This action might not be possible to undo. Are you sure you want to continue?

BooksAudiobooksComicsSheet Music### Categories

### Categories

### Categories

Editors' Picks Books

Hand-picked favorites from

our editors

our editors

Editors' Picks Audiobooks

Hand-picked favorites from

our editors

our editors

Editors' Picks Comics

Hand-picked favorites from

our editors

our editors

Editors' Picks Sheet Music

Hand-picked favorites from

our editors

our editors

Top Books

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Audiobooks

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Comics

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Sheet Music

What's trending, bestsellers,

award-winners & more

award-winners & more

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

of Statistics / La Revue Canadienne de Statistique, Vol. 33, No. 3, Dependence Modelling: Statistical Theory and Applications in Finance and Insurance (Sep., 2005), pp. 415-427 Published by: Statistical Society of Canada Stable URL: http://www.jstor.org/stable/25046188 . Accessed: 09/05/2011 12:31

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at . http://www.jstor.org/action/showPublisher?publisherCode=ssc. . Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Statistical Society of Canada is collaborating with JSTOR to digitize, preserve and extend access to The Canadian Journal of Statistics / La Revue Canadienne de Statistique.

http://www.jstor.org

de vie des Une seconde et femmes examine la pr?sence de d?pendance positive par quadrant entre les esp?rances hommes par pays.2005.g. il propose et justifie une inference reposant sur une m?thode simul?e ?multiplicateurs et une m?thode de bootstrap. risk management. In a financial setting. (1) is equivalent to P(X >x. multiplier method. Y < y) > P(X < x)P{Y < y).g. e. Levy (1992). Drouef-Mari & Kotz (2001) or Lai & Xie (2003). in finance. empirical process. he pro poses and justifies inference relying on a simulation-based multiplier method and a bootstrap method. type Kolmogorov-Smirnov la d?pendance Pour un tel test. 1.62P05. Denuit. Shaked & Shanthikumar (1994).. copula. restrictions that relate to the copula representation of positive quadrant dependence. No. no matter what the strength of this dependence is. insurance and risk management has emphasized the importance of see. Il explore aussi le comportement ? distance finie des deux m?thodes ? l'aide d'exp?riences deMonte-Carlo.62H20. Une premi?re illustration empirique est faite pour des donn?es am?ricaines de sinistres en assurance. Formally. Un test de type Kolmogorov-Smirnov R?sum? restrictions : L'auteur provenant consid?re de un test convergent de la repr?sentation pour de la d?pendance positive positive de par quadrant par quadrant l'ensemble sous forme des complet de copule. Neil & Straumann (2000). Abstract: The author considers secondary 60E15.Hie Canadian Journal of Statistics Vol. nonparametric estimator. INTRODUCTION The concept of positive quadrant dependence (PQD) was introduced by Lehmann (1966). for all (x. MSC 2000: Primary 62G30. Mc PQD.Y>y)> P(X > x)P(Y > y). He also explores the finite-sample behaviour of both methods with Monte Carlo experiments. A first empirical illustration is given for American insurance claim data. but for risk measures and derivative prices re lated to a portfolio of two PQD financial assets. For such a test. two random variables X and Y are said to be PQD if.. (1) Of course. the underestimation of most insurance premiums involving a portfolio of these two claims if the independence assumption ismade instead. Kolmogorov-Smirnov type of test of the complete set of a consistent. Dhaene & Goovaerts (1996). loss severity distribution. 33. A second one examines the presence of positive quadrant dependence in Ufe expectancies at birth of males and females across countries. One advantage of this dependence structure is that it allows the risk manager to compare directly the sum of PQD random variables with the corresponding sum under the independence assumption. Dhaene & Ribas (2001). 3. e. Recent work P(* < a. see. Pages 415-427 La revue canadienne de statistique 415 A Kolmogorov-Smirnov type test for positive quadrant dependence Olivier SCAILLET Key words and phrases: Bootstrap. immediately allows one to conclude PQD. Two random variables are said to be PQD when the probability that they are simultaneously large (or small) is at least as great as itwould be were they independent. y) R2. positive quadrant dependence. Implications of PQD are also found in reliability theory and several other fields. the same holds true. (2) . Embrechts. The comparison is in the sense of different stochastic orderings Inferring that two claims are expressing the common preferences of rational decision makers. We refer the reader to Denuit 8c Scaillet (2004) for an extensive discussion of examples of the application of PQD in finance and actuarial sci ences.

