An Introduction to Copulas : Which Copula is the right one?

Dependence Concepts Copula Families Elliptical Copula Archimedean Copula Kendall’s tau Spearman Rho Dependency Structure

이 석 형 (선임 검사역)
3786-7946(prime@fss.or.kr)

Expect the unexpected
1
For Credit Risk

Copula 함수의 이해
Copula Copula의 의 유용성 유용성 □ Risk Factor 간에 의존성(Dependency)을 측정 □ 선형 상관관계는 선형관계의 의존성을 나타내지만 실제 금융시장간의 의존성은 비선형관계(Non-linear dependency)를 인식 Elliptical distributions - linear dependence structure - correlation coefficient meaningful Non-elliptical distributions - alternative measures of dependence needed ⇒ Copulas □ 다변량 분포들은 Copula를 이용하면 쉽게 계산가능

역사 역사

□ 1940's : Hoeffding studies properties of multivariate distributions □ 1959 : The word copula appears for the first time (Sklar) □ 1999 : Introduced to financial applications (Embrechts, McNeil, Straumann) □ 2006 : Several insurance companies, banks and other financial institutions apply copula as a Risk management, Credit Derivatives, CDO, Options etc

2
For Credit Risk

Copula 함수의 이해
Copula 함수의 이해

Copula 함수는 복잡한 종속성 구조(dependence structure)를 고려하면서 다변량 누적분포함수를 추정하는데 유용한 도구 (is a function that links univariate marginals to their full multivariate distribution)

Copula 함수는 다변량 분포함수와 단변량 한계분포함수(margins)를 연결시키는 함수를 의미하며 데이터 사이의 종속성 구조와 개별자료의 분포를 분리하여 모형화함으로써 추정과 시뮬레이션을 용이 □ CDF

F ( x1 , x2 ) = P[ X 1 ≤ x1 , X 2 ≤ x2 ]
□ Copula fun (Sklar’s theorem)

F ( x2 ) = P[ X 2 ≤ x2 ]

F ( x1 ) = P[ X 1 ≤ x1 ]

C (u1 , u2 ) = F ( x1 , x2 )
□ Property of copula

C 2 ( F1 ( x1 ), F 2 ( x 2 )) = F ( x1 , x 2 )
xi = Fi (ui ) −1

C (u1 , u 2 ) C (u1 ,1) = u1

0 ≤ u1 , u2 ≤ 1

0 ≤ C (u1 , u 2 ) ≤ 1
P[ X 1 ≤ x1 ] = u1 P[ X 2 ≤ x2 ] = 1

Fi ( xi ) = ui

C (1, u2 ) = u2

C (u1 ,1) = F ( x1 , x2 )

x2 = ∞

C (u1 ,1) = F ( x1 , x2 ) = P[ X 1 ≤ x1 , X 2 ≤ ∞] = P[ X 1 ≤ x1 ] = u1

C(u1,0) = [0, u1 ]×[0,0] = 0
3
For Credit Risk

Copula 함수의 이해

• Standard Normal function : N(0,1)

• Copula Function

Fi ( xi ) = ui

45.0% 40.0% 35.0% 30.0% 25.0% 20.0% 15.0% 10.0% 5.0% 0.0%

A1

-4.0

-3.0

-2.0

-1.0

0.0

1.0

2.0

3.0

4.0

C(0.1,0.2) = F(−1.28, −0.84) = F(x1, x2 ) = P[ X1 ≤ −1.28, X 2 ≤ −0.84] = 6.90% = [0, 0.1] × [0, 0.2] /1000 □ 각 길이의 product
Where
For Credit Risk

F1 ( x1 ) = P[ X 1 ≤ x1 ] = 0.1

X 1 = − 1.28
4

Copula 함수의 이해
2 2 dimension dimension copula copula

5
For Credit Risk

Copula 함수의 이해

6
For Credit Risk

Copula 함수의 이해

7
For Credit Risk

Copula 함수의 이해

8
For Credit Risk

Copula 함수의 이해
□ Copula function vs Cumulative Distribution :Normal case
Normal Dist

