Different rating methods which allow for exceptional large claims are discussed. A robust Bayesian statistical model is proposed which can cope with non negative, skewed data. An example from fire insurance is analyzed. The performance of the posterior mean is compared to the performance of a robust credibility estimator.