You are on page 1of 19

# Which copula is the right one?

V. Durrleman, A. Nikeghbali & T. Roncalli∗
Groupe de Recherche Op´erationnelle
Cr´edit Lyonnais
France
August 25, 2000

Abstract
In this paper, we give a few methods for the choice of copulas in financial modelling.

1

Introduction

Copulas reveal to be a very powerful tool in the banking industry, more especially in the modelling of the different
sorts of existing risks. It allows a multidimensional framework, giving up the gaussian assumption which we
´, Durrleman, Nikeghbali,
know is incorrect in asset modelling and the study of extremal events. Bouye
Riboulet and Roncalli  review different financial problems and show how copulas could help to solve
them. For example, they used copulas for operational risk measurement and the study of multidimensional
stress scenarios. But one of the difficulty is in general the choice of the copula. This article deals with this
problem.
Let’s recall some elementary facts about copulas. All these results are developped in Deheuvels  and
Nelsen . Let (X1 , . . . , XN ) ∈ RN be a random vector with cumulative distribution function
F (x1 , . . . , xN ) = P (X1 ≤ x1 , . . . , XN ≤ xN )

(1)

and marginal functions
Fn (xn ) = P (Xn ≤ xn ) ,

1≤n≤N

(2)

A copula function C of F is defined as a cumulative distribution function of a probability measure with support
N
in [0, 1] such that:
1. ∀ 1 ≤ n ≤ N , 0 ≤ un ≤ 1
Cn (un ) = C (1, . . . , 1, un , 1, . . . , 1) = un

(3)

2. ∀ (x1 , . . . , xN ) such that xn is a continuity point of Fn (1 ≤ n ≤ N ),
F (x1 , . . . , xN ) = C (F1 (x1 ) , . . . , FN (xN ))

(4)

The following results about copulas are very useful. The complete proofs of these results are established in
Deheuvels .
∗ Corresponding author: Groupe de Recherche Op´
erationnelle, Bercy-Expo — Immeuble Bercy Sud — 4`e ´
etage, 90 quai de Bercy
— 75613 Paris Cedex 12 — France; E-mail adress: thierry.roncalli@creditlyonnais.fr

1

the convergence of F(m) to a distribution function F with continous margins Fn . . when m −→ ∞. 0 ≤ un . . 1] . What we propose in this paper is different methods to cope with the uncertainty about the real underlying dependence structure of the studied phenomena. . xn is a continuity point of Fn . . m ≥ 1 is a sequence of probability cumulative distribution functions in RN . . . . Fn −→ Fn punctually. . we give several methods to choose the ‘optimal’ copula. . uniquely defined on the image set (F1 (x1 ) . xN ) such that ∀ 1 ≤ n ≤ N . we present the parametric estimation with a given family. Every copula C is continuous and statisfies the following inequality (∀ 1 ≤ n ≤ N . . If h1 . In a second section. . . xn . . . . . one takes a parametric family of copulas among many existing others and fit it to the data by estimating the parameters of the family. . . . xN ) = c (F1 (x1 ) . The density of the joint distribution F is given by the following expression f (x1 . . . . (b) if C is the unique copula function associated to F. . . weak convergence of the associated probability measure (Deheuvels ). compact with any of the following topologies. So it appears that copulas are in fact the dependence structure of the model. all based on the empirical copula introduced be Deheuvels . non decreasing mappings of R in itself. uniform convergence on [0. . C(m) −→ C (with the topology of C). 2.1 Parametric estimation with a given family The method of the maximum likelihood Let us first consider the case where both the copula C and the margins Fn are continuous. . . . . . Thus the choice of the copula that is going to fit the data is very important. vN )| ≤ N X |un − vn | (5) n=1 3. . . The set C of all copula functions is convex. uN ) = ∂ C (u1 . . FN (xN )) of the points (x1 . . . XN ) is also a copula function of (h1 (X1 ) . If all the marginal functions are continuous. uN ) − C (v1 . vn ≤ 1) |C (u1 . . . and if C(m) is a copula function associated to F(m) . . 4. is equivalent to both conditions: (m) (a) ∀ 1 ≤ n ≤ N . equivalent on N C: punctual convergence. FN (xN )) N Y fn (xn ) (6) n=1 where fn is the density of the margin Fn and c is the density of the copula c (u1 . . un . . . This can provide biased results since according to the dependence structure selected the obtained results might be very different. If F(m) . . there does not exist a systematic rigorous method for the choice of copulas: nothing can tell us that the selected family of copula will converge to the real structure dependence underlying the data. . . any copula function of (X1 . Every distribution function F has at least one copula function. . . . . . uN ) ∂ u1 · · · ∂ un · · · ∂ uN 2 (7) . All the information about the dependence is contained in the copula function. hN (XN )). Fn (xn ) . 2 2. then the copula function of F is unique. . . © ª 5. hN are monotone. . Nevertheless. un . . . . . Then. .1. . Usually. . .

