CERAMATHS - DMATHS seminar : presentation by Olivier Lopez

The weekly seminar of the Department of Mathematics will host, on Thursday, December 1, Olivier Lopez (Sorbonne University, Institute of Statistics), for the following talk:

Generalized Pareto regression trees: applications to cyber insurance pricing and natural disaster costing

(co-authors: Maud Thomas, Sébastien Farkas, Antoine Heranval)

Summary: Extreme value theory aims at analyzing the tail of distributions in order to determine high quantiles, useful in particular in risk management (anticipating the level of a flood, an amount of financial losses in an unfavorable scenario etc.). The generalized Pareto distributions play, in this framework, a particular role because they allow to approach the tail of distribution of a large class of random variables. This work is part of the "tail index regression" framework, whose goal is to analyze how the tail distribution (i.e. the parameters of the generalized Pareto distribution allowing to approximate a distribution) depends on explanatory variables. We implement a regression tree method for which we show new convergence results. We consider two illustrations: the classification of individuals and incidents in cyber insurance, and the anticipation of the amount of a natural disaster hitting a certain region.