Dynamic modelling of extreme daily precipitation in Germany from 1951 to 2020

It is important to analyse long-term changes in heavy precipitation but current risk management requires more dynamic and reliable forecasting of changes in the right tail of precipitation distributions in shorter periods of time. Robust prediction of high quantiles, for one year ahead, may be very useful in current risk management. We study the predictive reliability of models based on 30-year rolling-window. The Peak Over Threshold method is used to estimate 20-year and 50-year return levels. Two recently developed automated and eff cient sequential procedures, the ForwardStop and the StrongStop, are used to select thresholds for raw data and after declustering of the data. The hit process is applied to validate results in the backtesting procedure. This procedure was repeated for over 400 sites in Germany to generate dynamic return-level maps of extreme precipitation.