The MELBS team winning entry for the EVA2017 competition for spatiotemporal prediction of extreme rainfall using generalized extreme value quantiles (Q1792637)

This paper outlines the winning approach used in the EVA2017 challenge, whose target was to predict extreme quantiles for rainfall at several sites in the Netherlands. The authors' approach is based on modelling the monthly maxima in rainfall using generalized extreme value distributions \(Z\sim\text{GEV}(\mu,\sigma,\xi)\), with distribution function \[ P(Z\leq z)=\exp\left[-\left\{1+\xi\left(\frac{z-\mu}{\sigma}\right)\right\}_+^{-1/\xi}\right]\,, \] where \(x_+=\max\{h,0\}\), and \(\mu\in\mathbb{R}\), \(\sigma\in\mathbb{R}^+\) and \(\xi\in\mathbb{R}\) are location, scale and shape parameters, respectively. Letting \(Z_{s,m}\) denote the monthly maximum rainfall at site \(s\) in month \(m=1,\ldots,12\), the authors considered the model \(Z_{s,m}\sim\text{GEV}(\mu_{s,m},\sigma_{s,m},\xi_{s,m})\). Concluding that this model is overparameterised, they eventually opted to combine the data from all sites, fitting the model \(Z_m\sim\text{GEV}(\mu_m,\sigma_m,\xi)\) to the combined monthly maximum data. The parameter \(\xi=0.25\) was fixed, and the parameters \(\mu_m\), \(\sigma_m\) fitted by maximum likelihood using a rolling window cross-validation approach.