who extend and justify the procedure of McFadden (1989) (see also Abadie 2002. The first one Section 3. In fact. their tests rely on distance tests or intersection-union pairwise comparisons made at a fixed number of arbitrary chosen points. Dardanoni & Forcina (1999).e.Y < y) C{P(X < x). as well as Linton. a member of the Farlie-Gumbel-Morgenstern family.. X and Y are PQD if. TEST STATISTIC AND ASYMPTOTIC PROPERTIES We a setting made of pairs of independent and identically distributed observations of a random vector taking values in R2. In Section 2. they are either based on tests for inequality constraints. From a technical point of view. Proofs are gathered in an Appendix. C(u. Note that parametric copulas may or may not exhibit PQD per se. for testing PQD. A second one examines the presence of PQD in life expectancies at birth of males and females across countries. (l)-(2). and analyze the asymptotic properties of the test for PQD. For example.. Note further that nega tive quadrant dependence (NQD) is defined analogously if we substitute < for > in (1). or observed returns of financial assets.denotes an independent version of the random vector (X. we describe the test statistic. This is not a desirable The current proposal does not feature since it introduces the possibility of test inconsistency. Davidson & Duelos (2000). the inequality (1) can be written in terms of the copula C of the two random variables. The margins are denoted by F and G.v)=uv. where (X. Y). Y)1.v) Recall that by definition >C?(u. Y) and (X. These data may correspond (Xl. In Section 2. This indicates that PQD is a property of the underlying copula. v) G [0. Their procedures are inspired by traditional stochastic dominance tests as in Ander son (1996). the copula C = P(X < x. It can thus be considered as a special case of comparisons of pairs of bivariate components..416 SCAILLET Vol. families that only allow PQD are the Cook-Johnson family and the Gumbel family. We give some concluding remarks and discuss some potential extensions for dimensions higher than two in Section 6. and (X. Yi).. Maasoumi & Whang 2005) leading to consistent tests of stochastic dominance.g. 3 which enjoys a similar interpretation (with "small" replaced by "large"). respectively. we discuss two practical ways to compute the P-values relies on a simulation-based multiplier method. However. we explore the finite-sample behaviour of both methods with Monte Carlo ex periments. suffer from this drawback.has independent i. Y)x. the Frank family.. that is similar to a Kolmogorov-Smirnov test. a(X) and b(Y) are PQD for any strictly increasing functions a and b. of the complete set of restrictions that relate to the copula representation (3) of PQD. for all (i?. Clearly. and is not influenced by the marginals. In this paper. A first empirical illustration is given for the US insurance claim data examined in Section 5. (3) is such that (Sklar 1959). induces PQD when the parameter is positive. since (1) is equivalent to the condition that. and NQD when the parameter is negative. The paper is organized as follows. (Xn.1]2. we differ from their work by the multivariate aspect of our distributional setting as well as the use of empirical copula processes In instead of univariate empirical processes. 4. Y) to Considering that of (X. On the contrary. Ob serve that Denuit & Scaillet (2004) have already suggested some nonparametric ways to test for PQD. Y)L have identical univariate marginals.. Yn) to either observed individual losses on insurance contracts. (X. while the second relies on a bootstrap method. We follow closely Barrett & Donald (2003). e. Y)1. No. their properties and their use.P(F < y)}. 33. PQD appears as a comparison of the joint distribution of (X. Joe (1997) and Nelsen (1999) for detailed explanations of copulas. consider . distributions with identical marginals in terms of stochastic dominance. see. we propose a consistent test of PQD. and only if. the amounts of claims reported by a given policy holder on different guarantees in a multi-line product. or the Gaussian family.

. v).v)C(u\v') and from the weak convergence properties of the empirical copula process Its proof..v.. where sup u. is true.i. 1.. C2(i?'. then lim P(reject Ho) < P(S > c) = a(c).1.v). v) = uv for all (it. the weak convergence of y/? (Cn C) towards a completely-tucked Brownian sheet was originally given by Deheuvels (1981)...l.. who developed a Kolmogorov Note Smirnov Since we wish test for independence. v) = uv C(u." where c is some critical value that the properties of the test. respectively. u. for some (u. v.. D(u. t/)fi(ti.vf) = = E{Gc(u... to test for PQD. p. t/) C2(ii.ti. V (?n (2004). whose covariance function is dC(u. given in the Appendix.ve[o.l)-C2(ti/. v){0(l.v)/du Ci(u. l]2) the set of all locally bounded real functions on [0. t/)?l(u. v.1]2. v) > uv for all (u. 1.v. 1) C2(u'. This lemma follows = C(u Au\v A v') ? C(u..t/)}.2005 Following Deheuvels DEPENDENCE POSITIVE QUADRANT (1979). v) < uv Sn = V? and a test based on the decision will be discussed The following later. . D) converges weakly in ???([0. 1) v. Let c be a positive finite constant.u'.i?'.vf x A y = min(x. namely Ho Hi we consider the test statistic : C(u.y). u\ v') Ci(u'.. v) = uv . y). v!. 1.v) e [0. Xn and Y?. v) G [0. v')Sl(u. v!.l] S = Gc(u.t. relies on Theorem 4 in Fermanian. where Fn and Gn are the empirical cumulative distribution functions computed from X\.v [0. l]2). = Assume that the copula derivatives function C has continuous partial ? = Then^(Dn C2(u. l]2). v) ? [0. t/)} -Ci(u. Radulovic & C).') /or a/Z i?. v. 389) Wegkamp see also Stute (1984) or G?nssler & holds true in a larger space under weaker assumptions.v)/dv to a tight mean zero Gaussian process Gc ttG{u. In the special case C = CL..v) = -f]l{Fn(Xi) n r?f < %Gn(Yi) < v}.u.t/)il(l. let us define the empirical copula function by 417 Cn(u. l]2. (u. l]2.i] rule "reject Ho if Sn > c.v) anddC(u. which shows that the result stated by van der Vaart & Wellner (1996. Proposition i) IfHo 1. v). vf) Ci(v!. l]2. with equality when C(u. i/)?l(u. u'. l]2. result characterizes sup Dn(u. ? that uniform almost sure convergence is a by-product of this type of weak convergence. : C(u. 1. v){??(w.1].. and likewise for nGn(1?). Let Dn(u. Stute (1987) for weak convergence of the same process in the Skorokhod space ?>([0.v. Observe that Cn is actually a function of the ranks of the observa tions since nFn(Xi) is the rank of Xi among X\. v) [0.v)Gc(uV)} ??(u. LEMMA 1. Ci(w.. Xn.i/)ft(l. wftere ?2(i?.i?'. t/) G [0. v) and denote by ???([0. t*'.Cn(u. Yn.

