Microsoft-led AI method doubles skill of 2-6 week weather forecasts, Nature study says

A team led by Microsoft Research has published a study in the journal Nature showing that a machine learning method can substantially improve weather forecasts for two to six weeks ahead, a range that has long been one of the hardest to predict. The paper, "Advancing subseasonal forecasting with machine learning", was published on Wednesday, October 14, 2026. It introduces a technique called probabilistic bias correction, or PBC, which learns from a forecasting model's past predictions to correct its systematic errors. Applied to forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF), the authors report that PBC doubled the subseasonal accuracy of ECMWF's AI forecasting system. The researchers have released the code as open source.

Weather forecasts are good for about a week and seasonal outlooks give broad trends for months ahead, but the gap in between, two to six weeks, has been a weak spot. At that range, small errors in a model grow quickly, models carry built-in biases, and the atmosphere itself becomes chaotic. Yet this is exactly the window many decisions depend on: when to sow or harvest, how much power and water to plan for, when to prepare for a heatwave or a cold spell, and how to position relief supplies before a flood.

PBC does not replace existing forecasting systems. Instead, it sits on top of them and learns, from years of their past forecasts and what actually happened, where they tend to go wrong. It then corrects new forecasts in the same way. Because it works on probabilities rather than single numbers, it can tell users not just what is most likely but how confident the forecast is. That makes it useful for planning, where knowing the odds matters as much as the headline prediction.

The result also shows how artificial intelligence is changing weather science. In recent years AI models have matched or beaten traditional systems for short-range forecasts. This study suggests that AI can also add value at longer ranges by improving both AI and physics-based models. Releasing the code openly means national weather services, researchers and companies can test it on their own forecasts.

For countries such as India, where farming, water supply and disaster preparedness depend heavily on rainfall and temperature, better forecasts a few weeks ahead could be especially valuable, if weather agencies adopt and validate such methods for local conditions.

Short-term: Weather agencies and researchers can now try the open-source method on their own forecasts. The study is likely to prompt comparisons with other AI forecasting tools in ongoing benchmarking efforts, such as ECMWF's AI Weather Quest challenge.

Long-term: If the gains hold up in daily use, more reliable two-to-six-week forecasts could help farmers, energy companies, insurers and disaster managers plan earlier and with more confidence.

Who is affected: National meteorological services, farmers and agricultural planners, power and water utilities, disaster management agencies, and AI researchers working on climate and weather.

Questions and Answers