The result of five years of operational medium-range forecasts at ECMWF is discussed. It is shown that a considerable improvement in predictive skill has taken place, resulting in more accurate forecasts and an extension in time of useful predictive skill. For the Northern Hemisphere extratropics, the improvement in skill during the winter season is about three days for the 500-mb geopotential compared to the very first experiments in the early 1970s. There are large geographical and temporal variations in the performance of the model, as well as large-scale errors of a systematic nature. Numerical experiments have indicated great sensitivity to the representation of orography and to tropical forcing.