Copula Fn

9
For Credit Risk

Copula 함수의 이해
Multivariate Multivariate Copula Copula Function Function

C (F1( x1 ), F2 ( x2 ), L , Fm ( xm ), ρ ) = Pr[U1 ≤ F1( x1 ), U2 ≤ F2 ( x2 ), L , Um ≤ Fm ( xm )]
−1 = Pr[F1−1(U1 ) ≤ x1 , F2−1(U2 ) ≤ x2 , L , Fm (Um ) ≤ xm ]

= Pr[X1 ≤ x1 , X 2 ≤ x2 , L , X m ≤ xm ] = F ( x1 , x2 , L , xm )
The The marginal marginal ditribution ditribution of of X Xii

C(F1(+∞), F2(+∞),L, F2(xi ),L Fm(+∞), ρ) = Pr[F1(+∞), F2(+∞),L, Fi (xi )L Fm(+∞), ρ] = Pr[X1 ≤ +∞, X 2 ≤ +∞, L , X i ≤ xi , L X m ≤ +∞] = Pr[X i ≤ xi ] = F ( xi )

F ( x1 , x2 , L , xm ) = C (F1( x1 ), F2 ( x2 ), L , Fm ( xm ), ρ )

C (u1 , u2 , L , um ) = Φ m (Φ −1(u1 ), Φ −1(u2 ), L , Φ −1(um ); ρ )
10
For Credit Risk

Copula 함수의 이해
포트폴리오의 포트폴리오의 함수 함수 :: Correlation Correlation 반영 반영 Copula 함수 : 다변량 균등분포(Uniform Distribution)확률 변수의 결합 누적 확률 분포

C (u1 , u2 ,........un ) = Pr[ X 1 ≤ x1 , X 2 ≤ x2 ....., X n ≤ xn ]
ui :
균등분포인 확률변수

C (u1 , u2 ,.........un )
u1 , u2 ,........un ∈ [0,1]
n

F1 ( x1 ), F2 ( x2 )...............Fn ( xn )
Fi ( xi ) = ui xi = Fi (ui ) −1

F ( x1 , x2 ,.........xn ) = C[ F1 ( x1 ), F2 ( x2 )...............Fn ( xn )]
다변량 누적 확률 분포 Sklar Representation 정리

C ( x1 , x2 ,......... xn ) = F [ F1−1 ( x1 ), F2 −1 ( x2 )...............Fn −1 ( xn )]
11
For Credit Risk

Copula 함수의 이해 Definition (Definition of copula function)
N-차원 Copula

C : [ 0 ,1] n → [ 0 ,1] 는 다음과 같은 특성을 만족하는 함수
∈ [ 0 ,1 ]에 대하여 C i ( u ) = C (1,..., 1, u ,1,..., 1) = u

(1) C 는 양의 확률(grounded)이고 n-증가 함수 (2) C 는 u

(3) C 는 n- 증가함수(increasing function)

f ( x1 ,...., x n ) 는

다음과 같은 2 부분으로 분류 (1)과 (2)의 결정하고 해당 모형에 모수를 결정

(1) Correlation structure : find the co-movement between variables may be viewed as a part of
dependence structure

(2) the product of marginal probability density functions
be achieved using regular statistical methods

: fit the individual marginal distribution and estimate their parameters (which can

* Sklar의 정리는 copula 함수를 이용하여 종속성을 모형화하는 기본 아이디어를 제공 어떤 다변량 분포함수에 대하여 단변량 한계분포와 종속성구조를 분리 12
For Credit Risk

Copula 함수의 이해 Theorem (Sklar’s Theorem)

F ( x1 ,..., x n ) = C ( F1 ( x1 ),..., Fn ( x n ))
∂F (x1,..., xn ) ∂C (F1(x1 ),..., Fn (xn )) f (x1,..., xn ) = = ∂x1...∂xn ∂x1...∂xn ∂C (u1,..., un ) ∂Fi (xi ) = ×∏ ∂u1...∂un ∂xi i
i

