In the last two decades, significant progress in science and technology has driven major advances in the modeling of natural hazards and their impact on the built environment. In addition, several major catastrophe events over the same period highlighted important issues that could affect the underwriting and decision-making processes for selecting and managing risk. In many cases, these events had a profound impact on insurance and reinsurance practices, expanding the use of models, analytics, and exposure-data quality analysis to understand the sources of risk and to model loss uncertainty. Furthermore, since the development of earlier versions of catastrophe models, technology and computing power have advanced enormously, allowing modelers to address key elements of model and loss uncertainty in a more systematic way.
Most catastrophe models implicitly account for the uncertainty associated with various model components. These uncertainties can be characterized as aleatory or epistemic. The aleatory uncertainty represents the inherent variability in the physical system, which cannot be reduced. In contrast, the epistemic uncertainty relates to the lack of knowledge about the damage and hazard model components (i.e., building characteristics, ground motion, central pressure, etc.) and thus can be reduced with additional information about the system.
A complete quantification of the contribution of each model component to the epistemic uncertainty requires a systematic approach and thus involves developing multiple alternative versions of the hazard and damage modules. By developing multiple versions, the modeler can capture the plausible range of parameters that control each model component. The risk model can then propagate those ranges into estimating variability of losses for a given portfolio. However, developing such a bottom-up quantification of the catastrophe model uncertainty is time consuming, and carrying out all the required analyses is also computationally expensive for practical applications. Nevertheless, modelers can now take advantage of the robust computational power developed in recent years to quantify systematically the uncertainty in the estimation of risk. This paper illustrates some practical approaches to capture uncertainties in risk models. The issues and challenges that remain in developing risk models and in quantifying the uncertainty, even with today's advancement in science and engineering, and vastly enriched technological environment, are also addressed in this paper. It draws on specific examples from the authors' experience in modeling earthquake and hurricane perils.
