Metodoloski zvezki, Vol. 13, No. 2, 2016, 101-116 Estimating the Coefficient of Asymptotic Tail Independence: a Comparison of Methods Marta Ferreira1 Abstract Many multivariate analyses require the account of extreme events. Correlation is an insufficient measure to quantify tail dependence. The most common tail dependence coefficients are based on the probability of simultaneous exceedances. The coefficient of asymptotic tail independence introduced in Ledford and Tawn ([18] 1996) is a bivariate measure often used in the tail modeling of data in finance, environment, insurance, among other fields of applications. It can be estimated as the tail index of the minimum component of a random pair with transformed unit Pareto marginals. The literature regarding the estimation of the tail index is extensive. Semi-parametric inference requires the choice of the number k of the largest order statistics that lead to the best estimate, where there is a tricky trade-off between variance and bias. Many methodologies have been developed to undertake this choice, most of them applied to the Hill estimator (Hill, [16] 1975). We are going to analyze, through simulation, some of these methods within the estimation of the coefficient of asymptotic tail independence. We also compare with a minimum-variance reduced-bias Hill estimator presented in Caeiro et al. ([3] 2005). A pure heuristic procedure adapted from Frahm et al. ([13] 2005), used in a different context but with a resembling framework, will also be implemented. We will see that some of these simple tools should not be discarded in this context. Our study will be complemented by applications to real datasets. 1 Introduction It is undeniable that extreme events have been occurring in areas like environment (e.g. climate changes due to pollution and global heating), finance (e.g., market crashes due to less regulation and globalization), telecommunications (e.g., growing traffic due to a high expanding technological development), among others. Extreme values are therefore the subject of concern of many analysts and researchers, who have come to realize that they should be dealt with some care, requiring their own treatment. For instance, the classical linear correlation is not a suitable dependence measure if the dependence characteristics in the tail differ from the remaining realizations in the sample. An illustration is addressed in Embrechts et al. ([9] 2002). To this end, the tail dependence coefficient (TDC) defined in 1 Center of Mathematics of University of Minho, Center for Computational and Stochastic Mathematics of University of Lisbon and Center of Statistics and Applications of University of Lisbon, Portugal; msferreira@math.uminho.pt 102 Marta Ferreira Joe ([17] 1997), usually denoted by A, is more appropriate. More precisely, for a random pair (X, Y) with respective marginal distribution functions (dfs) FX and FY, the TDC is given by A = lim P(Fy(Y) > 1 - t|Fx(X) > 1 - t), (1.1) whenever the limit exists. Roughly speaking, the TDC evaluates the probability of one variable exceeding a large value given that the other also exceeds it. A positive TDC means that X and Y are tail dependent and whenever null we conclude the random pair is tail independent. In this latter case, the rate of convergence towards zero is a kind of residual tail dependence that, once ignored, may lead to an under-estimation of the risk underlying the simultaneous exceedance of a large value. On the other hand, by considering that the random variables (rv's) X and Y are tail dependent when they are actually asymptotically independent, it will result in an over-estimation of such risk. The degree of misspecification depends on the degree of asymptotic independence given by the mentioned rate of convergence, denoted n in Ledford and Tawn ([18] 1996). More precisely, it is assumed that P (Fx (X) > 1 - t, Fy (Y) > 1 - t) = t1/n L(t), n e (0,1], (1.2) where L(t) is a slowly varying function at zero, i.e., L(tx)/L(t) ^ 1 as 11 0 for all x > 0. We call the parameter n the coefficient of asymptotic tail independence. Whenever n < 1, X and Y are asymptotically independent and, if n = 1, asymptotic dependence holds if L(t) ^ c > 0, as 11 0. In case X and Y are exactly independent then n = 1/2 and we can also discern between asymptotically vanishing negative dependence and asymptotically vanishing positive dependence if, respectively, n e (0,1/2) and n e (1/2,1). Observe that we can state (1.2) as