%) = c(u

f (x1,...., xn) ∏fi(xi )
i

% ) × ∏ fi (xi ) = c(u1,..., un ) × ∏ fi (xi ) = c(u
i

f ( x1 ,...., xn ) 는 F (⋅) 의
% = (u1,..., u2 ) u
%) c(u
For Credit Risk

함수(probability density function)

ui = Fi ( xi ) i = 1 ,..., n
** 모수 추정의 기초

는 Copula의 density function

13

Copula 함수의 이해

Copula function property (bi-variate)
– C(u, v) is between [0,1] and u, v are uniformly distributed [0,1] – C(0, v) = 0 – C(u, 1) = u; – C(u, v) is a 2-increasing function

Special copula functions
– Minimum copula : C(u, v) = max(u+v-1, 0) – Maximum copula : C(u, v) = min(u, v) – Gaussian copula : C(u, v) = Φρ(Φ-1(v), Φ-1(u)) – Student t-copula : C(u, v) = tρ,s(ts-1(u), ts-1(v)) – And more …

14
For Credit Risk

Copula 함수의 이해 Example 1: Independence copula □ If U ∼ U(0, 1) and V ∼ U(0, 1) are independent, then □ C(u, v) = uv = = P(U ≤ u)P(V ≤ v) = P(U ≤ u, V ≤ v) = H(u, v), □ where H(u, v) is the distribution function of (U, V). C is called the independence copula.

Example 2: Gaussian copula
φ −1 (u ) φ −1 (v )

C (u, v ) =

GA R

−∞

−∞

x 2 − 2R12 xy + y 2 1 exp{− }dxdy 2 1/2 2 2π (1 − R12 ) 2(1 − R12 )

Example 3: Student’s copula

C (u, v) =

t v ,R

−1 −1 tv (u) tv (v )

−∞

∫ ∫

−∞

x2 − 2R12 xy + y2 −(v +2) / 2 1 {1 + } dxdy 2 1/2 2 2π (1 − R12 ) v(1 − R12 )
15

□ ν is the degrees of freedom and R is the linear correlation coefficient.
For Credit Risk

Copula 함수의 이해 Gaussian copula Student’s copula
Ga m −1 −1 −1 Cρ ( X ) = Φv (Φv ( x1 ), Φv ( x2 ), L , Φv ( xm )) T m −1 −1 −1 Cρ ( X ) = tv (tv ( x1 ), tv ( x2 ), L , tv (xm ))

rho= 0.7

m=2인 경우

rho= 0.7, v = 4.

16
For Credit Risk

Copula 함수의 이해 Archimedian Archimedian Copula Copula □ Archimedean copula는 분석이 쉬울 뿐만 아니라(대부분의 Archimedean copula는 closed form expression을 가진다) 여러 가지 다른 종속성 구조를 제공

ϕ : [0,1] → [0, ∞]

ϕ −1 :[0, ∞] → [0,1]

ϕ

가 볼록성을 가지면 copula 함수는 다음과 같이 정의됨

17
For Credit Risk

Copula 함수의 이해
Elliptical Copula

Archimedean Copula

18
For Credit Risk

Copula 함수의 이해 □ Dependency Structure

Gaussian or t-Copula Dependent Symetary 상승과 하락이 대칭 Clayton Copula Gumbel Copula
For Credit Risk

Low tail dependence (하락 dependency) Upper tail dependence(상승 dependency)
19

Copula 함수의 이해 Example 4: Gumbel copula

Cθ (u, v ) = exp{−[(− ln u) + (− ln v ) ]

Gu

θ

θ 1 /θ

}

ϕ(t ) = (− ln t )θ

1≤θ <∞

1<θ < ∞

θ =3

θ →1
Independence

1 1 1/1 CθGu } =1(u, v) = exp{−[(− ln u) + (− lnv) ]

