Bayesian inference for extremes: accounting for the three extremal types (Q2488461)

The distribution of the observed exceeds of a high threshold is modelled by a non-homogeneous Poisson point process with generalized Pareto intensity \[ \lambda_\vartheta=\sigma^{-1} \left\{ 1+\xi\left((x-\mu)/\sigma\right) \right\}_{+}^{-(\xi+1)/\xi}. \] A Bayesian technique of \(\vartheta=(\mu,\sigma,\xi)\) estimation is described. A prior is constructed in such a way that it incorporates a priori information of three level quantiles of the estimated distribution and allows the case \(\xi=0\) (related to the Gumbel extreme type of the underlying distribution) with nonzero probability. A reversible jump Markov chain Monte Carlo is proposed for computation of the posterior distribution. Applications to sea-level data and data on rainfalls are described.