Microsoft unveils open-source ML tool it says improves 2–6 week weather forecasts

Microsoft and academic partners released a machine-learning postprocessing method they say improves forecasts in the traditionally hard-to-predict two-to-six-week range. The technique, which Microsoft labels Probabilistic Bias Correction (PBC), is being published as open-source software and is intended to sit on top of existing physics-based models to correct systematic errors.

Microsoft researchers reported that applying PBC to a leading European physics-based forecast from the European Centre for Medium-Range Weather Forecasts (ECMWF) improved accuracy for 98 percent of precipitation targets and 92 percent of temperature targets.

The company characterized the approach as cost-effective and said it could help operational users such as farmers and emergency responders plan with greater certainty.

Researchers at MIT and partner institutions described a similar machine-learning postprocessing concept and said their system achieved the first forecasts to beat climatological averages in the subseasonal two-to-six-week window. MIT reporting said the model doubled accuracy in that range, a timeframe meteorologists have long called difficult for prediction.

Microsoft and MIT collaborators published a paper on the method with authors from multiple institutions, and Matthew Chantry, ECMWF’s strategic lead for machine learning, praised the paper as a positive development for the community while cautioning that moving research into operational forecasting is challenging.

Microsoft and one paper author, Genevieve Flaspohler, emphasized the practical advantages of the new approach: they said the models are lightweight, inexpensive to train and therefore accessible to scientists and organisations without large computational resources.

That accessibility was highlighted as one reason the teams are releasing the method as open-source software, so local forecasting groups and decision-makers can experiment and adapt the postprocessing layer to specific regional needs.

The reporting ties the potential operational benefit to concrete users: Microsoft and MIT-affiliated accounts named farmers, energy operators and emergency managers as groups that could use improved lead time to prepare for extreme events.

The work challenges a longstanding boundary in forecasting: meteorological theory and practice have treated roughly two weeks as a limit for deterministic predictability, a point the MIT reporting referenced when describing the ‘‘predictability desert’’ of subseasonal forecasting.

Both Microsoft and the academic collaborators framed the new method as a way to extend useful probabilistic information beyond that conventional horizon by correcting biases and errors in existing ensemble forecasts.

Sources differ in emphasis and detail: Microsoft’s account focuses on the PBC label, the ECMWF test gains for precipitation and temperature targets, and open-source release; MIT reporting emphasizes beating climatological averages and a reported doubling of accuracy in the two-to-six-week range.

Independent operational rollout and real-world testing remain open steps: ECMWF’s Matthew Chantry noted that operationalising research ideas is ‘‘a most challenging step,’’ and the reporting makes clear that improved metrics in retrospective or experimental tests do not automatically translate into immediate changes in routine forecasts.