Estimation of the extremal index using censored distributions (Q83310): Difference between revisions

Classical extreme value theory is based on independent and identically distributed variables. Extreme events in time series usually violate such assumption. The dependence translates on occurrence of extremes close to each other, i.e., clustering of extremes occurs. A parameter describing clustering of such a type is called extreme index, denoted by \(\theta\), \(0 < \theta \leq 1\), with \(\theta=1\) being related to independence. A considerable literature is available on \(\theta\) estimators. Among those estimators, it is of special interest in the current paper, the Ferro and Segers interval estimator, \(\hat{\theta}_I,\) and the Sürveges and Davison \(K\)-gaps estimator, \(\hat{\theta}_{SD}\). This paper proposes a new estimator, \(\hat{\theta}_C,\) based on an artificial type I left censoring and maximum likelihood. This approach can be seen as a generalization to the \(K\)-gaps estimator. A local dependence condition \(D^{(k)}_n(u_n)\) is required so that all the steps behind both \(K\)-gaps and censored estimators can follow. The censored estimator variability is studied using the Fisher information measure. Asymptotic normality results are established for \(\hat{\theta}_C\). The \(\hat{\theta}_C\) performance is compared to the ones of \(\hat{\theta}_I\) and of \(\hat{\theta}_{SD}\), through the simulations of a bunch of classical stationary processes. Also, an application to the daily maximum temperatures at Uccle, Belgium, is presented, in comparing the three estimators. In general, the censored estimator performs better than the other two.