Bayesian analysis of extreme values by mixture modelling (Q1887257)

It is supposed that the observed exceedances over some fixed threshold are generated by a Poisson process with intensity \[ \Lambda_y= ( 1+\xi(y-\mu)/ \psi)^{-1/\xi}, \] where \(\mu\) is the location, \(\psi\) is the scale and \(\xi\) is the shape parameter. A Bayesian analysis technique is described for the parameter \(\eta=(\mu,\psi,\xi)\) with a prior according to the mixture distribution \(\sum_{j=1}^k w_j p_j(\cdot\mid\delta_j)\) (here \(w_j\), \(\delta_j\) and \(k\) are hyperparameters). This construction allows the description of heterogeneous data. An application to insurance claim data is presented. Markov Chain Monte Carlo is used for the prior evaluation.