Developing multivariable probabilistic flood loss models for companies

Abstract

Decision-making in flood risk management strongly relies on the accurate estimation of monetary flood loss. Recent advancements in the field promote the use of multivariable flood loss models that consider a multitude of damage controlling factors beyond inundation depth. However, the development of novel flood loss models excluded companies for the most part, albeit their considerable contribution to total flood damages. In this methodological study, we propose three probabilistic approaches to flood loss modelling for companies that intrinsically quantify prediction uncertainty. We fit a random forest, a Bayesian network and a Bayesian regression to company loss data for buildings (n=545), which stem from four post-event surveys after floods in Germany. Posterior predictive checks, which give insight on the plausibility of the proposed models, prove that all candidate models reproduce essential characteristics of the observed loss data properly. The predictive training errors suggest that the random forest and the Bayesian network outperform the Bayesian regression. We trace the difference in predictive training error back to distinct model structures and emphasize that the presented model checks represent the groundwork for a detailed model validation.