Weak convergence of a pseudo maximum likelihood estimator for the extremal index (Q111091)

Let \( X_1,X_2,\dots\) be a (strictly) stationary sequence of real-valued random variables with a marginal distribution function \( F \). Let \( \hat{F_n}(x)=n^{-1}\sum^n_{s=1}\mathbf 1(X_s\leq x)\) be the empirical distribution function of \( X_1,\dots,X_n\). The extremal index \( \theta \in(0,1]\) measures the tendency to form temporal clusters of extreme values. It can be interpreted as being equal to the reciprocal of the mean cluster size. The authors consider blocks estimators for \(\theta\): a. The disjoint blocks estimators. Divide the sample of length \(n\) into \(k_n\) blocks of length \(b_n\). For simplicity assume that \( n=b_nk_n\). For \( i=1,\dots,k_n\), let \( M_{ni}= \max\left\{ X_{(i-1)b_n+1},\dots,X_{ib_n}\right\}.\) Define the estimators by \[ \hat{\theta}^{\mathrm{N,dj}}_{n} = \left( \frac{1}{k_{n}} \sum^{k_{n}}_{i=1} \hat{Y}_{ni} \right)^{-1}, \qquad \hat{\theta}^{\mathrm{B,dj}}_{n} = \left( \frac{1}{k_{n}} \sum^{k_{n}}_{i=1} \hat{Z}_{ni} \right)^{-1}, \] where \( \hat{Y}_{ni} =-b_{n} \log \left(\hat{N}_{ni} \right) \) and \( \hat{Z}_{ni} = b_{n} \left(1-\hat{N}_{ni} \right) \), \( \hat{N}_{ni} = \hat{F}_{n}\left( M_{ni} \right) \). \newline b. The sliding blocks estimators. Divide the sample of length \(n\) into \(n-b_{n}+1\) blocks of length \(b_{n}\). For \( t=1,\ldots,n-b_{n}+1\), let \( M^{\mathrm{sl}}_{nt}= \max \left\{ X_{t},\ldots,X_{t+b_{n}-1} \right\}. \) Define the estimators by \[ \hat{\theta}^{\mathrm{N,sl}}_{n} = \left( \frac{1}{n-b_{n}+1} \sum^{n-b_{n}+1}_{t=1} \hat{Y}^{\mathrm{sl}}_{nt} \right)^{-1}, \qquad \hat{\theta}^{\mathrm{B,sl}}_{n} = \left( \frac{1}{n-b_{n}+1} \sum^{n-b_{n}+1}_{t=1} \hat{Z}^{\mathrm{sl}}_{nt} \right)^{-1}, \] where \( \hat{Y}^{\mathrm{sl}}_{nt} =-b_{n} \log \left(\hat{N}^{\mathrm{sl}}_{nt} \right) \) and \( \hat{Z}^{\mathrm{sl}}_{nt} = b_{n} \left(1-\hat{N}^{\mathrm{sl}}_{nt} \right) \), \( \hat{N}^{\mathrm{sl}}_{nt} = \hat{F}_{n}\left( M^{\mathrm{sl}}_{nt} \right) \). From the authors' abstract: Both disjoint and sliding blocks estimator for the extremal index are analyzed in detail. In contrast to many competitors, the estimators only depend on the choice of one parameter sequence. We derive an asymptotic expansion, prove asymptotic normality and show consistency of an estimator for the asymptotic variance. Explicit calculations in certain models and a finite-sample Monte Carlo simulation study reveal that the sliding blocks estimator outperforms other blocks estimators, and that it is competitive to runs- and inter-exceedance estimators in various models. The methods are applied to a variety of financial time series.