Microsoft's Probabilistic Bias Correction improves ECMWF forecasts for 98% of precipitation targets, runs on CPUs

Microsoft Research has released Probabilistic Bias Correction (PBC), an open-source machine learning tool that corrects systematic errors in existing weather models. According to the researchers, it improves forecast accuracy at lead times of two to six weeks. The findings were published October 14, 2026 in the journal Nature.

What PBC does

PBC is not a standalone weather model. It applies machine learning to identify systematic bias in an existing forecasting system, whether physics-based or AI-based, and then corrects for it.

Microsoft says both model types become less reliable beyond roughly two weeks. Both fill data gaps with approximations, for example in regions with sparse weather stations or for atmospheric phenomena that are not well understood. Those approximations create systematic bias, and the resulting errors compound as the forecast horizon extends.

Lester Mackey, senior principal researcher at Microsoft Research New England, gave one example: models built on historical records can skew toward colder temperatures than the climate is producing today, because greenhouse gas emissions are warming the planet.

Reported results

According to the study, PBC:

• Improved accuracy of a leading physics-based model from the European Centre for Medium-Range Weather Forecasts (ECMWF) for 98% of precipitation targets and 92% of temperature targets.
• Doubled the accuracy of ECMWF's new AI forecasting system in the study's analysis.

The ECMWF had already recognized the PBC research in its AI Weather Quest competition before publication. The paper's authors come from Microsoft, MIT, ECMWF, and other institutions.

Compute requirements

Mackey says the tool can run on CPUs alone. Physics-based forecasting typically requires supercomputers, and generative AI weather models depend on powerful GPUs. Parameter count, training compute, and the open-source license terms were not specified in the available reporting. No pricing applies, as the software is released as open source.

Early deployment

Genevieve Flaspohler, a paper author who worked on PBC as a Microsoft intern, is now executive director of the nonprofit Rhiza Research. With Gates Foundation funding, the group is working with the Kenya Meteorological Service Authority to test whether AI can improve early warnings. PBC is one of the methods under evaluation ahead of a rainy season expected to be more extreme during a record-breaking El Niño.

Matthew Chantry, ECMWF's strategic lead for machine learning, called the paper "a good example of new ideas being developed for the benefit and consideration of the community." He added a caveat: "turning ideas into reality [in] an operational process is a most challenging step."

Microsoft joins Google, Nvidia, and others developing AI tools for weather forecasting.

What this means

The notable choice here is architectural. Rather than compete with ECMWF's models or Google's and Nvidia's AI forecasters, PBC sits on top of them. That makes it complementary to whichever base model an agency already runs, and it avoids the cost of replacing operational infrastructure.

The CPU-only requirement matters for adoption. Weather agencies in developing countries are often priced out of GPU-heavy approaches, and the Kenya pilot is an early test of whether that gap can close.

Two caveats apply. The 98% and 92% figures come from the authors' own study, and Chantry's comment makes clear that operational deployment is a separate hurdle from published results. The sub-seasonal range of two to six weeks is also where forecasting skill is weakest, so even modest, reliable gains there would have practical value for agriculture and emergency planning. Independent replication on other base models and regions will show how well the results generalize.