Heavy tailed time series with extremal independence (Q2352978)

The purpose of this paper is to apply the conditional extreme value approach to the study of heavy tailed time series, having ``extremally independent'' finite-dimensional distributions. The effect of the extreme values at time zero on the future of the time series is measured in terms of regular variation. The authors consider a refined assumption which imposes the vague convergence of a ratio of probabilities (involving the first \(h+1\) terms of the time series and a sequence of scaling functions) to a nonzero Radon measure \(\mu_h\), for each positive integer \(h\). The measure \(\mu_h\) is assumed to be not concentrated at infinity. Under this assumption, the authors obtain a series of results related to the tail of products of regularly varying random variables, the conditional tail expectation, and the tail process. The conditional scaling functions are computed for Markov chains, exponential autoregressive models and stochastic volatility models. This study develops with finesse the scope of the conditional extreme value theory.