co.000210 0. 0. F2 (x2 )) − C (F1 (x1 − 1) . . F2 (x2 − 1)) 3 This (10) database is on the web site of the London Metal Exchange (http://www. .00 is the set of parameters of the margins and the copula. . .000247 0. N = 2. We consider the joint distribution of the asset returns Aluminium Alloy (AL) and Copper (CU). . . . FN xtN + ln fn xtn t=1 (8) t=1 n=1 with θ the K × 1 vector of parameters1 . the Radon-Nikodym density is Pr {(X1 . xN )} = N (−1) 1 X i1 =0 ··· 1 X i +···+iN (−1) 1 C (F1 (x1 − i1 ) . Fn (xn − in ) . If the margins are discrete + 2 (the support is {x− n . . 3 .000241 0. .uk) and it is available in a Gauss format on request. xn }. XN ) = (x1 . . Then. .T Let X = {(xt1 . . . .lme. .4 28. The expression of the log-likelihood is also ` (θ) = T X T X N ¡ ¡ ¢ ¡ ¢ ¡ ¢¢ X ¡ ¢ ln c F1 xt1 .158681 t-statistic -0. .52 0.597625 std. Let θˆML be the maximum likelihood estimator. . . F2 (x2 − 2)) + C (F1 (x1 − 1) . . J −1 (θ0 ) (9) with J (θ0 ) the information matrix of Fisher (Davidson and MacKinnon ). . we retreive the result Pr {(X1 . X2 ) = (x1 . . .013700 -0. . xtN )}t=1 denote a sample.err.34 0.9 p-value 0. . we use the LME database3 .000324 0. . . . XN ) = xt1 .5 -0. . it verifies the property of asymptotic normality and we have ´ √ ³ ¡ ¢ T θˆML − θ0 −→ N 0. . . u2 ) = −α ln 1−e − 1−e 1−e (13) 1 − e−α The log-likelihood of the observation t is then ¶ µ t ¶¶¶¸ · µ µ µ t x2 − m2 x1 − m1 +Φ − `t (θ) = ln α (1 − exp (−α)) exp −α Φ σ1 σ2 · µ µ µ t ¶¶¶ µ µ µ t ¶¶¶¸2 x1 − m1 x2 − m2 ln (1 − exp (−α)) − 1 − exp −αΦ 1 − exp −αΦ + σ1 σ2 " µ ¶2 # " µ ¶2 # 1 1 ¡ 2 ¢ 1 xt1 − m1 1 1 ¡ 2 ¢ 1 xt2 − m2 − ln (2π) − ln σ1 − + − ln (2π) − ln σ2 − (14) 2 2 2 σ1 2 2 2 σ2 We then obtain the following results Parameters m1 m2 σ1 σ2 α 1θ 2 If estimate -0. . . FN (xN − iN )) (11) iN =0 The expression of the log-likelihood becomes ` (θ) = T X © ¡ ¢ª ln Pr (X1 .017175 4.65 71.95 71. .00 0.000192 0. . Fn xtn .000258 0.00 0. F2 (x2 )) − C (F1 (x1 ) . . We assume that the margins are gaussians and that the dependence structure is the Frank copula: µ ¶ £¡ ¢ ¡ ¢¡ ¢¤ 1 −1 −α −αu1 −αu2 C (u1 . x2 )} = C (F1 (x1 ) . . . xtN (12) t=1 In order to clarify the underlying methods and to help the reader to reproduce the results. . . .

θn and α are the vectors of parameters of the parametric marginal distribution Fn and the copula C. θN . . . . Joe and Xu  suggest then to use the Jacknife method to estimate it. . θn .2. . . . The method consists in transforming the data (xt1 . because it requires to estimate jointly the parameters of the margins and the parameters of the dependence structure. θˆN . V −1 (θ0 ) (18) matrix of Godambe. . . . . . Using a close idea of the IFM method. Fn xn . Note also that the IFM method could be viewed as a special case of the generalized method of moments with an identity weight matrix. Let us define a score function in the following way g (θ) = ¡with 1V (θ0 ) the information ¢ ∂θ1 ` . . . . FN xtN . α (17) t=1 This two-step method is called the method of inference functions for margins or IFM method (Joe and Xu ³ ´ ˆ ˆ ˆ ). ∂θN `N . . . u ˆtN ) — using the empirical distributions — and then estimate the parameter in the following way: T X ¡ t ¢ α ˆ = arg max ln c u ˆ1 . θˆn . . We could also perform the estimation of the univariate marginal distributions in a first time θˆn = arg max `n (θn ) := arg max T X ¡ ¢ ln fn xtn . The IFM estimator θIFM is then defined as the vector θ1 . FN xN . θn t=1 (15) t=1 n=1 with θ = (θ1 . u ˆtn . . 4 . . α ˆ . Note that the IFM method could be t t ˆ viewed as a CML method with u ˆn = Fn xn . θN . θ1 . we remark that the parameter vector α of the copula could be estimated without specifying the marginals. θn . .2 The IFM method The problem with the ML method is that it could be computational intensive in the case of high dimension. The estimation of the covariance matrix requires to compute many derivatives. . . α + ln fn xtn . we use the notation utn to designe ³ ´ either u ˆtn or Fn xtn . In the rest of the paper. θn (16) t=1 and then estimate α given the previous estimates c α ˆ = arg max ` (α) := arg max T X ³ ³ ´ ³ ´ ³ ´ ´ ln c F1 xt1 . . . . . . . . . Fn xtn . The log-likelihood (8) could then be written as (Joe and Xu ) ` (θ) = T X T X N ¡ ¡ t ¢ ¡ t ¢ ¡ t ¢ ¢ X ¡ ¢ ln c F1 x1 . . . . α). . . The Godambe information matrix takes the form (Joe ): ¡ ¢> V (θ0 ) = D−1 M D−1 (19) h i h i > > where D = E ∂g (θ) /∂θ and M = E g (θ) g (θ) . we call it the canonical maximum ³ ´likelihood method or CML. xtN ) into uniform variates (ˆ ut1 . . . . . α (20) t=1 In this case. u ˆtN . ∂α `c . However. θN . θˆn . the copula representation splits the parameters into specific parameters for marginal distributions and common parameters for the dependence structure (or the parameters of the copula). Like the ML estimator. . . Because this method is based on the empirical distributions. θˆ1 . we could show that it verifies the property of asymptotic normality and we have ´ √ ³ ¡ ¢ T θˆIFM − θ0 −→ N 0. α ˆ could be viewed as the ML estimator given the observed margins (without assumptions on the parametric form of the marginal distributions). . .

