Predicting Incident Size from Limited Information

Abstract

Predicting the size of low-probability, high-consequence natural disasters, industrial accidents, and pollutant releases is often difficult due to limitations in the availability of data on rare events and future circumstances. When incident data are available, they may be difficult to fit with a lognormal distribution. Two Bayesian probability distributions for inferring future incident-size probabilities from limited, indirect, and subjective information are proposed in this paper. The distributions are derived from Pareto distributions that are shown to fit data on different incident types and are justified theoretically. The derived distributions incorporate both inherent variability and uncertainty due to information limitations. Results were analyzed to determine the amount of data needed to predict incident-size probabilities in various situations. Information re-quirements for incident-size prediction using the methods were low, particularly when the population distribution had a thick tail. Use of the distributions to predict accumulated oil-spill consequences was demonstrated.