To model for natural disasters, insurance companies rely on modeling expertise from firms such
By doing so, they can avail themselves of a deep well of scientific knowledge and models consistently improved to reflect the latest research. Yet, insurers will have to choose between the models or select multiple models a find a way to blend them. Oddly, this approach has become more difficult lately as the primary hurricane models from the vendors become more divergent, notes AIR in a new report, “Assessing U.S. Hurricane Risk: Do the Models Make Sense?”
“In the last two years, both AIR Worldwide and RMS released new versions of their hurricane models for the United States,” the report states. “Model users might reasonably expect that with the wealth of new data and new research on hurricane risk that has emerged over time—and which is available to all model vendors—modeled loss estimates would converge. Yet with this latest round of updates, we find ourselves more divergent in our views of risk than ever. This leaves it to companies to make sense of the differences, reconcile them with their own loss experience, and assess the implications on their risk management operations.”
To aid in this endeavor, the report lists several key requirements insurers should consider when evaluating a model. The model should: be consistent and unbiased when tested against a wide range of historical datasets, produce reasonable and unbiased loss estimates in real time, reflect basic physical principles of the underlying hazard and produce reliable and unbiased estimates of loss under today’s climate conditions. Moreover, the model should not require dramatic updates in response to a single actual event or even to individual hurricane seasons.
AIR proposes a series of simple tests insurers can employ to how well their catastrophe models reconcile with reality, for example comparing estimated loss to actual loss experience. Elsewhere, insurers can run tests to evaluate a model accuracy regarding frequency and severity. “While robust loss estimates are the most fundamental measure by which to judge a model, there are additional types of validation that model users can perform to ensure that their U.S. hurricanes model abides by basic physical properties observed in nature and are in line with over 110 years of observations,” the report states. “For example, a model’s stochastic catalog should contain more Category 1 and 2 hurricanes than major hurricanes (Category 3, 4, and 5). The reason is that a confluence of often independent meteorological factors is necessary for a major hurricane to make landfall in the United States.”