δ. ςn = Φ−1 (Fn (xn )) and Fn the univariate GH distribution4 . ρ) = ¢ 1 λn αn2 − βn2 2 ´ ³ p λn − 12 λn δn Kλn δn αn2 − βn2 n=1 αn ¡ N Y 1 N (2π) 2 (24) > and ς = (ς1 . δ) δ 2 + (x − µ)2  q  ×Kλ− 1 α δ 2 + (x − µ)2 exp (β (x − µ)) 2 with a (λ. Here are the IFM estimator for the gaussian and the student copulas: • Gaussian ¢ ¡ copula Let ςt = Φ−1 (ut1 ) . The IFM method could not be apply to the distribution (21). Even if this distribution has more parameters than the previous one. . α. because it reduces computational aspects.We remark that the copula representation presents some very interesting properties for the estimation. α. Then we have ρˆIFM = 4 The T 1X > ς ςt T t=1 t (27) corresponding density function is given by f (x) =   1 (2λ−1) 4 a (λ. α. then we have 353 parameters. β. Durrleman. . . . δ. ρ) |ρ| 2 n=1 "N ¶# µ µ q ¶ Y ¢ 1 ¡ 2 2 exp β > (x − µ) − ς > ρ−1 − I ς × (23) Kλn − 12 αn δn + (xn − µn ) 2 n=1 where A (λ. Φ−1 (utn ) . β. One of the important issue for the estimation is the existence of analytical solution of the IFM (or CML) estimator. . Riboulet and Roncalli  present a multivariate version of the Bouye univariate generalized hyperbolic distribution based on the gaussian copula. the IFM method consists in estimating the parameters of the margins (N maximum likelihood with 5 parameters per estimation) and the estimation of the parameters of the copula (which is straightforward in the case of the gaussian copula — see below). In the case of the copula-based distribution. we could perform the estimation more easily. Φ−1 (utN ) . we have 425 parameters for N = 25. ςN ) . . . α. δ. ρ) δ 2 + (x − µ) ρ−1 (x − µ) µ q ¶ ¡ ¢ > −1 2 ×Kλ− N α δ + (x − µ) ρ (x − µ) exp β > (x − µ) (21) 2 where K denotes the modified Bessel function of the third kind and ¡ 2 ¢1λ α − β > ρβ 2 ³ p ´ A (λ. δ. . α. . Consider for example the multivariate generalized hyperbolic distribution (Prause ) ³ ´ 14 (2λ−N ) > f (x) = A (λ. ´. . . This is for example a key point in the Finance industry. This distribution has 1 2 N [(N + 9)] parameters. ρ) = N N (2π) 2 αλ− 2 δ λ Kλ δ α2 − β > ρβ (22) The number of parameters is equal to 21 N [(N + 3)] + 3. Nikeghbali. δ) = √ 1λ α2 − β 2 2   p 1 2παλ− 2 δ λ Kλ δ α2 − β 2 5 (25) (26) . β. For example. α. β. Its density is "N # ´ 1 (2λn −1) Y³ 1 2 4 2 f (x) = δn + (xn − µn ) 1 A (λ. β. Suppose that N = 25. . β.

