Knowing tomorrow’s weather is useful. Knowing whether the coming month is likely to bring unusual heat, heavy rain, or cold could help communities prepare much earlier. New research highlights how artificial intelligence may improve forecasts in this difficult window by learning from the mistakes existing systems repeatedly make.
The approach, called probabilistic bias correction, focuses on forecasts looking two to six weeks ahead. Rather than replacing weather models, it adds a correction layer that learns from historical predictions and their errors. The goal is to improve the probabilities assigned to different possible weather outcomes.
The research team tested the method with both physics-based and AI forecasting systems from the European Centre for Medium-Range Weather Forecasts. In the reported evaluation, it doubled the AI system’s modest forecasting skill at these longer lead times. That does not mean every prediction became twice as accurate or that uncertainty disappeared.
The team also reported improvements across most evaluated temperature, precipitation, and atmospheric-pressure targets. Forecasts using the approach placed first across the evaluated variables and lead times in ECMWF’s 2025 real-time forecasting competition.
This period is often called the subseasonal range. It sits between familiar short-term forecasts and broader seasonal outlooks, creating a difficult challenge for researchers. Small errors grow as predictions extend further into the future, while systematic biases can repeatedly push a model’s output in the wrong direction.
The correction system examines those patterns of error and adjusts the forecasts. It works with probabilities rather than promising an exact description of the weather on a particular distant day.
MIT’s October 14 report also highlights improved predictions of extreme events. The researchers describe the method as an open-source, relatively low-cost addition to existing forecasting systems. Although supercomputing resources supported development, the resulting correction model does not require a supercomputer to run.
Potential applications include agricultural planning, energy management, and emergency preparation. The researchers report interest from Kenya’s meteorological office, but that interest should not be confused with confirmation of nationwide operational deployment.For decision-makers, a better probability estimate can be valuable even when certainty remains impossible. An energy operator might use an increased risk of cold weather to review supplies. A farmer could weigh a rainfall outlook alongside local conditions before scheduling fieldwork.
Those decisions still require judgment. Forecast performance must be assessed across locations, seasons, and the specific events users care about. A strong competition result is encouraging evidence, while dependable everyday service requires continued evaluation.
The research shows a practical role for AI: improving existing scientific tools by identifying recurring errors. Its promise lies in giving people more useful information and potentially more time to prepare.
