On the estimation of extreme tail probabilities (Q1364752)

Assume \(F\) is an unknown df and \(\overline F=1-F\) is approximated by an estimable quantity \(\overline F_\theta\). Choose \[ D_1(t;n,x)=E(\overline F_{\widehat\theta(t)}(x)-\overline F(x))^2 \] as a criterion for the approximation, where \(\widehat\theta(t)\) is an estimator of \(\theta\), depending on a tuning parameter \(t\). How to estimate in this case \(\overline F(x)\) at a point \(x\) well beyond the extent of the data \({\mathcal X}=\{X_1,\ldots,X_n\}\) is the main problem in the present paper. The authors suggest an approach founded on ``pulling back'' the problem to a location within the range of the data, by reducing the values of \(n\) and \(x\) to appropriate \(m\) and \(y\). Now one can approximate \(F\) by the empirical df and \(\overline F_{\widehat\theta}\) by its bootstrap version. The authors demonstrate this method in case when \(\overline F\) is regularly varying with remainder and \(\widehat\theta=\widehat\theta(k)\) depends on the \(k\) largest order statistics of the sample \({\mathcal X}\).