00 0.80 36. The IFM estimate of the ρ matrix for the Student copula is ρˆIFM = ρˆ∞ .016680 4. We then define µ t ¶ xn − m ˆn utn = Φ σ ˆn (29) (30) PT The paramater α is also estimated by maximising `c (α) = t=1 `ct (α) with £ ¡ ¡ ¢¢¤ `ct (α) = ln α (1 − exp (−α)) exp −α ut1 + ut2 − £ ¡ ¡ ¢¢ ¡ ¢¤2 ln (1 − exp (−α)) − 1 − exp −αut1 1 − ut2 The Godambe covariance matrix is then computed with the score function   ∂`1 (m1 . Durrleman. σn ) = − ln (2π) − ln σn − 2 2 2 σn t=1 We denote m ˆ n and σ ˆn the corresponding ML estimates. . 2 " µ ¶2 # T X 1 1 ¡ 2 ¢ 1 xtn − mn n ` (mn . Nikeghbali.3 p-value 0.156818 t-statistic -0. σ1 ) /∂m1  ∂`1 (m1 .7 28. 6 . . tν (un ) .81 41.42 0.000455 0. 4. Bouye Roncalli  propose to estimate the ρ matrix with the following algorithm: 1.42 0.440463 std. σ1 ) /∂σ1    2  g (θ) =   ∂` 2 (m2 . σ2 ) /∂σ2  ∂`c (α) /∂α (31) (32) The results are the following and are very closed to those given by the ML method.• Student copula ¡ ¢ t −1 t −1 t ´.578972 — the IFM and CML estimates are not very closed. . tν (uN ) . . ρˆm+1 is obtained using the following equation ρˆm+1 = 1 T µ ν+N ν ¶X T t=1 1+ ςt> ςt 1 > −1 ˆm ςt ν ςt ρ (28) 3. Parameters m1 m2 σ1 σ2 α estimate -0.0 -0. . Let us now consider the previous example. σ2 ) /∂m2   ∂` (m2 . because the IFM method does take into account the marginal distributions. . We have for n = 1.err. Let ρˆ0 be the IFM estimate of the ρ matrix for the gaussian copula.000325 0.000207 0. 0. Repeat the second step until convergence — ρˆm+1 = ρˆm (:= ρˆ∞ ). we obtain α ˆ CML = 3.013316 -0. .000256 0.000256 0.000320 0. . 2.00 0. Riboulet and Let ςt = t−1 ν (u1 ) .00 In the case of the CML method.

. In some cases. .i. We are now interested in modelling the dependence structure with consistency. Note that this estimate is closer to the CML estimate than the ML and IFM estimates. . We assume that F is continuous so that the copula associated to F is unique. . With the LME example. But the problem is that C is not unique. . 7 .3 Estimation based on the dependence measures We could also estimate the parameters such that they fit the dependence measures. Let Xt = t (X1t . .49958.2. XN ) ∈ RN be an i. PT If δu stands for the Dirac measure with u ³ ∈ RN . . x 1 n=1 N (rnt ) t the order statistic and {r1t . We note x . xn ] its empirical function. this is to say thanks to a copula that is going to converge to the real underlying dependence structure. the Spearman’s rho or the upper tail dependence measure. . Gaussian Gumbel FGM % ¡ ¢ ρ = 2 sin π6 % numerical solution 3% τ¡ ¢ ρ = sin π2 τ −1 α = (1 − τ ) 9 2τ λ X −1 ln 2 ln (2 − λ) X Otherwise. we define ´µ ˆ = T1 t=1 δXt the empirical measure associated n o QN (t) (t) ˆ with a sample of X. and F (x1 . for example the Kendall’s tau. .d. We define the point estimator θˆ as the solution of the following problem (Lehmann and Casella ) θˆ = arg min L (θ) (33) θ∈Θ with Θ the parameters space. gˆ the observed criterion values and g the criterion functions. In the case of one-parameter bivariate copula. we obtain %ˆ = 0. 3.1 The Deheuvels or empirical copula We present in this paragraph the notion of empirical copula as introduced by Deheuvels . 3 Non parametric estimation In this section. we could perform the estimation with only one dependence measure. xN ) = µ ˆ ]−∞.4390. rN } the rank statistic of the sample which are linked by the relationship xn = t ˆ of the empirical distribution xn . we don’t assume anymore that we have a parametric copula. . . analytical solutions are available (see the table below for some examples). It is possible to introduce the empirical copula of the sample as any copula C ˆ ˆ F. Let L (θ) be a loss function. . we use a root finding procedure. so it comes that α ˆ = 3. This is the case of the Frank copula (Frees and Valdez ): % τ 12 (D1 (α) − D2 (α)) α 4 = 1 − (1 − D1 (α)) α = 1− (35) with Dk (x) the Debye function (Abramowitz and Stegun ). . . people use a quadratic loss distribution > L (θ) = [ˆ g − g (θ)] W [ˆ g − g (θ)] (34) with W a weight matrix. . that’s why Deheuvels  proposes the following way to cope with the problem. In most cases. sequence with distribution F and margins Fn . .

. : 1 ≤ n ≤ N. The empirical copula C ˆ (T ) is any empirical copula of order T . T T ¶ (36) T N 1 XY 1[rt ≤t ] T t=1 n=1 n n = (37) is an empirical copula... . For example.... . T T T ¶ 2 X = ··· i1 =1 2 X µ i +···+iN (−1) 1 ˆ C iN =1 tn − in + 1 tN − iN + 1 t1 − i1 + 1 . Deheuvels [1979. .... tn = 0. . τˆ = ˆ c T − 1 t =1 t =1 i =1 i =1 T T T T T T T T 1 2 2 (40) (41) 2 We could also define the sample version of the quantile dependence function λ (u) = Pr {U1 > u|U2 > u} = ¯ (u. ˆ We could now define the analog of the Radon-Nikodym density for the empirical copula C µ cˆ t1 tn tN .. T T T by µ ˆ C t1 tN .. u) C 1−u (42) and the upper tail dependence measure λ = lim − λ (u) u−→1 8 (43) . 2... −ˆ c .. T T T =1 ¶ (39) Note that empirical copulas permits to define the sample version of dependence measures. The relationships between empirical copula distribution and frequency are µ ˆ C tn tN t1 .... T T T ¶ = t1 X ··· i1 =1 tN X iN µ i1 in iN cˆ . ˆ (T ) in order to define the order of the copula... ˆ c . The empirical measure µ ˆ (or the empirical distribution function F) by both ˆ n. the the Spearman’s rho and Kendall’s tau take the following form (Nelsen ) %ˆ = ¶ ¶ T T µ µ 12 X X ˆ t1 t2 t1 t2 C ..... ˆ c . If C for example). T T T ¶ (38) cˆ is called the empirical copula frequency (Nelsen ).. that is the dimension of the We introduce the notation C sample used to construct it. .... − T 2 − 1 t =1 t =1 T T T2 1 and 2 ¶ µ ¶ µ ¶ µ ¶¶ T T t1 −1 tX 2 −1 µ µ i1 i2 t1 i2 i1 t2 2 X X X t1 t2 . then C ˆ (T ) → C with the topology of C (uniform convergence 3.. .1981] obtains then the following conclusions: ˆ is uniquely and reciprocally defined 1.. ..ˆ ∈ C defined on the lattice Definition 1 Any copula C ½µ ¶ ¾ t1 tN L= ....... (a) the empirical measures of each coordinate F ˆ on the set L. (b) the values of an empirical copula C ˆ defined on L is in distribution independent of the margins of F.. ....

