Towards improved hailstorm and loss prediction using random forest in the Netherlands

Extreme hailstorms, linked to convective systems, can cause significant societal impact and account for ~30% of insured losses from 2007 to 2022 in the Netherlands. Stochastic models can quantify hailstorm impacts, but assessing the hazard and its consequences remains challenging due to a lack of consistent observations and detailed post-disaster losses. Our study addresses this research gap and investigates hailstorm occurrence and impacts by combining high-resolution radar-based data, meteorological observations, and reanalysis data with a unique asset-level insurance loss dataset for the Netherlands.Random forest models are used to identify proxy variables for hailstorm hazard and associated losses. Meteorological and convective variables are used to identify hail or damage days, testing different Maximum Expected Hail Sizes (MEHS) thresholds and train-test designs. Training the model solely on hail days and the entire hazard time series results in poor performance with correct predictions but many false positives. However, using a 70-30% (train-test) random selection from the entire hazard series enhances performance, especially when the training sample contains more positive observations. Random forest models, despite not providing direct hazard intensity information, effectively highlight influential proxies like CAPE and dewpoint temperature, which can be refined to enhance hailstorm prediction, frequency, and damage thresholds. Random forest models appear to be a promising option for further improving hazard and loss models. Our findings offer valuable insights to enhance hailstorm hazard and loss assessments.