A method of semi-automated classification of windstorms over France is proposed, in order to provide insurance companies with meteorological statistical values, that could lead to an evaluation of storm risks. Because not all storm events have the same origins or the same features, and do not always hit the same regions, a classification is necessary in order to assemble storm events with meteorological and geographical similarities. The classification is based on the expert knowledge of a weather forecaster, who first analyses 44 storm events and draws out the meteorological features of different types of storm. Afterwards, the results of the human typology are automated by determining decisive rules with criteria, which are calculated with only observed meteorological surface data over France. The automation of the human typology is obtained by taking into account the main results of statistical methods such as objective classification or discrimination and by validating them with physical considerations. We thus succeed in formulating the reasoning of the expert. The classification is then applied to the Météo-France records since 1962. We succeed in selecting 524 storm events and in classing them into seven main types, with enough events in each class to calculate statistical values. © 1998 Royal Meteorological Society