CU) data 3. we could compute also a confidence interval for λ Figure 1: Empirical copula of the (AL.We have ˆ λ ¡ ¢ µ ¶ ˆ t. Nguyen and Olsen  to characterize markov processes. t 1−C t T T =2− T 1 − Tt (44) ˆ = limt→T λ ˆ (t/T ). we use a grid equal to data is then 109. We present only the two dimensional case. Currie and Tawn ). By assuming ˆ (Coles. we use the idea of Li. we use these approximations to estimate non parametrically the underlying dependence structure. Mikusinski. and λ We consider now the example of the LME data. but the generalization is straightforward. 9 1 25 for the plots. Nevertheless. In the figure 2. we have reported the empirical quantile dependence function λ dependence function with the Frank copula and the parameter equal to the previous ML estimate.2 Copula approximation In this paragraph. The order T of the empirical . We have represented in the figure 1 the corresponding ˆ and the empirical copula5 . a normal distribution. These approximations are very interesting to study the properties of the ∗-product of copulas defined by Darsow. Their motivation was to present “certain approximations that lead naturally to a stronger convergence than uniform convergence”. Below. 5 We have 2713 observations in the database. Sherwood and Taylor  on approximation of copulas.

n (x) denotes the Bernstein polynomial Bi. it is well known that the copula Bn (C) converges to the copula C in the strong sense. v) = Bt1 .n (v) C i=1 j=1 i j .2.n (u) Bj. Bn (C) defined by Bn (C) (u.n (x) = µ ¶ n i n−i x (1 − x) i (45) Li. n n ¶ (46) is a copula too.T (v) C BT C T T t =1 t =1 1 (47) 2 ´ ³ ˆ (T ) is a copula that will (uniformly) converge to we know thanks to what we have said previously that BT C the real copula C ´ ³ ˆ (T ) −→ C (48) BT C 10 . v) = n X n X µ Bi.Figure 2: Quantile dependence function 3. t2 ˆ (T ) (u. Now.1 Approximation by Bernstein polynomials Let Bi. Mikusinski. Moreover. if we consider the two place function µ ¶ T X T ´ ³ X ˆ t1 . Sherwood and Taylor  showed that for any copula C.T (u) Bt2 .

we know that the approximated copula converge to the underlying dependence structure. v)| ≤ 2n−1 (51) If we now take the two place function ³ CT T X T ´ ³ ´Z X 2 ˆ ˆ C(T ) (u. t2 − 1 ∆t1 . In general. .T (x) dx 0 v χt2 . v) = n ∆i. Nikeghbali and Roncalli ). Sherwood and Taylor  give some properties of Cn (C): £ ¤ i 1. Nevertheless.n (y) dy (49) 0 i=1 j=1 where µ ¶ µ ¶ µ ¶ i−1 j i j−1 i−1 j−1 . we assume that we have a finite subset of copulas C˜ ⊂ C. n n ¶ 0 µ −C 2. these approximation methods do not preserve the upper tail dependence.n is the characteristic function of the interval i−1 n . We have reported in the figure 3 its approximation. n × £ j−1 j ¤ n . t2 − C ˆ (T ) t1 .T (y) dy (52) where µ ¶ ¶ ¶ ¶ µ µ µ ´ ³ ˆ (T ) = C ˆ (T ) t1 .1]2 |C (u. Another important point is that the convergence of the concordance measures is satisfied (Durrleman.2. If C˜ corresponds to the Archimedean family.2. 4 Selection criteria based on empirical copulas In this section. v) = T ∆t1 . However. One of the main interest of this construction is that the associated measure is continuous. one could use the perturbation method of Durrleman.t2 C(T ) 0 t1 =1 t2 =1 Z u χt1 . 11 .v)∈[0. Cn (C) is the copula of the doubly stochastic measure whose mass in every square of the form i−1 n . v) − Cn (C) (u. In the figure 4.2 The checkerboard approximation Another consistent way for modelling the dependence structure is to consider a checkerboard approximation Z u Z v n X n X 2 Cn (C) (u. Moreover. . n .n (x) dx χj. we have represented the measure for different orders T . n is uniformly distributed and equal to the total mass in that square in the doubly stochastic measure of the original copula ∆i. Nikeghbali and Roncalli  to obtain the desired upper tail dependence. we can deduce that CT C the real underlying copula ³ ´ ˆ (T ) −→ C CT C (54) 3.3 Applications We consider the previous empirical copula of the (AL. Li. And we would like to know which copula fits best the data. This is a real problem in financial modelling. t2 − 1 + C ˆ (T ) t1 − 1 .CU) data. we could solve the problem using results on Kendall’s processes. t2 − C ˆ (T ) t1 − 1 . Cn (C) approximate C uniformly sup (u.j (C) = C i j . we could use a distance based on the discrete Lp norm.j (C) χi. −C +C (50) n n n n n n £ ¤ i and χi. we consider the selection problem of ‘optimal’ copula.t2 C (53) T T T T T T T T ³ ´ ˆ (T ) is a copula which converges (uniformly) to From what we have said previously. In the general case. Mikusinski.3.

