A common belief is that a model of normalized hurricane losses (for instance given here) contains greater uncertainty then does more complex models of loss. Below, is a table showing the correlation of variation (CV), a common metric of uncertainty, for several different types of models. Highlighted in yellow, is the CV for Florida’s recorded economic hurricane losses between 1900-2008, adjusted for changes in society (e.g. wealth, inflation, and population). The second row shows the CVs for several commonly used models of historic (1900-2008) insured disaster losses estimated using assumptions about the interaction of the social and geophysical systems. The third row, gives CV’s for insured loss estimates using stochastic hurricane catalogs that use assumptions about interactions in the geophysical system in addition to the interaction between the social and the geophysical events. Data for the second and third rows were gathered from model submissions to the FCHLPM under 2009 standards.
By this metric, it appears that the measurable uncertainty in the normalized model is not much different from other models and generally lower than the stochastic models that are primarily used in insurance ratemaking. Also, the range of CV for these several commonly used models indicates that decisions about models are based on something other than measurable uncertainty (otherwise, the model with the lowest CV would always be chosen).
Unmeasurable uncertainty is inherent in all models. This type of uncertainty comes from a lack of complete knowledge about the world. For instance, what is the certainty about the accuracy of loss data pre-1950? There wasn’t much by way of insured hurricane loss data prior to the widespread sale of the homeowners insurance policy in 1950 and prior to about the mid 1980’s, hurricane damage losses were not consistently collected in any orderly way. So, judgment of this uncertainty is based on personal feelings about preferred means of estimating loss.
Many shy from the mash up of historical loss data from a variety of sources- newspapers, meteorological reports, etc.- and assumptions about the rate of population and wealth growth on regional and local scales. They prefer the consistent means of generating past loss estimated by using vulnerability functions. But even still, losses generated in this way are expected to resemble experienced losses- that is, what is accepted to have been experienced and constitutes a resemblance. Consequently, the knowledge limits that exist in the historical loss data exist in the modeled losses, in addition to the knowledge limits associated with predictive functions.
So the presumption that one model is more uncertain than another model is a reflection of preferred depictions of past and/or future reality. But who ought to have the power in constructing virtual realties or choosing among them? This matters because in the "real world" people stand to benefit or lose from the implementation of models in decision-making. For example, Aon Benfield claimed that record levels of reinsurer capital is due to a lack of catastrophic losses which is "a reasonable departure from the building view of a "new normal" higher level of global catastrophes."
But unlike the risk, the money is not virtual. It came from the pockets of policyholders.