Title: Estimating return levels from serially dependent observations
Authors: David Walshaw - Newcastle University (United Kingdom) [presenting]
Lee Fawcett - Newcastle University (United Kingdom)
Abstract: A practical overview of the problem of estimation of return levels for serially dependent observations is taken. Starting with the basic idea of an Exceedances-Over-Thresholds analysis, we consider the range of options available for dealing with the clustering of exceedances brought about by the temporal dependence in the process. We discuss the bias incurred by using a Peaks-Over-Thresholds (POT) approach, and the various alternatives for addressing this. Simply using all exceedances removes the bias but underestimates the variance of estimators. This can be dealt with by using composite likelihood based methods to inflate the variance estimates. However more sophisticated methods based on estimating the extremal index which characterizes the strength of extremal clustering enable us to produce model-based estimators which perform well, and can also be employed in the context of Bayesian inference.