CU) data with Bernstein polynomials Figure 4: Probability density function of the Bernstein copula with different orders 12 .Figure 3: Approximation of the empirical copula of the (AL.

. . Let X be a vector of N random variables. . we could first estimate the parameters by IFM ou CML method ˜ Note that if C is an absolutely two-dimensional copula. α = 2 and α = 3. the expression in order to reduce the cardinality of C. ut2 + ln ϕ0 ut1 ϕ0 ut2 − 3 ln −ϕ0 C ut1 . . ut2 We consider our previous example with three copula families. + ϕ (uN )) if ϕ (un ) ≤ ϕ (0) C (u1 . Genest and Rivest  have developed an empirical method to identify the copula in the Archimedean case. ϕn (u) κn−1 (u) n! (56) A non parametric estimate of K is given by T X ˆ (u) = 1 K 1[ϑi ≤u] T t=1 with (57) T ϑi = 1 X 1 t i t i T − 1 t=1 [x1 <x1 . . Another procedure consists in comparing u − K (u) and u − K ).xN <xN ] (58) ˆ Frees et Valdez  propose to use a The idea is then to choose the Archimedean copula which gives K. . .. . Ghoudi and Re K (u) = u + N X n (−1) n=1 with κn (u) = ∂u κn−1 (u) ∂u ϕ(u) and κ0 (u) = 1 ∂u ϕ(u) . Genest. ˆ ˆ (u) (Genest and Rivest QQ-plot of K and K. the Gumbel copula with parameters α = 1. In general.. . un . α = 1. C˜ consists not in one parametric family with different values of the parameters. ϕ0 (u) < 0 and ϕ00 (u) > 0 for all 0 ≤ u ≤ 1. C3 and C4 .1 The case of Archimedean copulas Genest and MacKay  define Archimedean copulas as the following:  N X   −1 ϕ (ϕ (u1 ) + .578972 . Frank. of the log-likelihood is ¡ ¡ ¢¢¢ ¡ ¡ ¡ ¢¢ ¡ ¡ ¢ ¡ ¢¢ (59) `ct (α) = ln ϕ00 C ut1 . for example Gumbel. but C˜ is the set of different families..462803 0. C the associated copula with generator ϕ and K the function defined by K (u) = Pr {C (U1 ..0 1 + u2 − 1 ³ ´ −1 −1 −α −αu −α ln 1 + (e − 1) (e − 1) (e−αv − 1) 13 α ˆ CML 1. UN ) ≤ u} (55) ´millard  showed that Barbe. . u2 ) α α 1 α ´ exp − ((− ln u1 ) + (− ln u2 ) ) ³£ ¤− α1 ´ −α max u−α . For each family. . Cook-Johnson. we notice that the procedure of Genest and Rivest  identifies more easily the ‘optimal’ copula and C2 seems to be the best dependence structure. Suppose that C˜ is the set of four archimedean copulas. etc. C2 . . The CML estimation gives the following results Copula ϕ (u) α C1 Gumbel (− ln u) C2 Cook-Johnson C3 Frank α−1 (u−α − 1) ´ ³ − ln exp(−αu)−1 exp(−α)−1 ³ C (u1 . We have represented in the figures 5 and 6 the two graphical procedures.1 of Nelsen  for a list of Archimedean copulas).5.708430 3. We note them respectively C1 . . . . uN ) = n=1   0 otherwise with ϕ (u) a C 2 function with ϕ (1) = 0. . (see the table 4. In general. We consider our LME example.4. + ϕ (un ) + .

