Extreme value frameworks for sub-hourly rainfall: comparison of predictive performance across Europe

Several extreme value frameworks are available for modelling rainfall extremes. These include classical asymptotic approaches, such as the Generalised Extreme Value (GEV) distribution applied to annual maximum series, as well as more recently proposed non-asymptotic methods, namely the Metastatistical Extreme Value (MEV) and the Simplified MEV (SMEV) distributions applied to ordinary events. While the non-asymptotic frameworks have been evaluated at daily and hourly timescales, they have not yet been systematically evaluated at sub-hourly timescales across climatic regimes. As a result, it remains unclear whether relative differences in predictive performance observed at longer timescales extend to sub-hourly durations.

We compare the predictive performance of the GEV, MEV, and SMEV distributions using sub-hourly rain gauge observations from 2,810 stations across six European countries. We conduct a cross-validation experiment in which at-site distribution parameters are estimated from a training subset and used to predict the return level associated with the most extreme event in an independent test subset. Performance is quantified as the root mean square error between predicted return levels and observed extreme events, computed over 1,000 iterations per rain gauge and duration.

Results show systematic differences in relative predictive performance across durations and regions, with SMEV being favoured at short durations (up to 3 hours) for the majority of rain gauges, MEV at longer durations, and GEV being competitive for a non-negligible fraction of rain gauges. Overall, no framework consistently outperforms the others across countries and durations, indicating that superior predictive performance of any one extreme value framework cannot be assumed across space or timescales.