Apr 2016-Mar 2017
Math-Fi seminar on 6 Oct.
2016.10.04 Tue up
- Date : 6 Oct. (Thu.)
- Place: W.W. 6th-floor, Colloquium Room
- Time : 16:30-18:00
- Speaker: Marcel Braeutigam
- Title: Detecting changes in extremal dependence
- Abstract: Change-point analysis is a standard tool employed in various applications to detect breaks in a given sample. Classical change-point analysis usually aims to detect changes in the mean, variance or the whole distribution. We however consider detecting changes in the tail behaviour of a given (iid) sample. We start by presenting an existing approach in the one dimensional case which uses extreme value theory. Building on this, our main contribution is to propose a test for a change-point in the bivariate extremal dependence. For this we exploit the fact that a sample of high threshold exceedances allows us to recover an estimate of the angular measure, a probability measure modelling the extremal dependence. This observation is used to propose both parametric and a non-parametric tests for changes in the angular measure. We provide a practical procedure showing how to use our tests and evaluate the approach on simulated as well as real-life data.