Beyond two weeks, the best weather forecasts often do worse than a simple average of past years. Small errors pile up day after day, and specialists speak of a “predictability desert” between two and six weeks out. A team of researchers from Harvard, the University of Toronto, MIT, and Microsoft Research has described in the journal Nature a tool that helps cross part of that gap.
The tool does not forecast the weather itself. It takes a model’s forecast, expressed as probabilities that a week will be warmer, cooler, drier, or wetter than usual, and corrects it. To do so, it compares 20 years of past forecasts with what actually happened, removes the errors the model keeps repeating, and takes the weather of recent weeks into account. It retrains every day, without a supercomputer, and its code is free to use.
Tested on three systems, it improved each of them, and it doubled the skill of an AI model three and four weeks ahead. It also did better at spotting extreme heat and rainfall around major floods. The authors remain cautious: the gains are modest, a forecast can still be wrong, and the tool does not replace better models.
We do not control the weather. But a few more weeks of preparation can help protect harvests, power grids, and lives.