A Statistical Characterization of a Simulated Canadian Annual Maximum Rainfall Field

The paper explores the use of a joint log Gaussian measurement distribution to represent a field of extreme values. A hierarchical Bayes prior distribution with estimated hyperparameters models uncertainty in that distribution’s unknown parameters. The resulting multivariate t distribution is fitted to the random maximum annual rainfall field simulated from a Canadian Climate model with 319 grid cells. To assess its performance in that application, we find its predictive credibility ellipsoids and confirm that their coverage fractions are close to their nominal credibility levels, using cross validation. Return values are estimated for cell marginal distributions and their overall accuracy is assessed by an index of discrepancy.