Weather forecasts have become dependable for about two weeks ahead, but between two and six weeks out they still fall short, both for traditional models that simulate the atmosphere with physics equations and for newer AI models that learn from past weather. In a study published in Nature on Wednesday, Oct. 14, researchers from Microsoft and partner institutions describe a fix they call probabilistic bias correction. Rather than build another forecasting model, it studies an existing model's past forecasts next to the weather that actually happened, learns the mistakes that model tends to repeat, and corrects them in each new forecast.
Forecasts for that stretch guide when farmers plant and which seeds they buy, and help agencies plan for wildfires, water, energy and storms. Forecasts that far out give odds rather than a single answer, such as the chance that a week will be much wetter or colder than usual, and the method corrects those odds themselves, where a forecaster in Kenya who has tested it says most older fixes just adjust how much rain a forecast predicts.
The researchers tested it on models from the European Centre for Medium-Range Weather Forecasts, based in Reading, England. The paper's tests scored forecasts three and four weeks ahead. In tests on 2025 forecasts, the method doubled the modest long-range accuracy of the center's AI forecasting system. In tests on forecasts for 2016 to 2024, it also improved the center's main physics-based forecast, even after the center's own standard correction, for 98% of the rainfall targets they measured, 92% of temperature targets and 91% of air pressure targets. In the center's real-time forecasting contest, the AI Weather Quest, the researchers' entry, which applied the method to two forecasting models, gave the best global forecasts for every weather measure and time range in the contest's first three-month season, in late 2025, beating the forecasts of six government weather centers, the European center's own among them, and 34 teams from around the world. The paper says it was also the top model across the contest's first full year, and Microsoft says it flagged two severe U.S. winter storms in early 2026 between 19 and 26 days ahead.
The method costs little to run compared with producing the original forecasts, which takes supercomputers, and the researchers say that makes it practical for weather offices with little money. The researchers have released the code as open source, free for anyone to use, study, change and share. Microsoft says Kenya's weather service is trying it out this rainy season to spot the chance of very heavy or very light rain. A principal meteorologist there says that across the Greater Horn of Africa it improved their forecasts more than any other method they tested. Microsoft notes it is still early in the growing season.
The paper says the method should be seen as a way to clean up a model's output, not a replacement for better weather models. Several authors work for Microsoft, two work at the European center that ran the contest, and one advises Kenya's weather service. Long-range forecasts remain far less certain than next week's, and it is not yet clear whether national weather services will build the method into the forecasts the public sees.