A0895
Title: Manifold MCMC algorithm for Gamma-GPD mixture model
Authors: Salah El Adlouni - Universite de Moncton (Canada) [presenting]
Abstract: Modelling extreme events is important in many domains, including environmental variables, civil engineering, reliability, financial risk, and computer security. The peaks over threshold approach (POT) describes the main characteristics of the observed extreme series, yet the threshold selection is challenging and might affect the results. Mixture models offer more flexibility to represent samples with heterogenous data. The Gamma-Generalized Pareto mixture model (GAM-GP) is presented for extreme risk estimation. The model is developed in its general form, where the observed events depend on multi-dimensional covariates and non-linear link functions. A new Monte Carlo Markov Chain algorithm, based on Riemannian Manifold, is developed to estimate the parameters in a Bayesian framework. Results show the capacity of the proposed algorithm to converge to the posterior distribution of the parameters even for the high dimension of the covariates space. The approach is illustrated on simulated data and a daily streamflow dataset for the Saint John River at Fort-Kent (upstream) New-Brunswick (Canada).