P (mi" (i-FXiX) ■ r-TiW)) > 0 =t-1/n L(1/t)' (13) and thus n corresponds to the tail index of the minimum of the two marginals standardized as unit Pareto. The tail index, also denoted extreme value index, quantifies the "weight" of the tail of a univariate distribution: whenever negative, null or positive it means that the tail of the underlying model is, respectively, "light", "exponential" or "heavy". In what concerns univariate extreme values, it is the primary parameter as it is implicated in all other extremal parameters, such as, extremal quantiles, right end-point of distributions, probability of exceedance of large levels, as well as return periods, among others. Therefore, the estimation of the tail index is a crucial issue, with numerous contributions in the literature. A survey on this topic can be seen, for instance, in Beirlant et al. ([2] 2004). Under a semi-parametric framework in the domain of heavy tails, the Hill estimator, introduced in Hill ([16] 1975), have proved to possess good properties, being an essential tool in any application on this topic. For a random sample (T1,..., Tn), the Hill estimator corresponds to the sample mean of the log-excesses of the k + 1 larger order statistics Tn:n ^ ... ^ Tn-k:n, i.e., 1 k T Hn(k) = H (k) := 1 J>g , 1 < k < n, (1.4) k - -, Tn—k:n i=1 Estimating the Coefficient of Asymptotic Tail Independence 103 Consistency requires that k must be intermediate, that is, a sequence of integers k = kn, 1 < k < n, such that kn —y to and kn/n — 0, as n — to. There is no definite formula to obtain k and it must be chosen not too small to avoid high variance but also not to large to prevent high bias. Figure 1 illustrates this issue, particularly the dashed line corresponding to a unit Frchet model where the tail index is 1. Observe also that there is a kind of stable area of the sample path around the true value of the tail index, where the variance is no longer high and the bias haven't started to increase. This disadvantage is transversal to the semi-parametric tools concerning extreme values inference. In the particular case of the Hill estimator, many efforts have been made to minimize the problem, ranging from bias-corrected versions to the implementation of procedures to compute k. The minimum-variance reduced-bias (MVRB) Hill estimator presented in Caeiro et al. ([3] 2005; see also Neves et al. [21] 2015) was developed for the Hall-Welsh class (within Generalized Pareto distributions), with reciprocal quantile function F-1(1 - 1/x) = CxY (1 + yftxp/p + o(xp)) ,x — to, (1.5) where y > 0 is the tail index of model F, C > 0, and ft = 0 and p < 0 are second order parameters. The MVRB Hill estimator is given by CHn(k) = CH(k) := H(k) - ) , 1 < k < n, (1.6) where ft and p3 are suitable estimators of ft and p, respectively. Details about these latter are addressed in Caeiro et al. ([4] 2009) and references therein. We will denote it "corrected Hill" (CH). Our aim is to compare, through simulation, several methods regarding the Hill and corrected Hill estimators applied to the estimation of n. We also consider the graphical and pure heuristic procedure presented in Frahm et al. ([13] 2005) in the context of estimating the TDC A in (1.1), also relying on the choice of k upper order statistics with the same bias/variance controversy. All the estimation procedures are described in Section 2. The simulation study is conducted in Section 3 and applications to real datasets appear in Section 4. A small discussion ends this work in Section 5. 2 Estimation methods In this section we describe the procedures that we are going to consider in the estimation of the coefficient of asymptotic tail independence n given in (1.3) and therefore corresponding to the tail index of T = min((1 - Fx(X))-1, (1 - Fy(Y))-1). (2.1) Coefficient n is positive and we can use positive tail index estimators such as Hill. Observe that T is the minimum between two unit Pareto r.v.'s Alternatively, we can also undertake 104 Marta Ferreira Figure 1: Hill plots of 1000 realizations of a unit Pareto (full line) and a unit Frechet (dashed line), both with tail index equal to 1 (horizontal line). a unit Frchet marginal transformation since 1 — FX (X) ~ — log FX (X). However, in the sequel, we prosecute with unit Pareto marginals, since the Hill estimator has smaller bias in the Pareto models than in the Frchet ones (see Figure 1; see also Draisma et al. [6] 2004 and references therein). In order to estimate