A proposal of a bivariate Conditional Tail Expectation

Transcription

1 A proposal of a bivariate Conditional Tail Expectation Elena Di Bernardino a joint works with Areski Cousin b, Thomas Laloë c, Véronique Maume-Deschamps d and Clémentine Prieur e a, b, d Université Lyon 1, ISFA, Laboratoire SAF c Université de Nice Sophia-Antipolis, Laboratoire J-A Dieudonné e Université Joseph Fourier, Grenoble

2 Given an univariate continuous and strictly monotonic loss distribution function F X, VaR α (X ) = Q X (α) = F 1 X (α), α (0, 1). Shortcoming of VaR measure: Var is not a coherent risk measure (in the "Artzner sense", see Artzner, 1999) Var does not give any information about thickness of the upper tail of the loss To overcome problems of VaR Conditional Tail Expectation (CTE): CTE α (X ) = E[ X X VaR α (X ) ] = E[ X X Q X (α) ], (e.g. for properties of the CTE α (X ) see Denuit et al., 2005).

3 Our Bivariate Conditional Tail Expectation In the literature A new bivariate Conditional Tail Expectation Denition (Generalization of the CTE in dimension 2) Consider a random vector (X, Y ) satisfying regularity properties, with associate distribution function F. Let L(α) = {F (x, y) α}. For α (0, 1), we dene ( ) E[ X (X, Y ) L(α) ] CTE α (X, Y) =. E[ Y (X, Y ) L(α) ] Distributional approach to risk model: Our CTE α (X, Y ) does not use an aggregate variable (sum, min, max...) in order to analyze the bivariate risk's issue. Conversely CTE α (X, Y ) deals with the simultaneous joint damages considering the bivariate dependence structure of data in a specic risk's area (α-level set : L(α)).

4 Our Bivariate Conditional Tail Expectation In the literature Several bivariate generalizations of the classical univariate CTE have been proposed; mainly using as conditioning events the total risk or some univariate extreme risk in the portfolio. We recall for instance: CTE α (X, Y ) sum = E[ (X, Y ) X + Y > Q X +Y (α) ], CTE α (X, Y ) min = E[ (X, Y ) min{x, Y } > Q min{x,y } (α) ], CTE α (X, Y ) max = E[ (X, Y ) max{x, Y } > Q max{x,y } (α) ]. Remark: Conditioning events are the total risk or some univariate extreme risk in the portfolio (aggregate variable).

5 Our Bivariate Conditional Tail Expectation In the literature Two dierent developments of research : 1 Problem of estimating the level sets L(c) = {F (x) c}, with c (0, 1), of an unknown distribution function F on R 2 + with a plug-in approach; in order to provide a consistent estimator for CTE α (X, Y ) [Article in a non-compact setting with applications in multivariate risk theory, with Laloë L., Maume-Deschamps V. and Prieur C., submitted]. 2 Analysis of the CTE α (X, Y ) as risk measure: study of the classical properties (monotonicity, translation invariance, positive homogeneity,...), behavior with respect to dierent risk scenarios and stochastic ordering of marginals risks [Article Some proposals about bivariate risk measures, with Cousin A.].

6 Estimation of the level sets of a density function: Polonik (1995), Tsybakov (1997), Baíllo et al. (2001), Biau et al. (2007). Estimation of the level sets of a regression function in a compact setting: Cavalier (1997), Laloë (2009). Estimation of general compact level sets: Cuevas et al. (2006). An alternative approach, based on the geometric properties of the compact support sets: Cuevas and Fraiman (1997), Cuevas and Rodríguez-Casal (2004).

7 Goal: We consider the problem of estimating the level sets of a bivariate distribution function F. More precisely our goal is to build a consistent estimator of L(c) := {F (x) c}, for c (0, 1). Idea: We consider a plug-in approach that is L(c) is estimated by L n (c) := {F n (x) c}, for c (0, 1), where F n is a consistent estimator of F.

8 From a parametric formulation of L(c) (see Belzunce et al. 2007). Case: Survival Clayton Copula with marginals (Burr(1), Burr(2))

9 We state consistency results with respect to two proximity criteria between sets: the Hausdor distance and the volume of the symmetric dierence (physical proximity between sets). Problem of compactness property for the level sets we estimate. Figure: (left) Hausdor distance between sets X and Y ; (right) λ(x Y ), where λ stands for the Lebesgue measure on R 2 and for the symmetric dierence.

