Bayesian hierarchical modelling of intensity-duration-frequency curves using a climate model large ensemble

Abstract. Accurate modelling of extreme precipitation is vital for predicting future risks and informing adaptation strategies. Here, we compare and evaluate six different extreme value statistical models for hourly to 48 h extreme precipitation in southern Germany, with a primary focus on duration-dependent Generalized Extreme Value (dGEV) distributions. To assess model performance, particularly in capturing tail behavior, we utilize the 50-member single model initial-condition large ensemble of the Canadian Regional Climate Model version 5 for the period 1980–2019. The large sample size of 2000 simulated years enables a robust sampling of extreme quantiles. Using a sub-sampling strategy with 30 to 100 years, we compare the efficacy of Bayesian methodology, in particular Bayesian hierarchical models, against frequentist models (L-moments and Maximum Likelihood Estimation – MLE) in representing the tail risk of 100-year return levels based on limited sample sizes. Hierarchical models allow us to give special emphasis on the dimensionality of the GEV shape parameter, a critical factor for tail behavior. Our findings reveal that a shape parameter varying over durations but fixed across space is beneficial for the prediction of the 100-year return level. The resulting Intensity-Duration-Frequency (IDF) curve shows the highest accuracy and smallest confidence intervals proving its robustness. Compared to the standard GEV estimated by L-moments, our proposed model can reduce the relative error of the 100-year return level from 18.1 % to 8.8 % based on a 30-year sample size. Furthermore, our analysis reveals fundamental limitations of the Anderson-Darling test for extreme value model selection, demonstrating its poor correlation with predictive skill for upper quantiles – a critical finding for climate risk applications.