Correcting the forecast, not rebuilding the atmosphere

A study published in Nature on October 14 reports a machine-learning layer that improves weather forecasts several weeks ahead. Called probabilistic bias correction, it learns from previous forecasts and the weather that followed them, then adjusts the probabilities produced by an existing system.

Researchers from several institutions, including Microsoft Research, MIT, the University of Toronto and ECMWF, developed the approach. A preprint first appeared in April and was revised in September. The journal publication and MIT's new account bring attention to existing research, rather than marking its first public disclosure.

What the model actually predicts

The authors' accessible preprint focuses its main comparisons on weekly temperature, precipitation and pressure. Its targets are days 19–25 and 26–32 after a forecast is issued, described as week three and week four. That is more specific than a general promise to predict weather six weeks ahead.

Instead of choosing one exact future, the system distributes probabilities across ranges derived from past weather. Two learning components correct systematic errors and incorporate recent observations and seasonal patterns. Their outputs are combined into a revised probability distribution.

In practical terms, the aim is a better estimate of how likely an unusually wet or warm week is. It is not a prediction of the exact hour a storm will reach a particular street.

Keep the baseline beside the gain

The authors report improvements for both physics-based and AI forecasting systems. For one AI system, their preprint reports a doubling of average ranked probability skill score from a modest baseline of 0.03 across the tested variables and horizons in 2025.

That score compares probabilistic forecasts with a historical climate baseline. Zero means no improvement over that baseline; positive values indicate better performance. Doubling a small positive skill score does not mean twice as many forecasts are correct, and it does not turn uncertainty into certainty.

The team also reports that its combined MicroDuet system led the September–November 2025 AI Weather Quest competition across the tested variables and lead times. This is a reported result for that competition and period, not a ranking of every forecast service for every use.

A useful research result, with work still to do

The authors have released code and instructions for generating corrected forecasts. The repository also describes data-access requirements. Code availability makes scrutiny easier, but this desk has not rerun the experiments or independently reproduced the reported scores.

The evidence reviewed here includes the authors' September preprint, their code repository, MIT's October 14 account and publisher-supplied publication metadata. The publisher's full-text page returned a browser challenge during direct checks; no audit of every change in the final journal version is claimed.

Better long-range probabilities could help planning. Whether they improve decisions in a particular region still requires local evaluation and a way to act on the information. This study offers a correction tool to investigate, not a replacement for official weather warnings.

Sources

  1. Authors: Advancing Subseasonal Forecasting with Machine LearningSeptember 4 preprint, first submitted April 17. Results sections 2.1-2.4 and discussion: targets, probabilistic correction, RPSS baseline and reported competition outcomes. Not an independent reproduction.
  2. MIT: subseasonal forecasting research accountOctober 14 institutional account of the newly published Nature study and intended planning applications. Precise test horizons and score interpretation are taken from the authors' methods, not an unqualified doubling-of-accuracy claim.
  3. Authors: probabilistic bias correction codePublic code and README link the method to the Nature DOI and describe implementation, reproduction instructions and data access. Code not executed for this report.
  4. Crossref: Nature publication metadataPublisher-deposited title, DOI and October 14 online publication date. Metadata verifies publication identity, not scientific validity.