= exp{ln(u) + ln(v )} = uv = Π

θ →∞
For Credit Risk

CθGu =∞ (u, v ) = min(u, v )
20

Perfect Dependence

Copula 함수의 이해 Gumbel Gumbel Copula Copula

21
For Credit Risk

Copula 함수의 이해 Example 5: Clayton copula (explicit)

CθClayton (u, v ) = (uθ + v θ − 1)−1 / θ

ϕ (t ) =
−1

t −θ − 1

θ

θ >0

Bivariate Clayton copula (δ = 3).
δ

θ →0 θ →∞
For Credit Risk

Independence Perfect Dependence
22

Copula 함수의 이해 Clayton Clayton Copula Copula

23
For Credit Risk

Copula 함수의 이해 Example 5: Frank copula (explicit)

Frank

⎡ (e −θ u − 1)(e −θ v − 1) ⎤ (u, v ) = − ln ⎢1 + ⎥ θ ⎣ (e −θ − 1) ⎦ 1

eθ t − 1 ϕ (t ) = − ln θ e −1
−1

−∞ < θ < ∞

24
For Credit Risk

Copula 함수의 이해 Dependency Dependency Structure Structure □ Upper tail dependence

λupper = lim Pr(Y ≥ FY−1 (u ) X ≥ FX−1 (u ))
u →1

1 − Pr(Y ≤ FX−1 (u)) − Pr(Y ≤ FY−1 (u)) + Pr( X ≤ FX−1 (u), Y ≤ FY−1 (u)) = 1 − Pr(Y ≤ FX−1 (u)) 1 − 2 u + C (u , u ) −1 −1 = lim Pr( Y ≤ F ( u )) = Pr( Y ≤ F (u )) = u X Y u →1 1− u −1 −1
□ Lower tail dependence

Pr( X ≤ FX (u ), Y ≤ FY (u ) = C (u , u )

λLower = lim Pr(Y ≤ FY−1 (u ) X ≤ FX−1 (u ))
u →0

□ The tail area dependency measure depends on the copula and not on the marginal distribution
25
For Credit Risk

P r ( X ≤ F X− 1 ( u ) , Y ≤ F Y− 1 ( u ) ) = P r ( Y ≤ F X− 1 ( u ) ) C (u, u ) = lim u →0 u

Copula 함수의 이해 Gaussian Copula : the coefficients of lower tail and upper tail dependence Dependency Structure Structure Dependency
Proof
For Credit Risk

λU = λL = 2lim x →∞ Φ( x 1 − ρ / 1 + ρ ) = 0
T-copula : the coefficients of lower and upper tail dependence

λU = λL = 2tν +1(− ν + 1 1 − ρ / 1 + ρ )
Clayton copula

λU = 0

λ L = 2 −1 / θ
λL = 0

Gumbel copula

λU = 2 − 21 / θ
Frank copula

λU = 0

λL = 0

1 − 2u + C (u , u ) 1 − 2u + exp{−[( − ln u )θ + ( − ln u )θ ]1/ θ } λU = lim = lim u →1 u →1 1− u 1− u θ 1/ θ 1 − 2u + exp{−{2( − ln u ) ] } = lim u →1 1− u 1/ θ −2 + 2 = lim = 2 − 21/ θ u →1 −1 26

Copula 함수의 이해 Dependence Dependence

Dependence

27
For Credit Risk

Copula 함수의 이해 Concordance Concordance function function
□ 두 Data Set(two pairs)에 대하여 다음과 같이 정의할 수 있음

(xi , y i ) ( xj, y j )
• 일치(concordance) • 불일치(disconcordance)

(x1 − x2 )(y1 − y2 ) > 0

( x1 − x2 )(y1 − y2 ) < 0

Concordance function

τ = Pr[( x1 − x2 )(y1 − y2 ) > 0] − Pr[(x1 − x2 )(y1 − y2 ) < 0]

