An automated technique is presented for determining an appropriate distributional model for the largest yearly wind speeds. With a view to assessing the validity of current probabilistic approaches to the definition of design wind speeds, this technique was used in a study of extreme wind speeds based on records taken at 20 U.S. weather stations. The following results were obtained: (1)At 83% of the stations not susceptible to experiencing hurricane-force winds, the series of the largest annual wind speeds were well fit by Type I probability distributions of the largest values; (2)the assumption that Type II distributions with γ=9 are generally representative of such stations was not confirmed; (3)type I probability distributions do not appear to describe correctly the behavior of extreme winds in regions subjected to special winds, e.g., hurricanes; and (4)in such regions, 20-yr data samples may provide a misleading picture of extreme wind behavior.