Some aspects of extreme value statistics under serial dependence (Q1003318)

This paper gives an overview of modern time series extreme value statistics, including the estimation of the tail index \(\gamma\) of a time series' stationary marginal distributions and the extremal index \(\vartheta\) of a stationary time series. Two strategies for estimation of \(\gamma\) are compared: direct estimation by Hill-type estimates under serial dependence and model-based estimation in which tail index estimates by nearly independent residuals (in some linear models) are used to construct an estimate for \(\gamma\). An example is presented which demonstrates that the second approach (which yields asymptotically more efficient estimates) can fail under moderate deviations from the theoretical model of dependence, which are difficult to detect by statistical means. Discussing the extremal index \(\vartheta\) and its estimation, the author stresses that in many applications \(\vartheta\) does not give the information one is actually interested in.