A two-step estimator of the extreme value index (Q1003330)

Estimation of the tail index \(\gamma\) by an i.i.d. sample \(X_1\),\dots, \(X_n\) is considered. The author proposes a ``two step'' estimator \(\hat\gamma_{STEP}\) which uses any \(\sqrt{k}\) consistent ``first step'' estimator \(\hat\gamma^{(1)}\): \[ \hat\gamma_{STEP}=2^{-1}(2\hat\gamma^{(1)}+1) WM_n^{(2)}(WM_n^{(1)})^{-2}-1, \] where \[ WM_n^{(j)}=\sum_{i=1}^k w_i^{(j)}(X_{n,n-i+1}-X_{n,n-k})^j, \] \[ w_i^{(j)}=(j\hat \gamma^{(1)}+1)^{-1}((i/k)^{j\hat \gamma^{(1)}+1}-((i-1)/k)^{j\hat \gamma^{(1)}+1}), \] and \(X_{n,k}\) are the order statistics of \(X_i\), \(i=1,\dots,n\). It is shown that \(\hat\gamma_{STEP}\) is asymptotically equivalent to the maximum likelihood estimator of \(\gamma\). Results of simulations are presented.