Figure 5: QQ-plot procedure of Frees and Valdez  Figure 6: Graphical procedure of Genest and Rivest  14 .

we obtain the following ˆ K = the distance is defined as d2 K. Cook-Johnson and Frank copulas 4. could define the ‘optimal’ copula as the copula which gives the minimum distance between K (u) and K 2 For example. . we could take the Kolmogorov distance. But we have to specify the distance we consider.63 × 10−4 for the Gumbel copula.4 × 10−4 for the Cook-Johnson copula values: d2 K. ³ ´ ˆ K = 1. With the LME data. etc.We have reported in the figure 7 the values of u − K (u) for these copulas. d2 K. Figure 7: Comparison of the Gumbel. . . Let us consider the L norm ÃZ ! 21 Z ° ¯ ¯2 ³ ´ ° ° ° ¯ ¯ ˆ j . which is then our ‘optimal’ copula. Because K (u) is a distribution. J) and a finite subset of K copulas C˜ = {Ck } empirical copulas C 1≤k≤K . the Hellinger distance. In the case of the L norm. Here we assume we have a finite subset of copulas C˜ ⊂ C.10 × 10−4 for the Frank copula. Ck = °C ˆ j − Ck ° := ˆ j (u) − Ck (u)¯ du (60) d2 C ··· ¯C 2 L [0. i2 ³ ´ Z 1h ˆ (u) du. and d2 K. it appears that any distance based on the Lp norm would not be proper. we ˆ (u). As there exists more than one empirical copula. All we know about any empirical copula is its values on the lattice L that we could compute thanks to the data. For example. let’s assume that we have different 2 ˆ j (j = 1. K (u) − K 0 ³ ´ ³ ´ ˆ K = 2. What we suggest here is to consider the distance between each considered copula and the empirical copula. and we are interested in knowing which one of the copulas in C˜ fits best the data (there might be parametric copulas or non parametric copulas or Fr´echet upper bounds for example).1]N 15 . .2 Selecting a copula among a given subset of copulas We are not interested here in the different ways of constructing an empirical copula (by interpolation or approximation for example). ˆ K = 11.

an explicit expression is very difficult to obtain. That’s why we suggest in these cases to estimate the parameter of the parametric copula C (u.. For high dimensions and some copulas (for example. T . 4. the MM1-MM5 families in Joe )..... . tn = 0. We then define the point estimator θˆ as the solution of the following minimization problem Ã θˆ = arg min θ∈Θ Xh i2 ˆ (T ) (u) − C (u.³ ´ ˆ j .. ..578972 0. As all empirical copulas take the same values on L. we obtain the following results for the LME data: 16 .462803 0.. T .708430 3.....47 × 10−3 4. the maximum likelihood method requires the cross derivatives of the copula.491794 ˆ (T ) . we could use a quasi-Newton algorithm like the BFGS method so that we don’t need the Hessian matrix anymore. θ) C ! 12 (62) u∈L To perform the numerical optimization...79 × 10−3 The Frank copula is then always our ‘optimal’ copula. Cook-Johnson and Frank copulas) and the Gaussian copula.. Ck can be different problems since the ‘optimal’ copula depends on j and thus the copula that minimize d2 C for different values of j. we suggest to take a distance based on the discrete Lp norm ° ° ³ ´ °ˆ ° ˆ (T ) . . T T T ¶¸2 ! 12 (61) ³ ´ ˆ (T ) . The cross derivatives could then be computed with a numerical algorithm based on finite differences. Ck . We obtain the following results: C1 C2 C3 C4 Copula Gumbel Cook-Johnson Frank Normal α ˆ CML 1.. However. For example..39 × 10 16. .Ck ) d¯2 (C T −3 7.. Moreover.5 × 10−3 4.3 Estimating parameters in ‘untractable’ cases We propose here a method for estimating parameters (of some parametric families of copulas) which are difficult to compute by maximum likelihood. these algorithms produces important numerical errors in high dimension... This is for example the case for some asymmetric extreme copulas because of the shape of the likelihood function. . And we suppose that the set C˜ corresponds to the set of the previous paragraph (Gumbel. T We use the LME example. but just the first order derivatives. T T T ¶ µ − Ck tn tN t1 . The advantage We consider that the best copula in the family C˜ is the copula which minimizes d¯2 C of this method ©¡ ¢ is that it does not depend ª on the behavior of the empirical copulas out of the lattice L = tN t1 : 1 ≤ n ≤ N. Ck . . θ) by taking the loss function L (θ) equal to the distance (61). this sort of estimation is coherent with what we have previously said about the choice of the dependence struture for modelling. we might encounter some If we consider that the best copula in C˜ is the copula which minimizes d2 C ³ ´ ˆ j . Ck d¯2 C = °C − C k° (T ) 2 Ã = L T X t1 =1 ··· T X tn =1 ··· T · X µ ˆ (T ) C tN =1 tn tN t1 . Moreover.