the unknown marginal df's FX and FY we consider their empirical counterparts (ranks of the components), more precisely, Ti(n) := min((n + 1)/(n + 1 — RX), (n + 1)/(n + 1 — RY)), i = 1,..., n where RX denotes the rank of Xi among (Xi,..., Xn) and RY denotes the rank of Yi among (Yi,...,Yn). The estimation of n through the tail index estimators Hill and maximum likelihood (Smith, [24] 1987) was addressed in Draisma et al. ([6] 2004). Other estimators were also considered in Poon et al. ([23] 2003; see also references therein) and more recently in Goegebeur and Guillou ([14] 2013) and Dutang et al. ([8] 2014). However, no method was analyzed in order to attain the best choice of k in estimation. In the domain of positive tail indexes, the Hill estimator is the most widely studied and many developments have been appearing around it. The main topics concern methods to obtain the value of k related to the number of tail observations to use in estimation and procedures to control the bias without increasing the variance. The corrected Hill version in (1.6), for instance, removes from Hill its dominant bias component estimated by H(k)(/3(n/k)p)/(1 — J5). In the following, we describe the methods developed in literature for the Hill estimator to compute the value of k, that will be used to estimate n (the tail index of rv T in (2.1)) in our simulation study. Based on Beirlant et al. ([1] 2002) and little restrictive conditions on the underlying Estimating the Coefficient of Asymptotic Tail Independence 105 model, we have T(n) H(i) / i \ -p Y :=(i + l)log T (n) — + = n + b(n/k)( -k) + ei,i = (2.2) n-(i+1):n (i + ) where the error term ei is zero-centered and b is a positive function such that b(x) — 0, as x —y . Extensive simulation studies conclude that the results tend to be better when p is considered fixed, even if misspecified. Matthys and Beirlant ([19] 2000) suggest p = -1. From model (2.2), the resulting least squares estimators of n and b(n/k) are given by C = Yk - C/(1 - p) and = (1-p)p21-2p)k E?=1 ((k)-P - 1-) Yi. (2.3) Thus, by replacing these estimates in the Hill's asymptotic mean squared error (AMSE) 2 i-p ' amse(h (-)) = + ( ^np1)2 we are able to compute klpt as the value of k that minimizes the obtained estimates of the AMSE and estimate n as H(k1pt). On the other hand, we can compute the approximate value of k that minimizes the AMSE, given by -opt - b(n/k)-2/(1-2p)k-2p/(1-2p^1/(1-2p). (2.4) See, e.g., Beirlant et al. ([1] 2002). Replacing again n and b(n/k) by the respective least squares estimates in (2.3) with fixed p = -1, we derive koptjk, for k = 3, ...,n, using (2.4). Then compute -2pt = median{koptjk, k = 3,..., |_fJ}, where [xj denotes the largest integer not exceeding x and consider n estimated by H(k2pt). Further reading of the methods is referred to Beirlant et al. ([1] 2002), Matthys and Beirlant ([19] 2000) and references therein. In the sequel, they are shortly denoted, respectively, AMSE and KOPT. The adaptive procedure of Drees and Kaufmann ([6] 1998) looks for the optimum k under which the bias starts to dominate the variance. The method is developed for the Hall-Welsh class of models defined in (1.5), for which it is proved that the maximum random fluctuation of \/i(H(i) - n), i = 1,k - 1, with k = kn an intermediate sequence, is of order \Jlog log n. More precisely, for p fixed at -1, we have: 1. Consider rn = 2.5 x j x n0 25, with j = k2^n,n. 2. Calculate j(rn) := min{k = 1,...,n - 1 : maxi rn}. If \/i|H(i) - H(k)| > rn doesn't hold for any k, consider 0.9 x rn to rn and repeat step 2, otherwise move to step 3. 3. For e 6 (0,1), usually e = 0.7, obtain 1/(1-0 1 . . „ / KrE ) \ k = kDK — l(2jj2)1/31 3 Ur(rn))£ 106 Marta Ferreira This method will be shortly referred DK. Sousa and Michailidis (2004) method is based on the Hill sum plot, (k, Sk), k = 1,...,n — 1, where Sk = kH(k). We have E(Sk) = kn, an thus the sumplot must be approximately linear for the values of k where H(k) k n, with the respective slope being an estimator of n. The method essentially seeks the breakdown of linearity. Their approach is based on a sequential testing procedure implemented in McGee and Carleton ([20] 1970), leaning over approximately Pareto tail models. More precisely, consider the regression model y = Xn + 5, with y = (Si,..., Sk)', X = [1 i]k=1 and 5 the error term. It is checked the null hypothesis that a new point y0 is adjacent to the left or to the right of the set of points y = (y1,..., yk), through the statistics k TS = s-21 (yo - £0)2 + - yn2 i= 1 where * denotes the predictions based on k + 1 and s2 = (k — 2)-1(y/y — r/X'y). Since TS is approximately distributed by F1k-2, the null hypothesis is rejected if TS is larger than the (1 — a)-quantile, F1jk_2;1_Q. The method, shortly denoted SP from now on, is described in the following algorithm: 1. Fit a least-squares regression line to the initial k = vn upper observations, y = [yi]k=1 (usually v = 0.02). 