Since Gc(ti.. Hereafter we rely on two different methods to simulate P-values.1) random variables that sample. Since the distribution of the test statistic depends on the underlying copula. v) i obtained from smoothed versions of the empirical copula process. one can use the nonparametric estimators ci. Of course.t.G-\v)}lfn{F-\u)} = C2. Consistent estimates are easily c?>n(t?.t.418 ?) If Wo isfalse. The second part of the result indicates that the test is capable of detecting any violation of the full set of restrictions of the null hypothesis. v) Ci(u. The first part of the result provides a random variable that dominates the limiting random variable corresponding to the test statistic under the null hypothesis.. The first part also implies that if one could find a c to set the a(c) to some desired probability level (say the conventional 0. then this would be the significance level for composite null hypothe ses in the sense described by Lehmann (1986).?(?. c2. in order to make the result operational. Indeed. v)Bc(l. we cannot i. and Hansen (1996). Furthermore. we need to find an appropriate critical value c. 3. the process generated from GCw(ti. then SCAILLET Vol.n(u. t/). The inequality tells us that the test will never reject more often than a(c) when the null hypothesis is satisfied. No. W)-L?Ui where = 1.n(t*. where Be v) = Bc(ti. Glidden (1999).01). this is not an easy task. is easily is a tight Brownian bridge on [0. are independent of the data independent and identically distributed N(0.) and = ^Hn{F-\u).G-l{v)}lgn{G-\v)}. See Barrett & Donald (2003) for a similar use in the context of a stochastic dominance test. SIMULATING P-VALUES 3.) = ^YfUix[l{Fn(Xi)<u1Gn(Yi)<v}-Cn(%v)] <*. directly simulate under the independence hypothesis. 3 n-?oo lim P(rejectHo) = 1. v) is a consistent estimate of Ci(u.2... 33. recall that the null hypothesis is not independence. Multiplier method. the probability of rejection will asymptotically be exactly a(c) when C = CL. In this section. we use a simulation-based method that exploits themultiplier central limit theory discussed in Section 2. For example.) ^Hn{F-l{u). Guay & Scaillet (2003) for other applications. .1.05 or 0.?(?.t. Such a procedure would not reflect the dependence of the distribution of the test statistic on the underlying copula. 1) C2(ti. y/n x x [l{Fn(Xi) <u}-u] [I{Gn(Yi) <v}-v]. t/)-L ?t/i . Therefore.e. The idea is to rely on artificial pseudo-random numbers to simulate a process that is iden tical but (asymptotically) Un denote a sequence of independent of Gc> To do this. let Ui.1]2 (see the proof of Lemma 1). v)Bc(u.9 of van der Vaart & Wellner (1996).. draw from the independent copula.

