A team led by Microsoft Research published a method in Nature on October 14, 2026 that sharpens weather forecasts for the hard-to-predict window of two to six weeks ahead. Called probabilistic bias correction, or PBC, it does not replace existing forecast systems. Instead it studies how a given model has erred in the past and adjusts its new probability forecasts accordingly. The code is open source on GitHub and Zenodo, and the authors include researchers from Microsoft, MIT and ECMWF, Europe's intergovernmental centre for medium-range forecasting.
The subseasonal range matters for farmers deciding when to plant, for energy and water planners, and for agencies preparing for floods or heat. Physics-based and AI models are now reliable out to about two weeks, but beyond that their errors compound and systematic biases take over. The paper argues that part of this error is predictable. PBC trains on a model's own historical forecasts and on past observations, using the ERA5 reanalysis as ground truth, and combines two learned corrections: one that recalibrates the forecast distribution directly and one that also draws on recent observed weather. The two are averaged into a single output.
The researchers tested PBC on three leading systems for weeks three and four ahead: ECMWF's operational dynamical ensemble, its subseasonal AI Forecasting System (AIFS-SUBS) and POET, a hybrid that post-processes the ECMWF ensemble with neural networks. Scored by ranked probability skill against climatology, PBC doubled the AI system's average 2025 skill from a low base of about 0.03. Applied to ECMWF's already debiased dynamical model, it improved skill for 91% of pressure targets, 92% of temperature targets and 98% of precipitation targets. Even POET gained 68% to 98% for precipitation, 15% to 23% for sea-level pressure and 5% to 7% for temperature in 2024.
The strongest evidence comes from a live contest. In ECMWF's 2025 AI Weather Quest, where teams submit global forecasts every week before the weather happens, a PBC-based entry called MicroDuet produced the best forecasts for every variable and lead time. It beat dynamical models from six operational centres, an international multi-model ensemble, ECMWF's own AI system and the systems of 34 teams worldwide.
Cost is a central selling point. Because the model is small, the team retrains it every day using only data that would have been available at the time, to mimic operational use. Lester Mackey of Microsoft Research New England told The Verge the tool can run on ordinary CPUs rather than supercomputers or GPU clusters. Co-author Genevieve Flaspohler, now at the nonprofit Rhiza Research, said its Gates Foundation-funded work with Kenya's national weather service is testing PBC ahead of a rainy season expected to be severe.
An outside view came from ECMWF itself. Matthew Chantry, the centre's strategic lead for machine learning, told The Verge the paper is a good example of ideas developed for the community, while cautioning that moving any method into an operational process is the most difficult step.
Several limits apply. The absolute skill numbers at three to four weeks remain small, so doubling a weak signal still leaves a modest forecast. The contest results cover the seasons the paper reports, and no weather agency has announced operational adoption. Several authors are Microsoft employees or shareholders, as the paper's competing-interests statement discloses.