2. Using the test statistic TS, determine if a new point y0 = yj for j > k, belongs to the original set of points y. Go adding points until the null hypothesis is rejected. 3. Consider knew = max(0, {j : TS < F1jk_2;1_Q}). If k^w = 0, no new points are added to y and thus move forward to step 4. Return to step 1. if knew > 0 by considering k = knew. 4. Estimate n by considering the obtained k. The heuristic procedure introduced in Gomes et al. ([15] 2013), searches for the supposed stable region encompassing the best k to be estimated. More precisely, we need first to obtain the minimum value j0, such that the rounded values to j decimal places of H(k), 1 < k < n, denoted H(k; j) are not all equal. Identify the set of values of k associated to equal consecutive values of H(k; j0). Consider the set with largest range ^ := kmax — kmin. Take all the estimates H(k; j0 + 2) with kmax < k < kmin, i.e., the estimates with two additional decimal points and calculate the mode. Consider K the set of k-values corresponding to the mode. Take H(k), with k being the maximum of K. Since it was specially designed for reduced-bias estimators, we shortly referred it as RB method hereinafter. Frahm et al. ([13] 2005) also presented a heuristic procedure that can be applied to all estimators depending on a number k of rv's whose choice bears the mentioned trade-off between bias and variance. Indeed is was developed within the estimation of the TDC A defined in (1.1). It was adapted to the Hill estimator in Ferreira ([11, 12] 2014, 2015) as follows: Estimating the Coefficient of Asymptotic Tail Independence 107 1. Smooth the Hill plot (k, H(k)) by taking the means of 2b + 1 successive points, H(1),...,H(n - 2b), with bandwidth b = [w x nJ. 2. Define the regions pk = (H(k),..., H(k + m — 1)), k =1,..., n — 2b — m + 1, with length m = [Vn — 2bJ. The algorithm stops at the first region satisfying k+m-1 Y^ |H(i) — H(k) | < 2s, i=k+1 where s is the empirical standard-deviation of H(1),..., H(n — 2b). 3. Consider the chosen plateau region pk * and estimate n as the mean of the values of pk* (consider the estimate zero if no plane region fulfills the stopping condition). The estimation of n through the plateau method was analyzed in Ferreira and Silva ([10] 2014) with respect to the sensibility of the bandwidth. The value w = 0.005 seems a reasonable choice (thus each moving average in step 1. consists in 1% of the data), also suggested in Frahm et al. ([13] 2005). In the sequel it will be referred as plateau method (in short PLAT). Both RB and PLAT are simultaneously graphical and free-assumption methods since they are based on the search of a plane region of the estimator's plot that presumably contains the best sample fraction k to be estimated through a totally "ad-hoc" procedure. The sumplot is also a graphical method and the remaining procedures are neither graphical nor free-assumption. 3 Simulation study In this section we compare through simulation the performance of the methods described above within the estimation of n through the under study estimators Hill in (1.4) and corrected Hill in (1.6). We have generated 100 runs of samples of sizes n = 100,1000, 5000 from the following models: • Bivariate Normal distribution (n = (1 + p)/2; see, e.g., Draisma et al. [6] 2004); we consider correlation p = —0.2 (n = 0.4), p = 0.2 (n = 0.6) and p = 0.8 (n = 0.9); we use notation, respectively, N(—0.2), N(0.2) and N(0.8). • Bivariate t-Student distribution tv with correlation coefficient given by p = — 1 (A = 2Fiv+1 (—V(v + 1)(1 — p)/(1 + p)), see Embrechts et al. [9] 2002; we have A > 0 and thus n = 1); we consider v = 4 and p = 0.25 (A = 0.1438) and v = 1 and p = 0.75 (A = 0.6464); we use notation, respectively, t4 and t1. • Bivariate extreme value distribution with a asymmetric-logistic dependence function = (1 — a1)x + (1 — a2)y + ((a1x)1/a + (a2y)1/a)a, with > 0, 108 Marta Ferreira dependence parameter a G (0,1] and asymmetric parameters a^ a2 G (0,1] (A = 2 —1(1,1), see