Then an application of the continuous mapping theorem shows that we get a simulated copy of the bounding random variable that appears in Proposition 1.. a test for PQD based on the rule "reject Ho if lim P(reject Ho) <a if Ho is true.. Assuming that a < " pn < a satisfies thefollowing n-?oo 1/2. F]*). The C* Let (X*. we need to rely on Monte Carlo methods to approximate the probability and a grid to approximate the supremum in amanner analogous to that in the previous subsection. fn and gn. that the replication number and the grid mesh can be chosen tomake the approximations as accurate as one desires given time and computer constraints. The following result provides the decision rule in this environment. is justified by the next statement. ifHo is false. n?fco = 1 lim P (reject Ho) ifHo isfalse. 1/2. . The P-value is simply approximated by 1 where Note R r=l the averaging is done on R replications.ve[o.2005 POSITIVE DEPENDENCE QUADRANT 419 where estimates of the joint cumulative distribution function H and themarginal densities / and g are obtained from kernel based estimators Hn.. we use Monte Carlo methods to approximate the probability and a grid to approximate the supremum. a test for PQD based on the rule "reject Ho if PROPOSITION 3. In practice.r is computed from a fine grid on [0.i] {C*(u. v) Cn(u. for which similar results can be shown to hold as well. and 5n. see Abadie (2002) and Barrett & Don ald (2003) for similar applications in the context of stochastic dominance tests. Maasoumi & Whang (2005) for its use in stochastic dominance tests. respectively. Bootstrap method. and the empirical copula function built from this bootstrap sample. PROPOSITION 2. v)}.v * where Pu is the probability function associated with the normal random variable U and is condi tional on the realized sample. ^ u.>Sn).. The multiplier method can be justified by showing that the simulated process converges weakly to an identical independent copy of the Gaussian process <Gc. Again. Let us further take S*n and define = y/n sup u. Assuming that a < pn < a" satisfies thefollowing 7l-?00 n?fco lim P(rejectHo) <a lim P(rejectHo) = 1 ifHo is true. K Then th? bootstrap method = p(s. see Fermanian & Scail let (2003) for details and proofs of asymptotic properties of such estimators. l]2. second method relies on the standard bootstrap. see Linton. The P-value can be estimated from Pn = Pu (sup GCn (U. An alternative resampling technique could be subsampling.V)> Sn }. (X*. Y*) be a random sample drawn from the observed pairs of data. 3.2.

samples.007 Table 2 gathers results concerning the power of the testing procedure when the true copula is a Frank copula. and the rejection rates are computed for the multiplier method and the bootstrap method with respect to the two conventional significance levels of a = 0. They match low and mod copula.21. 1000 Monte Carlo simulations are performed. Genest (1987) or Nelsen eter 0 are 6 G {-1.495. -3} for the Frank copula. Samples are generated with both margins corresponding to an exponential distribution with a unit parameter. a Gaussian copula.. -. for short) on a single claim.. 0 G {-.10. see.32.. The sample size is fixed at n = 200. -2. the true copula is the independent copula. In each case. No. -. we examine the performance of the Kolmogorov-Smirnov type test in small The grid is made of the values (u. Table 1: Observed significance level of the test when the underlying copula is the independent copula..003 . A prominent ex is the loss and allocated loss adjustment expenses (ALAE. The reported numbers show that both testing procedures tures. The replication number R to approximate the P-value is set equal to 1000..95}.31}. ?. ?. and 0 G {?.021 .103 . Various processes in casualty insurance involve correlated pairs of variables. erate negative dependencies r G {?.1. 33.027 n = 200 n = 400 . 0. v) evenly spaced inside {0. have nice power properties under different negative dependence struc 5. e. ample .002 .95} x estimator of the derivatives of the copula func 0. EMPIRICAL ILLUSTRATIONS 5.024 . ?.014 . Then Proposition 1 suggests that the test should reject the null hypothesis with a frequency close to the chosen nominal significance level... Note that the numerical results below remain exactly the same if we use other strictly monotonie continuously differentiable cumula as Gaussian or Student margins tomimic financial returns) and tive distribution functions (such generators. MONTE CARLO RESULTS SCAILLET Vol.01.020 n = 100 n = 200 . or a Farlie-Gumbel-Morgenstern copula inducing NQD (?).05 and a = 0. These parametric families are often used in actuarial and financial applications. 3 In this section. and are (1999).11.024 n = 50 n = 100 .01 n = 50 . while the nonparametric {0. a = 0.064 . This experiment should give us some idea about the validity of the asymptotic theory and the two in small samples.05.0...g.420 4.032 n = 400 Multiplier method Bootstrap method .05 Multiplier method Bootstrap method a = 0. The chosen values of the param easy to simulate.097 . ?1. This can be seen as mimicking the behaviour of claim or duration data.066 .945. The values shown in Table 1 indicate methods used to simulate the P-values that the test tends to over-reject with the multiplier method and under-reject with the bootstrap method.10.023 .17. American insurance claims.0. intrinsically In Table 1. but with a rejection rate converging in both cases to the chosen nominal significance level as n increases. The reason is that both procedures rely keep the same seeds in the pseudo-random on ranks. tion relies on a Gaussian product kernel and the quick standard rule of thumb (Silverman 1986) to select the two individual bandwidths.395} as exhibited by the corresponding true values of the Kendall tau.05.46} for the Gaussian for the FGM copula.004 .