= Pr[(x1 − x2)(y1 − y2) > 0] − {1 − Pr[(x1 − x2)(y1 − y2) > 0]}

= 2{Pr[x1 > x2, y1 > y2] + Pr[x1 < x2, y1 < y2] } −1
28
For Credit Risk

Copula 함수의 이해 Spearman rank correlation
Nonparametric Correlation

ρS ( X , Y ) = 12∫
0

1

∫ {C(u, v ) − uv}dudv
0

1

ρS ( X , Y ) = ρ(F1( X ), F2 (Y ))
= Cov[F1( X ), F2 (Y )]

σ 2[F1( X )]σ 2[F2 (Y )]

=

Cov[F1( X ), F2 (Y )] 1 1 12 12

= 12Cov[F1( X ), F2 (Y )]
= 12∫
1 0 0 1 1

∫ {C(x, y ) − F
0 1

1

F1 ( x )

( x )FF2 ( x )(y )}dxdy

= 12∫

= 12∫
0

0 1

∫ {C(x, y ) − xy}dxdy ∫ C(x, y )dxdy − 3
0
∞ ∞ −∞ −∞

F1( X ), F2 (Y )
1 2

: Standard uniform distribution

참고 : For Credit Risk

Cov[F1( X ), F2 (Y )] =

∫ ∫ {F (x, y ) − F (x)F (y )}dxdy
29

Copula 함수의 이해 Kendall rank correlation
Nonparametric Correlation

ρτ ( X , Y ) = 4∫
0

1

∫ C(u, v )dC(u, v ) − 1
0

1

ρτ ( X , Y ) = 2Pr[(x1 − x2 )(y1 − y2 ) > 0] − 1
= 2 E [1( x1 − x 2 )( y 1 − y 2 ) > 0 ] − 1

= 2E[1( x1 − x2 )> 0 1(y1 − y2 )>0 + 1( x1 − x2 )< 01(y1 − y2 )< 0 ] − 1
= 2[∫ ∫∫∫ 1( x1 − x2 )> 0 1(y1 − y2 )> 0 dF (x2 , y2 )dF (x1, y1 )
R

+ ∫ ∫∫∫ 1( x1 − x2 )< 0 1(y1 − y2 )< 0 dF (x2 , y2 )dF (x1, y1 )] − 1
R

= 2 2∫ ∫∫∫ 1( x1 − x2 )> 0 1(y1 − y2 )> 0 dF (x2 , y2 )dF (x1, y1 ) − 1
R

= 4∫
R
1

∫ F (x , y )dF (x , y ) − 1
1 1 1 1

= 4∫
0

∫ C (u, v )dC (u, v ) − 1
0

1

R

Pr[( x1 − x2 )(y1 − y2 ) > 0] + Pr[( x1 − x2 )(y1 − y2 ) < 0] = 1
For Credit Risk

30

Copula 함수의 이해

Kendall’s Tau for Archmedean Copula

ϕ(u) τ = 1 + 4∫ ' du 0 ϕ (u)
1

Clayton copula

τ =

θ
2+θ

Gumbel copula

τ =1−

1

θ

Frank copula

τ =1−4

θ

0

tdt exp(t ) − 1

θ2
31

For Credit Risk

Copula 함수의 이해

Testing for Elliptic Distributions - Do we really need to worry about using Copulae at all?

• Estimating Parametric and Non-parametric Copulae - Maximum Likelihood; • Full MLE or sequential / concentrated ML - Semi-parametric estimation • Impact of the misspecification of marginal distributions on the estimation of copula parameters and measures of dependence - Non-parametric methods • How to choose the Correct Copula? - AIC - Goodness of Fit Tests - Simulated Cox Tests • Copula Quantile Regression and Tail area Dependence - Quantile conditional relationships implied by a given copula - Quantile dependency measures
32
For Credit Risk

Copula 함수의 이해 Gaussian Gaussian copula copula의 의 시뮬레이션 시뮬레이션

33
For Credit Risk

Copula 함수의 이해 Student’s Student’s t-copula t-copula의 의 시뮬레이션 시뮬레이션

34
For Credit Risk

Sign up to vote on this title
UsefulNot useful