A weather forecast begins with numbers. Thousands of stations on land, balloons in the sky and satellites in space send measurements of temperature, wind and moisture. A computer takes these numbers and calculates what the air will do next. This is hard work, because the air over one country affects the air over another.

For many years, forecasts used equations that describe how air moves. These models are very accurate, but they need enormous machines and hours of computing time. A national weather service may run its model only a few times a day.

Now a different method is growing. Instead of solving equations, a program studies forty years of past weather and learns which patterns usually follow which. Once it has learned, it can produce a forecast in under a minute on a much smaller machine. Several tests have shown that these forecasts match the old models, and sometimes beat them.

This does not mean the old models will disappear. The learning program is only as good as the records it studied, so it may struggle with an unusual event that has never happened before. Most weather services therefore run both systems and compare them, and forecasters still decide what to tell the public.

Speed changes what forecasters can do. When one forecast took hours, a service could try only a few versions. When it takes a minute, the same team can run a hundred slightly different versions and see how many of them predict rain. That is why modern forecasts often give a percentage instead of a simple yes or no.