Microsoft Research and partners published a machine learning method called probabilistic bias correction, or PBC, in Nature on Oct. 14. It targets subseasonal forecasts, which cover 2 to 6 weeks ahead. PBC does not replace a weather model. It learns the systematic errors in an existing model's past forecasts and corrects them.

What the paper and the code confirm

Nature's paper reports these results, which are the authors' own and peer reviewed. ECMWF is the European Centre for Medium-Range Weather Forecasts.

  • PBC doubles the subseasonal skill of ECMWF's AI Forecasting System.
  • It improves ECMWF's operationally debiased dynamical model for 91% of pressure, 92% of temperature and 98% of precipitation targets.
  • In ECMWF's 2025 real-time AI Weather Quest, its global forecasts were the best for every weather variable and lead time.
  • That beat the dynamical models of six operational forecasting centres and the systems of 34 teams.

The GitHub repository carries an MIT License, copyright 2026 Soukayna Mouatadid, who is a paper author. Nature's code statement also points to a Zenodo archive. The repository holds code and setup files, with no model-weight files in its file list. We did not open the Zenodo archive.

The README describes the setup:

  • Data: ERA5 reanalysis through the Copernicus climate data store, and ECMWF forecast data through an IRI account or ECMWF's servers. Both need sign-up and have their own terms.
  • Tested on: Rocky Linux 8.10 and Python 3.13.3, on an 8-core machine with 80 GB of RAM.
  • Demo: one script generates climatology forecasts and, per the README, finished in 13 seconds on that machine.

What is claimed, not confirmed

Microsoft's own research story says PBC could help deliver more skillful 2-to-6-week forecasts. It describes the method as lightweight and says it is "showing promise" in real-world use, notably for monsoon forecasts for farmers in Kenya. That is Microsoft's account and its own framing.

The Verge reports that Microsoft claims to have solved the loss of reliability beyond two weeks, and that Microsoft's Lester Mackey says the tool can run on CPUs alone. Microsoft's story does not repeat the CPU claim. The README's tested setup lists an 8-core, 80 GB Linux machine and no GPU.

Both pieces say co-author Genevieve Flaspohler of Rhiza Research is piloting PBC with the Kenya Meteorological Service Authority. The Verge adds Gates Foundation funding. The paper's abstract does not mention a Kenya deployment, so this is claimed, not confirmed.

ECMWF's Matthew Chantry told The Verge the paper is a good example of new ideas for the community. He added that turning ideas into an operational process is "a most challenging step."

Why it matters

Forecasts beyond two weeks drive planting, water, energy and disaster planning, and skill drops sharply there.

Microsoft's story calls PBC lightweight, and co-author Flaspohler, quoted by The Verge, calls it lightweight and easy to train. If that holds, it could be a cheaper way for agencies to try than a new global model, but no cost comparison appears in the paper abstract or README. The open code lets agencies test it on their own data.

Who this is for

PBC is a research tool for forecasters and data scientists. It is not an app or a consumer weather feature. Nothing in the paper or README says it will appear in a Microsoft product or in ECMWF's operational forecasts.

Conflicts to know

Nature's competing-interests statement says several authors own Microsoft stock and some are Microsoft employees. The paper also notes that its first author did the work as a Microsoft Research intern.