Beirlant et al. [1] 2004; we have A > 0 and thus n = 1); we consider a = 0.7 and ai = 0.4, a2 = 0.2 (A = 0.1010) and a = 0.3 and ai = 0.6, a2 = 0.8 (A = 0.5182); we use notation, respectively, AL(0.7) and AL(0.3). • Farlie-Gumbel-Morgenstern distribution with dependence 0.5 (n = 0.5, see Dutang et al. [8] 2014); we use notation FGM(0.5). • Frank distribution with dependence 2 (n = 0.5, see Dutang et al. [8] 2014); we use notation Fr(2). Observe that the case N(0.8) is an asymptotic tail independent model close to tail dependence since n = 0.9 k 1. On the other hand, the cases t4 and AL(0.7) are tail dependent cases (n = 1) near asymptotic tail independence since A = 0.1438 k 0 and A = 0.1010 k 0, respectively. We consider these examples in order to assess the robustness of the methods in border cases. In Figures 2 and 3 are plotted, respectively, the results of the simulated values of the absolute bias and root mean squared error (rmse), for the Hill and corrected Hill estimators, in the case n = 1000. All the results are presented in Table 1 concerning the Hill estimator and Table 2 with respect to the corrected Hill. Observe that this latter case requires the estimation of additional second order parameters (^ and p). To this end, we have followed the indications in Caeiro et al. ([4] 2009). For the p estimation, there was an overall best performance whenever it was taken fixed at value —1, leading to the reported results. The largest differences between Hill and corrected Hill can be noticed in the above mentioned border cases, with the corrected one presenting lower absolute bias and rmse. The other models also show this difference but in a small amount. We remark that we are working with the minimum of Pareto rv's and the Hill estimator is unbiased in the Pareto case. The FGM and Frank models behave otherwise with a little lower absolute bias and rmse within the Hill estimator, for either estimated or several fixed values tried for p. The failure cases in the DK method (column "NF" of Tables 1 and 2) correspond to an estimate of k out of the range {1,..., n — 1}, which were ignored in the results. It sets up the worst performance, which may be justified by the fact that the class of models underlying the scope of application of this method excludes the simple Pareto law. The corrected Hill exhibits better results in general, particularly for methods KOPT, PLAT and AMSE, followed by SP and RB, in large sample sizes ^¿=1000). The PLAT procedure also performs well with the Hill estimator unlike the SP. For n = 100, we have good results within RB and SP based on corrected Hill. Once again, the PLAT method behaves well in both estimators. The border cases of weak tail dependence (t4 and AL(0.7)) are critical throughout all evaluated procedures and estimators. On the other hand, the methods are robust in the border case of tail independence near dependence expressed in model N(0.8). 4 Applications In this section we illustrate the methods with three datasets analyzed in literature: Estimating the Coefficient of Asymptotic Tail Independence 109 Figure 2: Simulated results of the absolute bias of Hill (full) and corrected Hill (dashed), for n = 1000, of the models (left-to-right and top-to-down): N(-0.2), N(0.2), N(0.8), tA, ti, AL(0.3), AL(0.7), FGM(0.5) and Fr(2). • I: The data consists of closing stock index levels of S&P 500 from the US and FTSE 100 from the UK, over the period 11 December 1989 to 31 May 2000, totalizing 2733 observed pairs (see, e.g., Poon et al. ([23] 2003)). • II: The wave-surge data corresponding to 2894 paired observations collected during 1971-77 in Cornwall (England); it was analyzed in Coles and Tawn ([5] 1994) and later also in Ramos and Ledford ([22] 2009) under a parametric view. • III: The Loss-ALAE data analyzed in Beirlant et al. ([2] 2004; see also references therein) consisting of 1500 pairs of registered claims (in USD) corresponding to an 110 Marta Ferreira Figure 3: Simulated results of the rmse of Hill (full) and corrected Hill (dashed), for n = 1000, of the models (left-to-right and top-to-down): N(-0.2), N(0.2), N(0.8), ta, ti, al(0.3), AL(0.7), FGM(0.5) and Fr(2). indemnity payment (loss) and an allocated loss adjustment expense (ALAE). The respective scatter-plots are placed in Figure 4. For the US and UK stock market returns, the largest values in each tail for one variable correspond to reasonably large values of the same sign for the other variable, hinting