kendall's τ,
1
τ (X, Y, Z) = (τ (X, Y ) + τ (X, Z) + τ (Y, Z)).
3
86
87. Arthur CHARPENTIER - École d'été EURIA.
Tail concentration functions
Venter (2002) suggest to use several Tail Concentration Functions
Denition 24. For lower tails, dene
L(z) = P(U z, V z)/z = C(z, z)/z = P r(U z|V z) = P r(V z|U z),
and for upper tails,
R(z) = P(U z, V z)/(1 − z) = P r(U z|V z).
Joe (1990) uses the term upper tail dependence parameter for
R = R(1) = limz→1 R(z), and lower tail dependence parameter for
L = L(0) = limz→0 L(z).
87
88. Arthur CHARPENTIER - École d'été EURIA.
Functional correlation measures
1 1
Consider also Kendall's tau, dened as −1 + 4 0 0
C(u, v)dC(u, v).
Denition 25. The cumulative tau can be dened as
z z
J(z) = −1 + 4 C(u, v)dC(u, v)/C(z, z)2 .
0 0
88