A machine-learning method designed to correct predictable errors in weather models has improved forecasts for weeks three and four, a period when reliability usually falls sharply. In a peer-reviewed study published on 14 October, researchers tested probabilistic bias correction (PBC) against global forecast records and in a real-time international competition.
The approach does not replace the underlying weather model. Instead, it learns how that model has performed in the past and adjusts the probabilities it assigns to possible conditions. The researchers describe it as a post-processing layer that can be applied to both physics-based and artificial-intelligence forecasting systems.
Correcting probabilities, not rebuilding the forecast
Weather models can become less reliable at lead times of two to six weeks as small errors accumulate and systematic biases distort their output. PBC is intended to address the latter problem. It uses historical forecasts and observations to learn where a model tends to be wrong, then corrects its probability distributions for temperature, precipitation and atmospheric pressure.
The study compared PBC with several existing systems and correction methods. In tests spanning multiple years, it improved the operationally bias-corrected global model for 91 per cent of pressure targets, 92 per cent of temperature targets and 98 per cent of precipitation targets. The authors also report that PBC doubled the modest skill of a leading artificial-intelligence system for subseasonal forecasts.
These percentages describe the share of evaluated targets where the method improved forecast skill; they do not mean that forecasts were correct 91 to 98 per cent of the time. Forecast skill was assessed against a climatological baseline, and performance varied by weather variable, location and forecast lead.
A separate evaluation used results from a 2025 real-time forecasting competition. The team’s combined system ranked first for all weather variables and lead times in that competition, outperforming six operational forecasting centres, an international multi-model ensemble and entries from 34 teams. Competition results are a test under defined conditions, not proof that every local forecast will improve.
Extreme-weather signals and practical limits
The researchers also examined forecasts of unusually high or low temperatures, heavy rainfall and recorded flood events. Their analysis found improved skill for extreme events compared with the underlying model, including better flood-forecast skill in the event catalogue. Such results matter because a useful warning several weeks ahead could give planners more time to consider preparations, but the study does not establish that PBC has already prevented damage or improved outcomes in a particular disaster.
The framework is designed as a relatively low-cost addition to existing systems. The paper reports that its corrections can be retrained as the input models improve, rather than requiring a forecasting centre to build a new global model from scratch. The researchers have made the software available as open-source code.
There are important boundaries to the findings. The method improves probabilistic guidance over weeks three and four; it does not provide a precise day-by-day forecast six weeks in advance. Forecast skill is not uniform everywhere, and historical patterns may not perfectly represent future conditions. Operational users would still need to assess local performance and uncertainty before relying on the output for decisions.
One early application is being tested with meteorologists in Kenya for weekly rainfall guidance. That work is at a pilot stage. The research suggests PBC could support agricultural, energy and emergency planning where subseasonal forecasts are relevant, but it does not show that the method has been adopted as a routine public-warning service.
By targeting recurring errors rather than attempting to simulate the atmosphere anew, PBC offers a way to extract more information from existing forecast systems. Its broader value will depend on continued testing across regions, weather variables and real operational settings.