an asymptotic independence but not exactly independence. In the wave-surge data, the dependence seems a bit more persistent within large values, as well as in Loss-ALAE data. The Hill and corrected Hill sample paths of n estimates are pictured in Figure 5. Table 3 reproduces the estimates obtained with each method and estimators under study. The estimation results found in literature for the financial (I), environmental (II) and insurance datasets (III) are respec- Estimating the Coefficient of Asymptotic Tail Independence 111 tively approximated by 0.731, 0.85 and 0.9. The results seem to be in accordance with the simulation study. k k k Figure 5: From left to right: sample paths of Hill (full;black) corrected Hill (dashed;grey) of datasets I, II and III. 5 Discussion In this paper we have analyzed some simple estimation methods for the coefficient of asymptotic tail independence, with some of them revealing promising results. However, the choice of the estimator is not completely straightforward. It can be seen from simulation results that the ordinary Hill estimator may be still preferred over the corrected one in some situations. Also in boundary cases of tail dependence near independence, there are still some worrying errors to correct. These will be topics of a future research. 112 Marta Ferreira Acknowledgment The author wishes to thank the reviewers for their constructive and valuable comments that have improved this work. This research was financed by Portuguese Funds through FCT - Fundacao para a Ciencia e a Tecnologia, within the Project UID/MAT/00013/2013 and by the research centre CEMAT (Instituto Superior Tecnico, Universidade de Lisboa) through the Project UID/Multi/04621/2013. References [1] Beirlant, J., Dierckx, G., Guillou, A. and Starica, C. (2002): On Exponential Representation of Log-Spacings of Extreme Order Statistics. Extremes, 5, 157-180. [2] Beirlant, J., Goegebeur, Y., Segers, J. and Teugels, J.L. (2004): Statistics of Extremes: Theory and Applications. J. Wiley & Sons. [3] Caeiro, F., Gomes, M.I. and Pestana, D.D. (2005): Direct reduction of bias of the classical Hill estimator. Revstat, 3(2), 111-136. [4] Caeiro, F., Gomes, M.I. and Henriques-Rodrigues, L. (2009): Reduced-Bias Tail Index Estimators Under a Third-Order Framework. Communications in Statistics -Theory and Methods, 38(7), 1019-1040. [5] Coles, S.G. and Tawn, J.A. (1994): Statistical methods for multivariate extremes: an application to structural design (with discussion). Appl. Statist., 43, 1-48. [6] Draisma, G., Drees, H., Ferreira, A. and de Haan, L. (2004): Bivariate tail estimation: dependence in asymptotic independence. Bernoulli, 10(2), 251-280. [7] Drees, H. and Kaufmann, E. (1998): Selecting the optimal sample fraction in univariate extreme value estimation. Stochastic Process Appl., 75, 149-172. [8] Dutang, C., Goegebeur, Y. and Guillou, A. (2014): Robust and bias-corrected estimation of the coefficient of tail dependence. Insurance: Mathematics and Economics, 57, 46-57. [9] Embrechts, P., McNeil, A. and Straumann, D. (2002): Correlation and dependency in risk management: properties and pitfalls. In: Risk Management: Value at Risk and Beyond, M.A.H. Dempster, Ed. Cambridge University Press, 176-223. [10] Ferreira, M. and Silva, S. (2014): An Analysis of a Heuristic Procedure to Evaluate Tail (in)dependence. Journal of Probability and Statistics, Vol. 2014, Article ID 913621, 15 pages. [11] Ferreira, M. (2014): A Heuristic Procedure to Estimate the Tail Index. Proceedings of the 14th International Conference in Computational Science and Its Applications -ICCSA 2014, June 30 - July 3 (2014), Guimares, Portugal, IEEE-Computer Society, 4264a241, 241-245. Estimating the Coefficient of Asymptotic Tail Independence 113 [12] Ferreira, M. (2015): Estimating the tail index: Another algorithmic method. ProbStat Forum, 08, 45-53. [13] Frahm, G., Junker, M. and Schmidt R. (2005): Estimating the tail-dependence coefficient: properties and pitfalls. Insurance: Mathematics & Economics, 37(1), 80100. [14] Goegebeur, Y. and Guillou, A. (2013): Asymptotically unbiased estimation of the coefficient of tail dependence. Scand. J. Stat., 40, 174-189 [15] Gomes, M.I., Henriques-Rodrigues, L., Fraga Alves, M.I. and Manjunath, B.G. (2013): Adaptive PORT-MVRB estimation: an empirical comparison of two heuristic algorithms. Journal of Statistical Computation and Simulation, 83(6), 11291144. [16] Hill, B.M. (1975): A Simple General Approach