10 Let R 2 + := R 2 + \ (0, 0), F the set of continuous distribution functions f : R 2 + [0, 1] and F F. Given an i.i.d sample {X i } n i=1 in R2 + with distribution function F, we denote F n (.) = F n (X 1, X 2,..., X n,.) an estimator of F. Dene, for c (0, 1), the upper c-level set of F F and its plug-in estimator : L(c) := {x R 2 + : F (x) c}, In addition, given T > 0, we set L n (c) := {x R 2 + : F n (x) c}, {F = c} = {x R 2 + : F (x) = c}. L(c) T = {x [0, T ] 2 : F (x) c}, L n (c) T = {x [0, T ] 2 : F n (x) c}, {F = c} T = {x [0, T ] 2 : F (x) = c}. For any A R 2 + we note A its boundary.

11 For r > 0 and λ > 0, dene E := B({x R 2 + : F c r}, λ), m := inf x E ( F ) x, M H := sup x E (HF ) x, B(x, ξ) is the closed ball centered on x and with positive radius ξ, ( F ) x is the gradient of F evaluated at x, (HF ) x is the matrix norm induced by Euclidean distance of the Hessian matrix in x. Hausdor distance between A 1 and A 2 is dened by: d H (A 1, A 2 ) = inf{ε > 0 : A 1 B(A 2, ε), A 2 B(A 1, ε)}, where B(S, ε) = x S B(x, ε); A 1 and A 2 are compact sets in (R 2 +, d).

12 H: There exist γ > 0 and A > 0 such that, if t c γ then T > 0 such that {F = c} T and {F = t} T, d H ({F = c} T, {F = t} T ) A t c. Assumption H is satised under mild conditions. Proposition Let c (0, 1). Let F F be twice dierentiable on R 2 +. Assume there exist r > 0, ζ > 0 such that m > 0 and M H <. Then F satises Assumption H, with A = 2. m

13 From now on we note, for n N, and for T > 0, F F n = sup F (x) F n(x), F F n T = sup F (x) F n(x). x R 2 + x [0,T ] 2 Theorem (Consistency Hausdor distance) Let c (0, 1). Let F F be twice dierentiable on R 2 +. Assume that there exist r > 0, ζ > 0 such that m > 0 and M H <. Let T 1 > 0 such that for all t : t c r, L(t) T1. Let (T n) n N be an increasing sequence of positive values. Assume that, for each n, F n is continuous with probability one and that Then for n large enough, F F n 0, a.s. d H ( L(c) T n, L n(c) T n ) 6 A F F n T n, a.s., where A = 2 m.

14 d H ( L(c) T n, Ln(c)T n ) converges to zero and the quality of our plug-in estimator is obviously related to the quality of the estimator F n. Note that in the case c is close to one the constant A = 2 could be m large. In this case, we will need a large number of data to get a reasonable estimation. In order to overcome the problem of F n is continuous with probability one it can be considered a smooth version of estimator.

15 Let us introduce the following assumption: A1 There exist positive increasing sequences (v n) n N and (T n) n N such that v n [0,T n] 2 F F n p λ(dx) P n 0, for some 1 p <. Theorem (Consistency volume) Let c (0, 1). Let F F be twice dierentiable on R 2 +. Assume that there exist r > 0, ζ > 0 such that m > 0 and M H <. Let (v n) n N and (T n) n N positive increasing sequences such that Assumption A1 is satised and that for all t : t c r, L(t) T1. Then it holds that p n d λ (L(c) T n, L n(c) T n ) P 0, n with p n an increasing positive sequence such that p n = o ( 1 p+1 vn /T p p+1 n ).

16 Theorem Consistency volume provides a convergence rate, which is closely related to the choice of the sequence T n. A sequence T n whose divergence rate is large implies a convergence rate p n quite slow. Theorem Consistency volume does not require continuity assumption on F n. Analysis in the case: F n the bivariate empirical distribution function. Problem: optimal criterion for the choice of T n related with the p in Assumption A1.

