Tail index estimation for heavy tails; accommodation of bias in the excesses over a high threshold (Q1003332)

Estimation of the tail index \(\gamma\) is considered by an i.i.d. sample. A new PORT-MP (peaks over random threshold, modified Pareto) estimator is proposed of the form \[ \hat\gamma_n=k^{-1}\sum_{i=1}^k p_{ik}\ln(1+\hat\alpha (X_{n-i+1:n}-X_{n-k:n})), \] where \(X_{i:n}\) are the order statistics of the sample, \(\hat\alpha\) is an estimate of the approximating generalized Pareto distribution parameter \(\alpha\), \(k\) is the sample proportion (a smoothing parameter) and \(p_{ik}\) are weights depending on the second-order parameters \((\beta,\rho)\) of the sample distribution. To obtain a feasible estimator, these parameters are estimated by the sample. Asymptotic normality of \(\hat\gamma_n\) is demonstrated. Its behavior on finite size samples is analyzed via simulations.