Tail index estimation and an exponential regression model (Q1582511)

A regression model for spacings of extreme order statistics from a Pareto type distribution is proposed. Namely, let \(X_1, X_2,\dots,X_n,\dots\) be a sequence of positive independent and identically distributed random variables with distribution function \(F.\) It is supposed that \(F\) is of Pareto type, i.e. there exists a positive constant \(\gamma\) for which \(1-F(x)=x^{-1/\gamma}l_{F}(x),\) where \(l_{F}(x)\) is a slowly varying at infinity function. The problem of estimating the tail index \(\gamma>0\) from a sample size is considered. The regression approach is helpful in constructing point estimates and confidence intervals for \(\gamma.\) It can also be used in the adaptive estimation of the optimal number of order statistics to be used when applying the \textit{B.M. Hill} [Ann. of Statist. 3, 1163-1174 (1975; Zbl 0323.62033)] estimator or other ``classical'' estimators. Other important problems for which this approach offers some promising perspectives are the estimation of high quantiles, the incorporation of covariate information, and the extension towards time series models.