The data consist in n = 1466 uncensored observed values of the pair (LOSS. (bootstrap method) for Sn .21 r = -0. -0.753 FGM .998 G 1.000 1.000 previous Monte Carlo experiments.974 FGM . Frees & Valdez (1998) choose the Pareto distribution tomodel themargins. Life expectancies at birth. and select Gumbel and Frank copulas (on the basis of a graphical procedure suitable for Archimedean copulas).999 . the bivariate kernel and the bandwidths are again chosen as in the previous Monte Carlo experiments.31 a = 0.953 .208 G . we rely on a nonparametric approach to assess PQD.000 F 1.2.000 = 1.994 FGM 1. These data are available at http://www.177 FGM .gov/cia/publications/factbook/.495 F .000 .202 F .431 . for example. at birth of This second empirical illustration aims to detect a PQD behaviour in life expectancies males and females across 225 different countries.ALAE).2005 Here ALAE POSITIVE DEPENDENCE QUADRANT 421 are a type of insurance company expense that is specifically attributable to the set tlement of individual claims. for the computation of reinsurance premi ums (where the sharing of expenses between the ceding company and the reinsurer has to be decided on) and for the determination of the expense level for a given loss level (for reserving an appropriate amount to cover future settlement expenses).680 . -0. The joint in parametric settings of those two variables has already been examined by Frees & modeling Valdez (1998).993 . the number of replications. the number of replications.686 .000 G .000 1.896 . Again.0356.973 . A slightly different type of data (life expectancy on total population versus difference be tween life expectancy of males and females) has been examined in Amblard & Girard (2005) in the context of semiparametric estimation of bivariate copulas under a PQD assumption.000 FGM 1.05 Multiplier method Bootstrap method a = 0.979 F .421 G . Both models express PQD by their estimated parameter values. The grid.430 .493 FGM . Klugman & Parsa (1999) opt for the Inverse Paralogistic for the losses and for the Inverse Burr for ALAEs. the estimated value of the dependence parameter entails PQD for losses and ALAEs.01 Multiplier method Bootstrap method F .628 .000 = which means that we cannot reject PQD. and Klugman & Parsa (1999). The grid. They use the Frank copula.851 F 1.924 G .000 .Gaussian (G) and Farlie-Gumbel-Morgenstern alternatives. (FGM) copula T= -0. and comprise general liability claims randomly choosen from late settlement lags. Table 2: Power of the test for Frank (F). (bootstrap method) for Sn 5. In the following. We find pn = 1.odci.963 . We find pn (multiplier method) and pn = which means that we cannot reject PQD.991 .345 .000 (multiplier method) and p* = 1.0208.863 G . such as lawyers' fees and claims investigation expenses.998 The data used in these empirical studies were collected by the US Insurance Services Of fice. the bivariate kernel and the bandwidths are chosen as in the = 1.000 1.11 r = -0. This assessment has many implications in insurance. We refer to Denuit & Scaillet (2004) for further discussion and the practical implications on the design of reinsurance treaties when PQD is present.

uh 1. A d-dimensional random vector Y is said to be positively lower orthant dependent (PLOD. .ud)-Cn(ui. then Y is said to be positively orthant dependent (POD... Note also that there is no obvious ranking across these sorts of tests.ud the aforementioned results remain valid with {C?(uu.. ud) ]T Ci(ui. we have considered a Kolmogorov-Smirnov type test for PQD.. Let us further remark that it is rather straightforward to extend the procedure to accommodate dimensions Newman see........ .422 SCAILLET Vol.. This test is con sistent since it is based on an examination of the complete set of restrictions that result from the copula representation of PQD. 2 + C(l ? i?2. udAud)-C(uu for each ui.u'de Sn = [0.. (4) upper orthant dependent d = JJ(l-ti<). (4) and (5) are no more equivalent when d > 3. in short) if d C(tii.... l]d. but other are available. We could also design tests based on Cram?r-von Mises type functionals... test has been designed in the spirit of a Kolmogorov-Smirnov functional. (5) where C denotes the survival copula associated with C. This will lead to tests for positive orthant dependences.ti<f)>C?(tii. l]2 with covariance = ud)} C{u!Au[... 3 6... v). CONCLUDING REMARKSAND EXTENSIONS In this paper.l]d... 1 u2.... ?=i Indeed we have that ud)Bc{l.1-^3) wi.. u'd). the empirical copula process converges weakly to d Gc(tii. 1)..-.. we have C (ui. However the bootstrap procedure would then need to be modified tomake it consistent.. The results of this paper carry over in that case. where Be E{Bc(ui. When (4) and (5) simultaneously hold.. 1.ud)}. For example..ud)e in short) if [0. bridge on [0... . 33..1].... such as i max{0.ud)> while it is said to be positively CXK .. v)}rw(u.1 ix2)+ C(l -1*1..^)G[0..l] The PUOD case is more delicate When d = 3...ttd) V(ui.. Dn(u.u2... e.-. we could opt for a weighted supremum test statistic possibilities y/? swp(Dnw) for some nonnegative weighting function w(u.. The extension of the testing procedure in the PLOD case is immediate... see Nelsen (1999). Two empirical examples have illustrated its practical use in detecting PQD in American insurance claim data and life expectancy data. 2=1 (PUOD.. 1 -1?3) C(l ?xi.... Hence y/? sup ui.U2>) = ui +u2 + C(l - to handle.. ud)C(u[. Of course.... C(tii... ud) = Y[uh t=i V(tii. see Nikitin (1995).. (1984) as described in the next lines.1 u3) 4-U3 ..g. No. [0. higher than two... function ... is a tight Brownian ud)BcK. v) du dv Jo Jo fl for some positive r. as which func The tional yields asymptotic efficiency depends on the alternative being tested. Ud) = Bc(wi. in short).. One could rely on the link between C and C...

