Probabilistic Regional Conditioning of Natural Hazard Loss Models

Dennis Johannes Wagenaar, Maricar Rabonza, Mariano Balbi, Agustin Bertero, Tian Ning Lim, David Lallemant

Natural hazard risk models underpin decisions from insurance pricing to infrastructure investment, yet their accuracy depends on vulnerability functions rarely calibrated to local conditions. The most accurate vulnerability functions capture regional building characteristics through multi-variable models or large engineering-based loss-function databases, but need detailed asset-level data that is rarely available. This paper introduces a method that condenses such models and databases into single-variable loss functions tailored to a region, without asset-level data collection. Rather than weighting all loss functions equally, as conventional blending does, the method assigns each a probability based on how well it matches the regional building stock, inferred from samples or expert judgement, and updates these probabilities using Bayes theorem as building data becomes available. Applied to a wind loss-function database and a multi-variable flood loss model, regional conditioning reduces absolute and bias errors versus equal-weight blending, especially where building stock differs most from the original model context.