A Climate-Informed Generalized Extreme Value Model for Global Precipitation Extremes

Quantifying the intensity and frequency of extreme precipitation remains a fundamental challenge in climate science, particularly in regions with limited observational records. While Global Climate Models (GCMs) often estimate extremes with bias and Machine Learning (ML) approaches lack interpretability, we investigate whether the complicated spatial variability of extremes can be captured by a low-dimensional climate manifold. We propose a modelling framework based on Extreme-Value Theory (EVT) to assess the annual maximum 1-day (Rx1day) and 5-day (Rx5day) precipitation using three physically interpretable covariates. We construct a non-stationary Generalized Extreme Value (GEV) model where location and scale parameters are driven by the mean and standard deviation of precipitation in the wettest month, and structurally constrained by the Köppen–Geiger climate class. The model is fitted to 85 years of ERA5 reanalysis data, and uncertainty is quantified through bootstrapping. Validation against empirical quantiles demonstrates that this simple, low-order framework can successfully reproduce the spatial patterns and magnitude of local precipitation extremes. These findings suggest that precipitation extremes can be understood in terms of basic hydroclimatic constraints, providing a theoretical baseline for benchmarking complex models and assessing the predictability of extremes in global models, with potential applications in flood management, infrastructure design, and (re)insurance pricing in data-poor locations.