ud) Cn(uu. Under continuous differentiability ofthe copula function..ud6[0.l] {C (uiJ. function is a tight Brownian bridge on [0..v) u. 1) C2(u. APPENDIX All A. v) = 5 > 0.v supy/n{Dn(u. covariance function.v) u. upon computation A.v Hence ofS. This means that the properties of a testing procedure based on Sn = y/? should be similar.. l]2.t.l. l]2. v) and Theorem 4 of Fermanian.. + sup y/nD(u. ? - v) < 0 for all sup \fn{Dn(u. for which D(U. v)Bc(u\ v')} =C(u Aw>A v') - C(u.ud)}. .. u'yv. The weak convergence ofthe empirical copula process can then be invoked again to lead to the conclusion that y/? (C n ? C ) has a Gaussian limit..2005 POSITIVE QUADRANT DEPENDENCE 423 (see Georges et al. Proof of Proposition 1..v) u.Ud)} Finally to derive a test for POD. one could rely on a test statistic equal to the supremum of the bivariate vector made of V?{C-L(uu. l]2) to Wegkamp wards Gc(u. then there is some (u. v)Bc(u. for each i?.)] cifn(t*.. Radulovic & Wegkamp (2004).. where Be v) = Bc(u. v).. v) G [0.. Theorem 4 of Fermanian. 2001 for a translation formula in the general case) and build the estimator C n obtained from substituting Cn for C. v) = 4= ? - fi x [l{Fn(Xi) < u. Then the result follows using the inequality Sn > y/?Dn(?. v)C(u\ v').v)}. v' G [0. v) Ci(u.1]. Proof of Lemma 1.ud)-Cn(ui. v) G [0. say (?. sup ui. we obtain Sn < < of Sn and the fact that under Ho.. A3.ui Wn r~? V 2=1 . the results follows from the weak convergence of y/? (Dn ? D) and the definition Proof of Part (ii): If the alternative is true.ud)-Cn(uu. .. t. ofthe Proof of Part (i): From the definitions (tz. This yields the result. limits are taken as n -> co..)4= Y... Proof of Proposition!. u\ v') E{Bc(u.. Gn(Yi) <v}x [HFn(Xi) <u}-u] C(ti... v.1Ud)} and parallel the previous developments. v)Bc(h v). Let us write GCn (u. and y/n{C (uu .v)} D(u.2. Radulovic & (2004) states that the empirical copula process converges weakly in ???([0.v D(u.. D(u. l]2 with covariance = tt(u.

v)}^=? v) where 0 is the zero function. {Cn(u.i] IV \Cn(u. v)}?= ? * so that Ui =*0.v)}?Y.v [0.1 of van der Vaart & Welltier (1996) that the asymptotic equicontinuity conditions for the three empirical and multiplier processes are equivalent. respectively. l]2).424 SCAILLET Vol.9. almost surely where BLi is the set of bounded Lipschitz functions on ???([0. i=l But this holds for almost all samples.v)-C(u. /?OO = ||I7||2.i]l r = Pu\ U* sup [o. 3 c2>n(u. n i i i i -C(u. .v)^= ??? x [l{Gn(Yi) <v}- v].1).v)-C(u.9. amember Ui4 0.v)-C(u. C(u.7 of van der Vaart & Wellner (1996). sup \Cn(u. 33. we have that condi Then since the Ui are independent tional on the sample and identically distributed N(0. No.v)\ f I1 n ~[ -7= y^Ui\>e\ I 1 I J < Consequently 1{ -+0.i / we deduce y/P(\U\ >x)dx < oo and E max JO l<i<n \Ui\/y/? -> 0. This means that the sum of first three terms converge weakly to an independent copy G'c of Gc. from the multiplier inequalities of Lemma 2.Ui. l]2)) which {Cn(u. this leads to the almost sure conditional convergence sup heBLi \Euh(Ccn) - Eh(Gc)\ -? 0. - 1 C(u. Radulovic has the property that sample & (2004) implies that almost every observed sup u. First Wegkamp consider the last term.v)\}2E(?jru2) for this sample we have that {Cn(u.v)}-j= Y^Ui\>e\ J Pu\ sup \{Cn(u.v) Vn ~7 I Uue[o. v) For the first term since ||I7||i E\U\ < oo. = v)} ?j= n 22 Ui => 0> almost surely.l] v) \Cn(% - C(% v)\?> 0. implies that of the space ???([0. Note that Theorem 4 of Fermanian. {Cn(ii. v) - C{u.As in Theorem 2.

