Comparison of local and basinwide methods for risk assessment of tropical cyclone landfall

Two statistical methods for predicting the number of tropical cyclones (TCs) making landfall on sections of the North American coastline are compared. The first method-the local model-is derived exclusively from historical landfalls on the particular coastline section. The second method-the track model-involves statistical modeling of TC tracks from genesis to lysis, and is based on historical observations of such tracks. Identical scoring schemes are used for each model, derived from the out-of-sample likelihood of a Bayesian analysis of the Poisson landfall number distribution. The track model makes better landfall rate predictions on most coastal regions, when coastline sections at a scale of several hundred kilometers or smaller are considered. The reduction in sampling error due to the use of the much larger dataset more than offsets any bias in the track model. When larger coast sections are considered, there are more historical landfalls, and the local model scores better. This is the first clear justification for the use of track models for the assessment of TC landfall risk on regional and smaller scales.