This study was designed to examine the performance of a nonparametric forecasting method first developed in meteorology, the "method of analogues," in predicting influenza activity. This method uses vectors selected from historical influenza time series that match current activity. The authors applied it to forecasting the incidences of influenza-like illnesses (ILI) in France and in the country's 21 administrative regions, using a series of data for 938 consecutive weeks of ILI surveillance between 1984 and 2002, and compared the results with those for autoregressive models. For 1- to 10-week-ahead predictions, the correlation coefficients between the observed and forecasted regional incidences ranged from 0.81 to 0.66 for the method of analogues and from 0.73 to -0.09 for the autoregressive models (p < 0.001). Similar results were obtained for national incidence forecasts. From the results of this method, maps of influenza epidemic forecasts can be made in countries in which national and regional data are available.