The Pulse
Microsoft Uses Bias Correction to Improve 2-to-6-Week Forecasts
Microsoft Research says its Probabilistic Bias Correction system improves extended-range forecasts of temperature, precipitation and pressure. The approach adjusts existing weather models’ probability estimates and is being tested with Keny
AI.info Team · Written with AI ·
“Our new pipeline (PBC) takes your favorite model as input and learns in what ways it tends to make errors. It then corrects those errors, yielding a better output forecast for the sub-seasonal range.”
Lester Mackey, senior principal researcher, Microsoft Research New England
Microsoft Research says its Probabilistic Bias Correction (PBC) framework improves weather forecasts for the difficult two-to-six-week window by adjusting existing models’ errors rather than replacing them with a new forecasting system. The company published the results on October 14, 2026, alongside details of early testing and a pilot with Kenya’s meteorological service.
Correcting forecasts, not replacing models
Weather models tend to lose skill beyond roughly two weeks, when small differences in atmospheric conditions can lead to widely different outcomes. PBC targets a more specific problem: persistent errors in a model’s probability estimates. It learns from past forecasts and observations, then adjusts the probabilities while retaining useful signals from the original model.
The approach combines two corrections. One learns location-, date- and probability-specific adjustments from previous forecast errors; the other blends the forecast with recent observations and historical climate patterns. The resulting probabilities are constrained to form a valid forecast distribution. That focus matters because forecasts several weeks out describe chances of broad conditions, not a reliable prediction of the weather on a particular day.
The reported gains are measured against climate baselines
The research team tested PBC on forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF), including its dynamical model and AI Forecasting System. In the paper, the researchers report that PBC improved the operationally corrected ECMWF dynamical forecast across 98% of precipitation targets, 92% of temperature targets and 91% of pressure targets. They also report that PBC doubled the AI system’s average subseasonal skill.
Those percentages describe the share of forecast targets where skill improved; they are not claims that forecasts became that many percentage points more accurate. The study evaluates weekly averages for temperature, precipitation and sea-level pressure in weeks three and four, against a climatological baseline. The authors also test against historical forecasts from 2016 through 2024, while the AI-system comparison uses 2025 data.
A PBC-based entry called MicroDuet, built by the MicroEnsemble team, ranked first across the weather variables and lead times in the opening season of ECMWF’s AI Weather Quest. Microsoft says the entry outperformed systems from six operational forecast centers and 34 competing teams. The researchers also report that MicroDuet anticipated two U.S. winter storms 19 to 26 days ahead, though such case studies do not by themselves establish how often the system will outperform other forecasts in future conditions.
Kenya’s forecast pilot puts rainfall probabilities to work
Microsoft describes an early deployment with the Kenya Meteorological Service Authority, where forecasters are testing PBC for weekly rainfall outlooks. The forecasts aim to help identify chances of unusually heavy or light rain during a rainy season. For smallholder farmers who rely on rainfall, that information can inform choices about seeds, planting and fertilizer.
“Most bias correction just adjusts how much rain a forecast predicts. PBC corrects the probabilities themselves, and across the Greater Horn of Africa it improved our forecasts more than any other method we tested,” said Phillip Okello, principal meteorologist at the Kenya Meteorological Service Authority. He also identified a remaining task: sharpening forecasts of heavy rain and dry spells, the conditions that matter most to farmers.
Open code, limited customer access
Microsoft says PBC is available as open-source software, while its separate sub-seasonal-to-seasonal model offers weekly probabilistic guidance up to six weeks ahead. That model runs in real time and is in private preview for selected customers. The distinction is important: PBC is a correction framework that can be applied to existing models, while Microsoft’s S2S system is a forecasting product.
The results make a case for post-processing as a practical way to improve longer-range forecasts, but the method addresses predictable, systematic errors rather than the uncertainty that no correction can remove. Its value will depend on whether the gains hold across regions, seasons and operational settings—and whether forecasters can turn probability improvements into guidance people can use. For now, the evidence includes competition rankings, retrospective tests and an early Kenya pilot; the framework’s code is publicly available for researchers to examine.