Weather is too chaotic to be predicted more than ten days or so ahead. So why should we believe the global rainfall forecast, pictured here, made a few months in advance?

Well, this isn't really a weather forecast, in which the evolution of individual weather systems is predicted. It is a climate forecast, described elsewhere in this issue (Stockdale, T. N., Anderson, D. L. T., Alves, J. O. S. & Balmaseda, M. A. Nature 392, 370–373; 1998), which uses a global model of the coupled oceanic and atmospheric circulations to calculate the statistical behaviour of large-scale weather patterns and the main physical influences on those statistics. The model calculates not only the mean (predictable) shift in climate for a season, but also estimates the random (unpredictable) component. Forecasts are therefore expressed as probabilities. For example, the model predicts that it is about 70% certain that southeast China will continue to experience the unusually wet weather it has been suffering of late.

The forecast was completed at the end of January 1998, and the map shows predicted total rainfall anomalies for this year's March-April-May season, relative to long-term mean rainfall for the same season; blue is wet, red is dry. Uncoloured areas represent a low-confidence forecast. As well as southeast China, the central United States and western central Eurasia are predicted to be unseasonally rainy, whereas the forecast is that central Mexico and Indo-China will undergo shifts to an unusually dry season. Although some of these predictions are perhaps not so surprising in the wake of the present El Niño, Australia — which would normally suffer drought after an El Niño — seems to be in for a relatively wet time.

The high rainfall that contributed to last summer's severe and unexpected floods in central Europe was predicted using this approach. But a successful rainfall prediction is not necessarily a sign of a reliable forecasting system. On the one hand, a forecast is only a probabilistic statement (so a good forecast might get the rainfall wrong); on the other, the model might forecast the correct rainfall for the wrong reasons. The European Centre for Medium-Range Weather Forecasts, where the map was produced, is investing substantial resources in addressing this ‘verification’ issue, largely by using past datasets to increase model confidence, and is now in a quasi-operational forecast phase. Go and look at the up-to-date seasonal forecasts (http://www.ecmwf.int), but don't bet your house on them.