W. Valdez for kindly providing the Loss-ALAE data. Since Gfc is a Gaussian process indexed by two parameters within the compact set [0. and c(a) being a continuity point of the distribution of S. proof to get the final result.). Then lim Pf?ect H0\H0) n-?oo = = < lim P{Sn > cn(a)} lim P{Sn > c(a)} + = a. P{S > c(a)} n?*oo n?>oo lim [P{5n > cn(a)} .4.l] Gcn (uiv) =^ u. let Pn (t) be the cumulative distribution function of the process (conditional on the original sample) generated by sup Gcn The Continuous Mapping Theorem yields sup u. which were collected by the US Insurance Services Office (ISO).l] G>c(u> v)> almost surely (6) where the latter random variable is an independent copy of S. instead of almost sure convergence. as well as Professors E. ACKNOWLEDGEMENTS The author would like to thank Christian Genest and two referees for constructive criticism. A. The author acknowledges financial support by the Swiss National Science Foundation through theNational Center of Competence: Financial Valuation andRisk Management (NCCR FINRISK). and we can continue as in the of Proposition 2 but using convergence in probability. Theorem 6 of Fermanian. where Ex.2005 POSITIVE QUADRANT DEPENDENCE425 Now to show the result concerning the asymptotic behaviour of the P-values. Proof of Proposition 3. almost surely (7) by (6) and the aforementioned n-+oo properties of P?. On the other hand. He is further grateful to the participants at seminars atULB and CERN as well as at theDeMoSTAFI conference for their comments. Note that the median of the dis tribution P?(t) of sup G^ is strictly positive and finite. Hence we deduce from the Continuous S?=^supG?(ti. A.P{5n > c(a)}] where the last statement comes from (7).? [0. P? is absolutely continuous (Tsirel'son 1975). Frees and E. Part of his research was done when he was visiting THEMA and ECARES. while c(a) defined by P{S > c(a)} = a is finite and positive for any a < 1/2 (Proposi tion A. where the latter random variable is an independent copy of 5. ? Radulovic & Wegkamp (2004) states that \fn (C* Cn) converges weakly to an independent on the sample in that copy Gq of Gc in probability conditionally heBL! sup \Ex. Part (i) of Proposition 1.y is the expectation Theorem that Mapping given the original sample.v SUP [0. . The author also wishes to express his deep gratitude to Bru no R?millard for his valuable help in designing the simulation-based multiplier method in an appropriate way.1]2.7 of van der Vaart & Wellner (1996). The event {pn < a} is equivalent to the event {Sn > cn(a)} where inf{? : Pn(t) > 1 ? a} = cn(oi) -? c(a).t. Let C* be the empirical copula associated to the bootstrap sample. Part (ii) of Proposition 1 and c(a) < oo ensure that n-?oo lim P(rejecfHol#i) = 1.2.Yh(GCn) Eh(Gc)\^ 0.

