A Bayesian spatio-temporal model for precipitation extremes -- STOR team contribution to the EVA2017 challenge (Q1792629)

This paper outlines one approach which was used in the EVA2017 challenge, whose target was to predict extreme quantiles for rainfall at several sites in the Netherlands. Letting \(R_{j,m}\) denote the daily rainfall amount at site \(j\) in month \(m=1,\ldots,12\), the authors model the transformed random variable \[ \tilde{R}_{j,m}=\log\left(1+R_{j,m}\right) \] using a Bayesian hierarchical model. Their model assigns a probability \(p_{m,j}\) of observing a non-zero amount of rainfall. The amount of rainfall (for days on which this is positive) is then modelled using an extremal mixture model; given a threshold \(u_{j,m}\), rainfall below this threshold is modelled using a truncated Gamma distribution, and rainfall above this threshold using a generalised Pareto distribution. The prior in the Bayesian model aims to exploit the spatial and seasonal structure by assuming that parameters for neighbouring sites and consecutive months are likely to be similar.