Estimating uncertainty in the extreme value analysis of data generated by a hurricane simulation model

Extreme value analyses of any environmental phenomenon are fraught with difficulties, but the additional difficulty of collecting reliable data during hurricane events makes their analysis even more complicated. A widely accepted procedure is to use calibrated hurricane models to simulate hurricane events. The simulated data can then be subjected to standard extreme value procedures. The estimation uncertainties which arise from such analyses depend upon (1) the extent to which the hurricane models are physically realistic, (2) the length of the simulated series, which consists of about 1,000 or even 10,000 simulated events, and therefore introduces negligible errors, and (3) the length of the historical record on which the simulations are based, which usually consists of about 50 events. In this paper, we propose the use of resampling schemes in an attempt to obtain some reasonable measure of uncertainties due to the relatively short length of the historical record. An intuitive, "naive" procedure is first described, which leads to an alternative approach that has connections with the statistical procedure of bootstrapping. Standard application of these procedures for extremes induces bias, and we propose a simple, though nonstandard method for reducing this effect. The results are illustrated in detail for a dataset of simulated hurricane wind speeds corresponding to a location in Florida and are also summarized for a sequence of 55 locations along the U.S. Gulf and Atlantic coasts.