Modeling storm damage risk in Germany

Extratropical cyclones generate high societal costs across Europe, prompting numerous studies that aim to model their economic impacts. The majority of existing building damage models are limited to the maximum wind gust as their sole predictor, applied either directly or through a derived metric (e.g., the cubic exceedance of the 98th percentile). When these models are applied to insurance loss data on the district level for Germany, the resulting spatial patterns are counter‑intuitive: the highest modeled vulnerability appears in coastal regions that are typically best adapted to wind risk, while the lowest vulnerability is found in areas with the weakest adaptation pressure. This discrepancy raises doubts about the adequacy of the current modelling approach.

In our study we employ a Generalized Additive Model (GAM) based on logistic regression to estimate storm damage risk for Germany. The model is trained with ERA5 meteorological variables and daily monetary damage data ranging from 1997 to 2023 supplied by the German Insurance Association (GDV) for the 400 German districts. Beyond the daily maximum gust speed, we test additional predictors, including daily maximum instantaneous wind speed, gust factor (the ratio of maximum gust speed to maximum wind speed), storm duration and precipitation amount.

Wind speed improves model skill relative to gust speed and produces vulnerability maps that better align with expectations based on societal adaptation patterns. A model that combines wind speed, gust factor, and storm duration yields the highest predictive performance, while precipitation adds no value. Although ERA5 wind speed and gust speed are highly correlated under normal conditions, this correlation weakens significantly during storm events. Consequently, we argue that both wind speed and gust speed variables should be retained in storm damage models. Using the extended model, we identify the districts in central Germany as the most vulnerable to storm damage, overturning the earlier, coastal‑biased results. Our findings demonstrate that relying solely on maximum gust speed overlooks important aspects of storm impacts. Incorporating multiple storm characteristics, particularly wind speed, gust factor, and duration, significantly enhances the explanatory skill of damage models.

In the future we plan to apply this damage model to climate model output data to assess projected storm damage risks under future climate scenarios.