Frank's family of bivariate distributions. & S. Glidden (1999). Cr?dit Lyonnais. C. J. A. 64. Cambridge University J. Wiley.). Goovaerts (1996). Forcina (1999).-G. North American Actuarial Journal. Correlation and dependency in risk management: Prop erties and pitfalls. P. 549-555. France. of Multivariate Analysis. 1137? 1153. Correlation and Dependence. Bulletin de l'Acad?mie royale de Belgique. Statistical inference for stochastic dominance and for themeasurement of poverty and inequality. Guay & O. DMV Seminar 9. 141-156. Basel. GRO Working Paper. E. is not identified under the null hypothesis.Wegkamp (2004). New York. 1-25.413-430. ed. 422-450. 68.1985). 86. An asymptotic decomposition formultivariate distribution-free tests of independence. 28. Indirect inference. Anderson G. Scaillet (2004). Multivariate Survival Modelling: A Unified Approach with Copulas. curve orderings. and Graphical tests Statistics. Roncalli (2001). L. Scaillet Journal of Business B. Fitting bivariate loss distributions with copulas. Some concepts of dependence.-D. D. & O. New York. Press. The Annals of Mathematical Statistics. Springer. September P.122-132. Georges. Lehmann (1966). Statistics. Birkh?user. Weak convergence of empirical copula processes. Klugman & R. Paris. D. Nonparametric tests for positive quadrant dependence. G?nssler & W. Econometrica. Hansen metrica. Lamy.1435-1464.139-148. P. pp. A. Classe des sciences.1183-1193. S.426 REFERENCES SCAILLET Vol. and Economic when Inference a nuisance parameter 64. The Econometrics Journal.305-308. Lehmann (1986). Econometrica. Journal ofFinancial Econometrics. Dependency of risks and stop-loss order. Estimation procedures for a semiparametric family of bivariate copulas. D. Econo parameter and threshold moving average models. Denuit. Nicolas. . Chapman & Hall. (1996). Concepts of stochastic dependence in reliability analysis. Bernoulli. 2. Bootstrap tests for distributional treatment effects in instrumental variable models. Ribas (2001). dominance. No. P. Parsa (1999).71-104. Davidson & J. D?sseldorf 8-13. W. Radulovic 10. Fermanian.284-292. Joe (1997). Checking the adequacy of the gamma frailty model for multivariate failure times. (2003). Dardanoni 49-75. Kotz (2001). Jour nal of theAmerican Statistical Association. 1A. M.Dhaene & M. for stochastic 14. ed. Genest (1987). C. Scaillet (2003). Barrett V. 11. Nonparametric (2003). M. Xie (2003). Imperial College Press. D. H. G. 24. Insurance: Mathematics and Economics. 25-54. Biometrika. L. Risk Management: Value at Risk and Beyond (M. Amblard & S. London.Deheuvels Journal (1981). Multivariate Models and Dependence Concepts. Frees & E. 65.847-860. Nonparametric estimation of copulas for time series.!1. Denuit & O. A.274-292. Drouet-Mari & S.!. Fermanian 5. Abadie (2002). 33. Inference R.1-15. 26. & M. Consistent tests for stochastic for Lorenz dominance. Understanding relationships using copulae. Dempster. Deheuvels (1979). La fonction de d?pendance empirique et ses propri?t?s: Un test non param?trique d'ind?pendance. Journal G. Seminaron Empirical Processes (Schlo?Mickein. Biometrika. 2. Duelos (2000). of Computational (1996). 97. H. Embrechts. InHandbook of Reliability Engineering (H.-D. London. Testing Statistical Hypotheses.-Y. Valdez (1998). J. E. 3 A. Lai & M. Journal of Risk. J. Does positive dependence between individual risks increase stop-loss premiums? Insurance: Mathematics and Economics.201-212. E.A. McNeil & D. 381-393. Straumann (2000). Donald & A.). nuisance 21. Dhaene & C. Econometrica. A. Pham. 37. C.102-113. Quibel & T. Girard (2005). Second edition.ASTIN Bulletin. Stute (1987). J. P. E.

pp. C. New York.735-765.cn HEC Gen?ve and FAME. Levy (1992). 12. IMS Lecture Notes-Monographs Series. Density Estimation for Statistics and Data Analysis. Springer. Received Accepted 24 January 26 May 2005 2004 Olivier SCAILLET: scaillet? hec.847-856. Fonctions de r?partition ? n dimensions et leursmarges. eds. Nikitin (1995). Stochastic Orders and theirApplications. The density of the distribution of themaximum of a Gaussian process. 555-593. ed. The Annals of Probability. Seo. 72. Maasoumi schemes. In Studies in theEconomics of Uncertainty (T. B. Asymptotic Efficiency ofNonparametric Tests. New York. Testing for stochastic dominance.2005 POSITIVE DEPENDENCE QUADRANT 427 H. V. Asymptotic independence and limit theorems for positively and negatively depen dent random variables.). In Inequalities in Statistics and Probability (Y. of Economic for stochastic dominance under general sampling D.-J. M. A. E. Academic Press. Theory of Probability and theirApplications. R. Springer. Lecture Notes in Statistics 139. An Introduction toCopulas. Cambridge University Press. Suisse . Y. 127-140. M. B.unige. London.). A. Review & Y. testing Studies. Silverman (1986). Tong. 361-379. L. Whang Consistent (2005). Universit? de Gen?ve Boulevard Carl-Vogt. 38. Linton. McFadden (1989). Shaked & G J. 113-134. 8. O. van der Vaart & J. Stochastic dominance and expected utility: Survey and analysis. L. Chapman & Hall. vol. Fomby & T. The oscillation behavior of empirical processes: themultivariate case. 5. Stute (1984). New York. Springer. Nelsen (1999). Wellner (1996). W. Newman (1984). Shanthikumar (1994). Sklar (1959). 229-231. 16. pp. Weak Convergence and Empirical Processes. New York. 102 CH-1211 Gen?ve 4. Tsirel'son (1975). Management Science. Publications de l'Institut de statistique de l'Universit? de Paris. A.

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd