Estimation of the tail parameter in the domain of attraction of an extremal distribution

Abstract The problem of estimating the extremal parameter for a distribution with regularly varying tail is revisited by proposing three new estimators and studying their rate of convergence. In the case of positive extremal parameter, the optimal rate in the sense of Hall and Welsh [Ann. Statist. 12 (1984) 1079–1084] is attained for a generalization of Pickands well-known estimator and for a new estimator. The optimal number of top order statistics involved in the computation of the latter is also studied by using an optimality criterion based on a bias-variance trade-off.