The use of physico-mathematical models for the numerical prediction of weather changes on the global scale is described. The accuracy of the predictions is assessed in relation to the limitations of both the observational data and the representation in the model of the many interactive physical and dynamical processes that govern the evolution of the major features of the atmospheric circulation. The concept of predictability in relation to the evolution of atmospheric disturbances is discussed. Given an adequate global coverage of observations and continued improvement in the models, it should be possible to extend the range of useful forecasts up to about 14 days; this may prove to be the limit of deterministic predictability set by the random nature of atmospheric fluctuations. However, some relatively stable atmospheric states, such as those that produce long dry summers, may possess greater predictability. Furthermore, it may well be possible to predict the general character of the weather for some weeks or months ahead even if the day-to-day variations are unpredictable this far ahead. Because the longer-range forecasts are bound to show considerable variability in skill and reliability, it will be desirable to assign them a confidence rating based on the rate at which an ensemble of model forecasts diverge when starting from slightly different initial states.