Biases Beyond the Mean in CMIP6 Extreme Precipitation: A Global Investigation

Abstract Climate models are crucial for assessing climate variability and change. A reliable model for future climate should reasonably simulate the historical climate. Here, we assess the performance of CMIP6 models in reproducing statistical properties of observed annual maxima of daily precipitation. We go beyond the commonly used methods and assess CMIP6 simulations on three scales by performing: (a) univariate comparison based on L‐moments and relative difference measures; (b) bivariate comparison using Kernel densities of mean and L‐variation, and of L‐skewness and L‐kurtosis, and (c) comparison of the entire distribution function using the Generalized Extreme Value ( ) distribution coupled with a novel application of the Anderson‐Darling Goodness‐of‐fit test. The results reveal that the statistical shape properties (related to the frequency and magnitude of extremes) of CMIP6 simulations match well with the observational datasets. The simulated mean and variation differ among the models with 70% of simulations having a difference within 10% from the observations. Biases are observed in the bivariate investigation of mean and variation. Several models perform well with the HadGEM3‐GC31‐MM model performing well in all three scales when compared to the ground‐based Global Precipitation Climatology Centre data. Finally, the study highlights biases of CMIP6 models in simulating extreme precipitation in the Arctic, Tropics, arid and semi‐arid regions.