Spatial risk zonation of belgian flood reinsurance: A clustering approach to catastrophe model output

Details

Supervisors

Eddy Vanbeneden

Faculty

Faculté des sciences

Degree label

Master [120] en sciences actuarielles, à finalité spécialisée

Abstract

enThis thesis develops a data-driven framework for the geographic zonation of Belgian inland flood risk, grounded in the output of a probabilistic catastrophe model. The standard segmentation frameworks used in the (re)insurance industry, such as CRESTA zones and administrative province boundaries, follow political and historical lines rather than the hydrological processes that govern flood risk generation. The central contribution of this work is the use of a Damage Distribution Table (DDT), a model output recording the spatial footprint of each simulated flood event as provincial loss proportions, as the empirical basis for clustering. This allows a data-driven algorithm to recover the underlying hydrological structure of Belgian flood risk directly from the model's own loss estimates, without imposing any prior geographic convention. Five clustering families are implemented and compared, including K-Means, Gaussian Mixture Models, Hierarchical Energy Clustering, and K-Groups, evaluated against both geometry-based and energy distance-based internal validation indices. The DISCO F-ratio and normalised H-matrix, , are introduced as cluster quality measures sensitive to differences in shape and tail behaviour rather than geometric compactness alone. All algorithms consistently recover a three-cluster structure aligned with the primary drainage basins of Belgium: a Meuse cluster concentrated in Liège and Namur, a central Scheldt cluster concentrated in Oost-Vlaanderen, and a low-risk Scheldt cluster. The Meuse basin emerges as the dominant source of tail risk, accounting for the entire portfolio Solvency Capital Requirement at the 200-year return period.