this is the Normal copula which appears to be ‘optimal’. that is we have assumed that the margins correpond to the empirical ones. the IFM and ML estimators are biased (see the figure 9). They are all based (sometimes not directly) on the Deheuvels or empirical copula. We assume that the bivariate distribution F correspond to the Normal copula with parameter 0.79 × 10−3 α ˆ L2M 1. References  Abramowitz. If the margins are not well specified.462803 0.1 × 10−3 3.5 × 10−3 4. in this case. Stegun . Dover 17 . ninth edition. we assume that the margins of AL and CU are gaussians. done by choosing different parameter families and by doing a ML optimization for each family. But one of the difficulty is in general the choice of the copula. Handbook of Mathematical Functions. 4.Ck ) d¯2 (C T −3 7.01 × 10 12. We suppose now that we fit a distribution with a Normal copula and two gaussian margins. Because the margins are wrong.06 × 10−3 α ˆ L2M is the point estimator based on the discrete L2 norm. To clarify this point. 5 Conclusion Copulas are a powerful tool in financial modelling. We could certainly notice that the gaussian margins are not appropriate. we then obtain the following values for the distances: Estimation method CML IFM ML L2M ˆ (T ) . We remark that this is not the case of the CML estimator. In the previous example. We remark that we improve significantly the distance.Ck ) d¯2 (C T −3 5. we see clearly that the impact of the margins is very important.708430 3. we consider a simulation study. we could then find an optimal copula into the subset C˜ which is in fact irrelevant to give the good dependence structure.4 The influence of the margins on the choice of the dependence structure ˜ we have to reduce the cardinality of the subset of C in a first time. M.39 × 10 16. Handbook of Mathematical Functions. If we compare the level curves of the Deheuvels copula.93 × 10−3 We remark that the discrete L2 norm for IFM and ML methods are larger. we give a few methods to solve this problem. In the case of the Frank copula. In the second section.528938 ˆ (T ) . However.554518 1. With this example. we have perform the estimation step using the CML method.Ck ) d¯2 (C T −3 4. we obtain the figure 8.A. That’s why estimation based on CML method is very important.389611 0.491794 ˆ (T ) .5 × 10−3 12. Moreover. If we notice that there exist significant differences between CML and ML (or IFM) methods.C1 C2 C3 C4 Copula Gumbel Cook-Johnson Frank Normal α ˆ CML 1. ninth edition.93 × 10−3 3.8 × 10−3 3. CML Frank copula and ML Frank copula. we are going to see that the margins play an important role to define the optimal copula.47 × 10 11.47 × 10−3 4.5 and two margins F1 and F2 which are a student distribution (F1 = t2 and F2 = t3 ).578972 0. Dover. and I. In this article.056350 3. This is In order to define a ‘tractable’ set C. that indicates that the margins are not well appropriate.

CML Frank copula and ML Frank copula Figure 9: Comparison of the density of the CML IFM and ML estimators when the margins are wrong 18 .Figure 8: Comparison of empirical copula.

.. . P. Caract´erisation compl`ete des lois extrˆemes multivari´ees et de la convergence des types extrˆemes. J. Mikusin ˇ ep´an (Eds. Publications de l’Institut de Statistique de l’Universit´e de Paris. C. Chapmann & Hall. 29-50  Durrleman. ETHZ. P. R. Working Paper  Embrechts. Illinois Journal of Mathematics. . Department of Mathematics and Statistics.F. Valdez. New York ´ ski. in V. Roncalli . An Introduction to Copulas. A. Dissertation.D. 65. Journal of Multi Barbe. Correlation and dependency in risk management: properties and pitfalls. Dordrecht ´.. P. Xu . On Kendall’s process. V. C. Nguyen and E. 2. Taylor . C. Acad´emie Royale de Belgique – Bulletin de la Classe des Sciences – 5e S´erie. W. Sherwood and M.J. University of British Columbia. Copulas and markov processes. Nikeghbali. North American Actuarial Journal. Departement of Mathematik.. 88. P. E. 36-4. E. Estimation and Inference in Econometrics. H. The joy of copulas: Bivariate distributions with uniform marginals. V. K. Lancaster University. Rivest . Casella . 26. 1034-1043  Joe.W. Distributions with Given Marginals and Moment Problems. 1-25  Genest. and Risk Measures. .A. A. 197-229 ˇ e ˘pa ´ n . Cr´edit Lyonnais. Z¨ urich. Oxford  Deheuvels. 1-36  Deheuvels. P. Statistical inference procedures for bivariate Archimedean copulas. H. University of Freiburg 19 . Lectures Notes in Statistics. and L. Kluwer Academic Beneˇs and J.. MacKinnon . V. and G. Ghoudi and B. Cr´edit Lyonnais. Dordrecht  Nelsen. Working Paper  Coles.B. McNeil. Tawn . St˘ Publishers.. 166  Lehmann. 73. Groupe de Recherche Op´erationnelle.L. Copulas approximation and new families. Straumann . Multivariate Models and Dependence Concepts. X. 23. Understanding relationships using copulas. 40. Riboulet and T. R. P.T. Financial Derivatives.´millard . A. La fonction de d´ependance empirique et ses propri´et´es — Un test non param´etrique d’ind´ependance. Springer Verlag. and T. Publications de l’Institut de Statistique de l’Universit´e de Paris. Olsen . S.  Li. St Academic Publishers. E. Journal of the American Statistical Association. Working Paper  Frees.). and J. MacKay . H. and J. Groupe de Recherche Op´erationnelle. 274-292  Deheuvels. 58.. The estimation method of inference functions for margins for multivariate models. and E. New York  Prause. K. G. Copulas for finance  Bouye — A reading guide and some applications. Nikeghbali. . Technical Report. . 280-283  Genest. Theory of Point Estimation. Distributions with Given Marginals and Moment Problems. On approximation of copulas. London  Joe. A non parametric test for independence. Oxford University Press. Durrleman. Roncalli . Re variate Analysis. Currie and J. 139. Department of Statistics. B. Kluwer  Beneˇ s.. and D. Monographs on Statistics and Applied Probability. The Generalized Hyperbolic Model: Estimation. .J. American Statistician. 600-642  Davidson. and J. second edition. Working Paper  Darsow. Genest. Dependence measures for extreme value analyses. and J. Springer Verlag.