Microsoft Research, in collaboration with external partners including MIT and the European Centre for Medium-Range Weather Forecasts (ECMWF), has developed a machine learning framework called Probabilistic Bias Correction (PBC) to enhance the reliability of sub-seasonal weather forecasts. This new tool, detailed in a paper published in the journal Nature on October 14, 2026, addresses a persistent challenge in weather prediction: the decline in accuracy for forecasts extending beyond two weeks. PBC operates by identifying and correcting systematic biases in existing weather models, rather than creating an entirely new forecasting system. The software for PBC is now available as open-source.

Both traditional physics-based models and newer AI-driven weather models typically exhibit reduced accuracy in the two-to-six-week timeframe. This limitation stems from compounding errors, inherent model biases, and the complex, chaotic nature of atmospheric interactions. PBC aims to mitigate these issues by learning from a model's past performance and adjusting its probabilistic forecasts. Lester Mackey, a senior principal researcher at Microsoft Research New England, noted that PBC can run on CPUs, eliminating the need for high-performance GPUs or supercomputers often required by other advanced forecasting systems. This reduced computational demand could make more accurate long-range forecasting accessible to a wider range of users, including regions with less advanced weather observation infrastructure.

In evaluations, PBC demonstrated significant improvements when applied to the ECMWF's leading physics-based model. It boosted performance for 98% of precipitation targets, 92% of temperature targets, and 91% of pressure targets. The framework also doubled the accuracy of ECMWF's AI Forecasting System (AIFS). Furthermore, a PBC-based solution, dubbed MicroDuet by a team from Microsoft and its collaborators, secured top honors in the ECMWF's AI Weather Quest, a real-time competition focused on improving probabilistic sub-seasonal forecasting.

The practical application of PBC is already being explored. In Kenya, the nonprofit Rhiza Research is testing PBC with local weather authorities, supported by the Gates Foundation, to enhance agricultural rainfall forecasts and early warnings for extreme weather events. Dr. Phillip Okello, a principal meteorologist at the Kenya Meteorological Department (KMSA), observed that PBC improved monsoon forecasts more than other methods tested in the Greater Horn of Africa, particularly by correcting probabilities themselves rather than just adjusting rainfall amounts. Microsoft Weather, the team behind AI weather forecasts for various Microsoft products, is also integrating PBC into its sub-seasonal-to-seasonal (S2S) model, which provides weekly probabilistic guidance up to six weeks ahead.

The enhanced accuracy in the two-to-six-week window has implications for sectors such as agriculture, disaster response, and energy management, allowing for more informed decision-making regarding planting and harvesting, wildfire management, and resource allocation. Microsoft continues to investigate further applications and extensions of PBC, including identifying which weather models could benefit most from its machine learning corrections.