An impact-based extreme event catalogue on southwest Germany: Overview, Clustering and Triggers

Multiple hazards of different types, such as heat waves, floods, or storms, occurring either simultaneously or in serial clusters can significantly enhance adverse effects on society, economy, and the environment compared to single events. The disastrous flood in Western Europe in 2021 once again showed that natural hazards can lead to severe building damage and thus pointed to the importance of insurance coverage against such events.

To better understand how multiple hazards translate into impact, we propose an economic approach using a unique residential building insurance dataset for southwest Germany ranging from 1986 to 2020. This dataset includes both the number of damage claims reported to a building insurance company and insured losses on a daily resolution, aggregated over the federal state of Baden-Wuerttemberg. This study area is chosen because of the high insurance coverage and therefore high reliability of the data to capture the most important events compared to other states in Germany. In the first step, an event catalogue regarding serially clustered events was elaborated using different methods and statistics. Only convective storms, winter storms and floods are taken into account as these events cause most of the economic damage compared to other events, such as heat waves. To filter smaller events with limited impact and to remove high-frequency clustering, various methods to aggregate the loss events over several days are applied and compared, such as runs declustering using the Peak-Over-Threshold method and an aggregation method considering a fixed number of days, which is common in the insurance industry. After further separating the events according to the relevant seasons, we apply and compare three different clustering methods to the filtered economic dataset: (a) the Poisson regression method, (b) Ripley’s K, and (c) the counting method.

Results show that a high percentile (e.g., 95th or 99th) is needed to analyse the dataset with regard to the most damaging events. This is because the dataset shows a strongly right-skewed distribution. Furthermore, it is found that a small number of high-impact events dominate the overall damage. We show that different hazard types exhibit different behaviours regarding economic metrics (e.g., average loss or correlation between damage claims and insured loss). It is also found and discussed that the degree of clustering depends on the method chosen. For this reason, we performed sensitivity tests and applied different methods to estimate the reliability of the results. To better differentiate between the meteorological event types (e.g., pluvial vs. fluvial floods and convective gusts vs. windstorms), the dataset is further filtered with precipitation data and a dataset on turbulent wind gusts. Building on the final event set with the different event types, the time frames identified by the analyses above are combined with large-scale weather patterns that were dominant at the times when the loss events occurred. This is done to identify relevant relationships of extreme events and their clusters to large-scale processes and mechanisms (e.g., weather regimes or teleconnection patterns).