CMStatistics 2017: Start Registration
View Submission - CMStatistics
B0765
Title: Peaks-over-threshold inference for spatio-temporal processes, with an application to European windstorms Authors:  Raphael de Fondeville - EPFL (Switzerland) [presenting]
Anthony Davison - EPFL (Switzerland)
Abstract: Classical spatio-temporal models for extremes rely on block maxima, but this approach is limited by computational considerations to a few dozen variables. In order to get a better understanding of extremal dependence and reduce model uncertainties, exploitation of gridded datasets, for example from climate models, is necessary. $\mathcal{R}$-Pareto processes based on a peaks-over-threshold approach, use single extreme events, generalize the notion of exceedance, and have relatively simple mathematical expressions. For spatio-temporal modelling, we focus on the Brown--Resnick model, which relies on classical Gaussian models widely used in applications. An efficient algorithm for censored likelihood allows us to perform inference with hundreds of locations. For higher dimensions and generalized risk functionals, we develop an estimator based on the gradient score, whose numerical complexity is similar to likelihood-based inference methods for Gaussian fields. We develop a spatio-temporal model for extreme winter storms over Europe and apply our method on `three-second wind gusts' from the reanalysis dataset ERA-Interim, which covers the period $1979$--$2016$. We can then use the model to generate new extreme storms with previously unobserved intensities and spatial pattern.