to Inference About the Tail of a Distribution. Ann. Stat., 3, 1163-1174. [17] Joe, H. (1997): Multivariate Models and Dependence Concepts. Harry Joe, Chapman & Hall. [18] Ledford, A. and Tawn, J. (1996): Statistics for near independence in multivariate extreme values. Biometrika, 83(1), 169-187. [19] Matthys, G. and Beirlant, J. (2000): Adaptive Threshold Selection in Tail Index Estimation. In: Extremes and Integrated Risk Management, (Edited by P. Embrechts), 37-49. Risk Books, London. [20] McGee, V.E. and Carleton, W.T. (1970): Piecewise Regression. Journal of the American Statistical Association, 65, 1109-1124. [21] Neves, M., Gomes, M.I., Figueiredo, F. and Prata-Gomes, D. (2015): Modeling Extreme Events: Sample Fraction Adaptive Choice in Parameter Estimation. Journal of Statistical Theory and Practice, 9(1), 184-199. [22] Ramos, A. and Ledford, A. (2009): A new class of models for bivariate joint tails. Journal of the Royal Statistical Society, Series B, 71, 219-241. [23] Poon, S.-H., Rockinger, M. and Tawn, J. (2003): Modelling extreme-value dependence in international stock markets. Statistica Sinica, 13, 929-953. [24] Smith, R.L. (1987): Estimating tails of probability distributions. Ann. Statist., 15, 1174-1207. [25] Sousa, B. and Michailidis, G. (2004): A Diagnostic Plot for Estimating the Tail Index of a Distribution. Journal of Computational and Graphical Statistics, 13(4), 1-22. a •<—I Ë £ SP KOPT AMSE RB DK PLAT n = 100 abias rmse k abias rmse k abias rmse k abias rmse k abias rmse k NF abias rmse N(-0.2) 0.0449 0.0590 90 0.0387 0.1232 12 0.0258 0.0579 68 0.0286 0.0470 69 0.0350 0.2883 3 4 0.0111 0.0780 JV( 0.2) 0.0574 0.0698 89 0.1202 0.2002 15 0.0878 0.1224 64 0.0532 0.0714 75 0.0388 0.4878 4 2 0.0384 0.1042 N( 0.8) 0.1372 0.1460 93 0.1881 0.2726 16 0.1935 0.2402 77 0.1323 0.1397 75 0.1320 0.4158 8 7 0.1133 0.1440 ti 0.4187 0.4223 96 0.4121 0.4458 20 0.4309 0.4362 79 0.4155 0.4188 76 0.3007 0.5849 3 5 0.3539 0.3734 h 0.2266 0.2323 96 0.1605 0.2297 14 0.2318 0.2344 95 0.2144 0.2199 76 0.1923 0.3481 12 5 0.1300 0.1507 AL{ 0.7) 0.4642 0.4658 94 0.4625 0.4895 18 0.4784 0.4863 92 0.4572 0.4594 78 0.3447 0.6026 4 3 0.4199 0.4342 AL( 0.3) 0.2825 0.2855 98 0.1686 0.2364 17 0.2877 0.3024 73 0.2498 0.2556 74 0.1991 0.3459 14 6 0.1585 0.1864 FGM{ 0.5) 0.0383 0.0578 90 0.0507 0.1683 12 0.0163 0.1117 56 0.0362 0.0585 75 0.0508 0.3649 6 8 0.0302 0.1052 Fr( 2) 0.0805 0.0954 88 0.2065 0.1762 13 0.0320 0.1265 61 0.0839 0.0960 77 0.0041 0.3391 5 5 0.0764 0.1293 n = 1000 abias rmse k abias rmse k abias rmse k abias rmse k abias rmse k NF abias rmse JV(—0.2) 0.0425 0.0546 819 0.0059 0.0515 121 0.0378 0.0474 652 0.0437 0.0455 755 0.0242 0.3225 48 2 0.0247 0.0399 N( 0.2) 0.0462 0.0642 826 0.0370 0.0687 171 0.0519 0.0690 111 0.0394 0.0432 754 0.0223 0.3651 39 0 0.0297 0.0452 JV( 0.8) 0.1178 0.1266 866 0.0832 0.0907 277 0.1231 0.1239 920 0.0926 0.0940 625 0.0991 0.3588 84 1 0.0716 0.0784 U 0.3921 0.4013 893 0.3303 0.3339 220 0.3703 0.3737 460 0.4056 0.4061 822 0.0431 0.6092 29 1 0.3114 0.3172 h 0.1975 0.2095 933 0.0777 0.0896 238 0.1530 0.1562 509 0.1886 0.1906 779 0.0479 0.1042 78 0 0.0554 0.0664 AL{ 0.7) 0.4518 0.4544 941 0.3906 0.3931 197 0.4245 0.4270 592 0.4392 0.4398 643 0.1613 0.6207 45 4 0.3827 0.3864 AL{ 0.3) 0.2369 0.2597 885 0.1282 0.1356 303 0.1821 0.1859 496 0.1940 0.1945 580 0.0800 0.1506 108 1 0.0868 0.0961 FGM( 0.5) 0.0358 0.0430 846 0.0303 0.0525 178 0.0429 0.0600 630 0.0487 0.0516 762 0.0216 0.3347 50 0 0.0415 0.0532 Fr(2) 0.0630 0.0859 696 0.0305 0.0791 132 0.0409 0.1136 405 0.0952 0.0963 786 0.0380 0.3451 50 3 0.0691 0.0795 n = 5000 abias rmse k abias rmse k abias rmse k abias rmse k abias rmse k NF abias rmse N(-0.2) 0.0485 0.0515 4369 0.0217 0.0280 629 0.0424 0.0445 3353 0.0399 0.0406 3135 0.0920 0.3383 572 1 0.0214 0.0271 JV(0.2) 0.0486 0.0490 4804 0.0288 0.0346 847 0.0410 0.0422 3684 0.0384 0.0391 3590 0.0601 0.4406 402 1 0.0261 0.0330 JV(0.8) 0.1253 0.1261 4902 0.0725 0.0745 1343 0.1021 0.1043 3357 0.0907 0.0915 3052 0.0696 0.2242 737 0 0.0585 0.0625 U 0.4103 0.4117 4853 0.2709 0.2745 548 0.2746 0.2829 648 0.4106 0.4107 4418 0.0636 0.4472 34 1 0.2653 0.2688 h 0.2075 0.2090 4902 0.0499 0.0543 1062 0.0804 0.0843 1442 0.2039 0.2043 4573 0.0209 0.0393 235 0 0.0201 0.0328 AL{ 0.7) 0.4594 0.4595 4999 0.3428 0.3448 457 0.3558 0.3633 1178 0.4411 0.4413 3222 0.1898 0.5659 20 2 0.3511 0.3534 AL(0.3) 0.2694 0.2712 4950 0.0956 0.0989 969 0.1100 0.1137 1101 0.1989 0.1998 3024 0.0499 0.0641 298 0 0.0529 