17 Estimation of CTE α (X, Y ) Denition Consider a random vector (X, Y ) with associate distribution function F F. For α (0, 1), we dene the estimated bivariate α-conditional Tail Expectation ĈTE α (X, Y ) = n i=1 X i 1 {(Xi,Y i ) L n(α)} n i=1 1 {(X i,y i ) L n(α)} n i=1 Y i 1 {(Xi,Y i ) L n(α)} n i=1 1 {(X i,y i ) L n(α)}. (1) We introduce truncated versions of the theoretical and estimated CTE α : CTE T α (X, Y ) = E[(X, Y ) (X, Y ) L(α) T ], T ĈTEα (X, Y ), using L(α) T and L n (α) T i.e. the truncated versions L(α) and L n (α).

18 Estimation of CTE α (X, Y ) Theorem (Consistency of ĈTE α (X, Y )) Under regularity properties of (X, Y ), Assumptions of Theorem Consistency volume and with the same notation, it holds that T β n T n CTE n α (X, Y ) ĈTEα (X, Y ) P 0, (2) n r 2(1+r) where β n = min{ pn, a n }, with r > 0 such that the density f (X, Y ) L 1+r (λ) and a n = ( ) o n. The convergence in (2) must be interpreted as a componentwise convergence. In the case of a bounded density function f (X, Y ) we obtain β n = min{ p n, a n }.

19 On the choice of T n ( ) In the case of the bivariate empirical d.f. F n : β n = o n r 2 r 6 (1+r) 3 (1+r) /T n. ( In the case of a bounded density function f (X,Y ), β n = o n 1 6 /T 2 3 n ). Obviously it could be interesting to consider the convergence: T n CTEα (X, Y ) ĈTEα (X, Y ). the speed of convergence will also depend on the convergence rate to zero of CTEα (X, Y ) CTE T n α (X, Y ), then of P[(X, Y ) L(α) \ L(α) T n ]. We remark that (P[X T n or Y T n ]) 1 is increasing in T n, whereas the speed of convergence is decreasing in T n. This kind of compromise provides an illustration on how to choose T n, apart from satisfying the assumptions of consistency results above.

20 Study of CTE α (X, Y ) as risk measure (stochastic order for random variables and random vectors) CTE α (X, Y ) and Kendall distribution function or bivariate probability integral transformation (BIPIT) F (X, Y ). In this work we provide asymptotic results for a xed level c... Problem of constant A for c 1: Plug in method with Ln (c) := {(x, y) R 2 + : F (x, y) c}, for c 1, with F a tail estimator of F (Bivariate extreme value theory; rst part of my PhD thesis).

21 Artzner, P. and Delbaen, F. and Eber, J. M. and Heath, D., Coherent measures of risk. Math. Finance 9(3), (1999), A. Baíllo, J.A. Cuesta-Albertos, A. Cuevas, Convergence rates in nonparametric estimation of level sets. Statistics and Probability Letters 53, (2001), F. Belzunce, A. Castaño, A. Olvera-Cervantes, A. Suárez-Llorens, Quantile curves and dependence structure for bivariate distributions. Computational Statistics & Data Analysis 51, (2007), G. Biau, B. Cadre, B. Pelletier, A graph-based estimator of the number of clusters. ESAIM, Probability and Statistics 11, (2007), J. Cai, H. Li, Conditional tail expectations for multivariate phase-type distributions. Journal of Applied Probability 42(3), (2005), L. Cavalier, Nonparametric estimation of regression level sets. Statistics 29, (1997), A. Cuevas, R. Fraiman, A plug-in approach to support estimation. Annals of Statistics 25, (1997),

22 A. Cuevas, W. González-Manteiga, A. Rodríguez-Casal, Plug-in estimation of general level sets. Australian and New Zealand Journal of Statistics 48, (2006), A. Cuevas, A. Rodríguez-Casal, On boundary estimation. Advances in Applied Probability 36, (2004), M. Denuit, J. Dhaene, M. Goovaerts, R. Kaas. Actuarial Theory for Dependent Risks. Wiley, (2005). W. Polonik, Measuring mass concentration and estimating density contour clusters - an excess mass approach. Annals of Statistics 23, (1995), A. Rodríguez-Casal, Estimacíon de conjuntos y sus fronteras. Un enfoque geometrico. Ph.D. Thesis, University of Santiago de Compostela, (2003). A.B. Tsybakov, On nonparametric estimation of density level sets. Annals of Statistics 25, (1997),