0.0642 FGM (0.5) 0.0391 0.0422 4562 0.0277 0.0387 705 0.0415 0.0460 2053 0.0487 0.0494 3655 0.0421 0.3120 190 0 0.0313 0.0379 Fr(2) 0.0831 0.0842 4854 0.0620 0.0684 617 0.0862 0.0926 1590 0.1027 0.1030 3650 0.0035 0.2501 286 0 0.0692 0.0738 -xt" Table 1: Simulation results from Hill estimator, where abias denotes the absolute bias, NF the number of fails and k correspond to the mean of the k values obtained in the 100 runs. SP KOPT AMSE RB DK PLAT n = 100 abias rmse k abias rmse k abias rmse k abias rmse k abias rmse k NF abias rmse N(-0.2) 0.0186 0.0653 91 0.0427 0.1287 12 0.0032 0.0738 57 0.0137 0.0603 74 0.0416 0.2880 2 5 0.0076 0.0795 JV( 0.2) 0.0164 0.0977 90 0.1085 0.2044 15 0.0458 0.1295 58 0.0202 0.0961 74 0.0514 0.4604 3 1 0.0241 0.1130 N( 0.8) 0.0594 0.1066 93 0.1717 0.2675 17 0.1014 0.1860 66 0.0658 0.1050 77 0.1025 0.4099 7 6 0.0959 0.1436 ti 0.3446 0.3618 96 0.3846 0.4268 20 0.3649 0.3810 66 0.3566 0.3704 71 0.2871 0.6015 3 2 0.3361 0.3610 k 0.0952 0.1261 96 0.1369 0.2112 15 0.1104 0.1337 78 0.1118 0.1387 75 0.1437 0.3297 5 0 0.0850 0.1215 AL{ 0.7) 0.3995 0.4123 93 0.4528 0.4846 18 0.4245 0.4410 60 0.4122 0.4227 76 0.3313 0.5980 4 4 0.4046 0.4237 AL( 0.3) 0.0437 0.1355 96 0.1187 0.2105 21 0.0781 0.1698 66 0.0609 0.1418 71 0.1537 0.3491 7 3 0.0865 0.1519 FGM{ 0.5) 0.0659 0.1121 89 0.0439 0.1749 13 0.0199 0.1345 55 0.0565 0.1036 72 0.0468 0.3775 3 8 0.0393 0.1170 Fr( 2) 0.1237 0.1549 88 0.0199 0.1794 13 0.0733 0.1718 58 0.1210 0.1482 73 0.0048 0.3401 4 7 0.0912 0.1499 n = 1000 abias rmse k abias rmse k abias rmse k abias rmse k abias rmse k NF abias rmse JV(—0.2) 0.0165 0.0357 819 0.0008 0.0514 120 0.0119 0.0495 463 0.0204 0.0286 948 0.0103 0.3473 9 2 0.0206 0.0367 N( 0.2) 0.0200 0.0539 808 0.0305 0.0662 169 0.0273 0.0608 515 0.0179 0.0342 913 0.0432 0.3660 18 1 0.0222 0.0442 JV( 0.8) 0.0353 0.0552 848 0.0545 0.0674 253 0.0450 0.0505 527 0.0359 0.0438 837 0.1318 0.4158 23 7 0.0514 0.0622 ti 0.3255 0.3343 893 0.3061 0.3109 197 0.3275 0.3317 296 0.3471 0.3489 838 0.0806 0.6097 23 0 0.3042 0.3100 k 0.0514 0.0680 924 0.0278 0.0474 238 0.0525 0.0617 331 0.0667 0.0731 827 0.1303 0.2741 54 2 0.0276 0.0439 AL{ 0.7) 0.3937 0.3969 941 0.3751 0.3786 183 0.3920 0.3948 365 0.4009 0.4023 935 0.1170 0.6324 14 5 0.3781 0.3817 AL{ 0.3) 0.0063 0.0538 857 0.0371 0.0572 210 0.0610 0.1185 241 0.0239 0.0409 797 0.1388 0.2870 42 2 0.0413 0.0559 FGM(O.S) 0.0547 0.0649 846 0.0356 0.0572 180 0.0617 0.0671 600 0.0657 0.0698 904 0.0288 0.3346 42 0 0.0446 0.0585 Fr{ 2) 0.0854 0.1104 668 0.0371 0.0841 140 0.0845 0.1253 516 0.1172 0.1200 814 0.0442 0.3355 62 1 0.0729 0.0843 n = 5000 abias rmse k abias rmse k abias rmse k abias rmse k abias rmse k NF abias rmse JV(—0.2) 0.0199 0.0240 4368 0.0156 0.0248 584 0.0206 0.0254 1823 0.0208 0.0225 4686 0.0520 0.3914 59 0 0.0200 0.0248 JV(0.2) 0.0173 0.0223 4804 0.0223 0.0304 865 0.0193 0.0232 2720 0.0159 0.0205 4766 0.0292 0.4502 71 1 0.0212 0.0289 JV(0.8) 0.0324 0.0346 4902 0.0458 0.0495 1110 0.0481 0.0502 1291 0.0308 0.0337 4466 0.0992 0.4592 78 3 0.0475 0.0521 ti 0.3349 0.3360 4853 0.2495 0.2535 454 0.2549 0.2620 487 0.3447 0.3451 3940 0.1093 0.5322 39 2 0.2666 0.2696 ti 0.0446 0.0473 4901 0.0008 0.0191 747 0.0127 0.0216 979 0.0535 0.0566 3782 0.0849 0.2497 311 2 0.0032 0.0214 AL{ 0.7) 0.3967 0.3971 4999 0.3303 0.3325 397 0.3210 0.3279 465 0.4003 0.4007 4203 0.1226 0.6323 26 3 0.3509 0.3529 AL(0.3) 0.0157 0.0260 4902 0.0315 0.0401 550 0.0414 0.0485 633 0.0194 0.0292 4520 0.1619 0.3958 210 1 0.0355 0.0427 FGM(0.5) 0.0561 0.0602 4562 0.0316 0.0431 727 0.0511 0.0567 2440 0.0619 0.0634 4196 0.0484 0.2880 218 0 0.0328 0.0399 Fr(2) 0.1186 0.1208 4854 0.0693 0.0758 691 0.1225 0.1162 2431 0.1264 0.1270 4100 0.0333 0.2429 241 1 0.0703 0.0757 Table 2: Simulation results from corrected Hill estimator, where abias denotes the absolute bias, NF the number of fails and k correspond to the mean of the k values obtained in the 100 runs. 116 Marta Ferreira H (k) I k II k III k DK D.651D 21 D.8255 83 D.7827 78 SP D.6D25 2592 D.5922 2893 D.6584 1499 KOPT D.6733 744 D.9137 738 D.8444 135 AMSE D.6494 955 D.7D76 1244 D.685D 1172 RB D.6D41 2477 D.5967 2772 D.7428 7D8 PLAT D.7148 - D.8755 - D.811D - CH (k) I k II k III k DK D.7654 5 D.4521 1 D.7D44 27 SP D.6725 2592 D.8581 2893 D.8671 1499 KOPT D.7D7D 585 D.8991 412 D.8661 176 AMSE D.6925 726 D.8997 596 D.8386 678 RB D.6652 2264 D.83DD 2D4D D.8671 1499 PLAT D.7261 - D.89D8 - D.8524 - Table 3: Estimates of n and respective values k, of datasets I, II and III.