Microsoft and MIT researchers report gains in forecasts two to six weeks ahead

Microsoft and MIT researchers have reported improved weather forecasts two to six weeks ahead using a machine-learning technique that corrects existing predictions. Their study was published in Nature on October 14, 2026. The reported gains concern a forecasting window that matters for planning but has long been difficult for weather models to predict reliably.

The method, called probabilistic bias correction, or PBC, works with forecasts produced by other systems. It learns from historical probabilistic forecasts and observations of what weather actually occurred, then adjusts the distribution of future predictions. It is therefore a way to improve an input model’s output, rather than a separate weather model replacing that input.

Why forecasts two to six weeks ahead are difficult

Nature describes the two-to-six-week horizon as subseasonal forecasting. At that range, errors compound, systematic biases persist and atmospheric chaos limits predictability. Those problems make it harder to turn a forecast into a useful estimate of conditions weeks ahead, even when shorter-range forecasts perform well. Correcting a model’s recurring errors offers one route to better predictions without claiming to remove that underlying uncertainty.

The paper says PBC can be used with any input model. That matters because forecasting centres already use different approaches, including physics-based systems and newer AI systems. An adjustment that can sit on top of those forecasts could be applied across more than one modelling approach, although the published results establish performance in the settings studied rather than operational use everywhere.

What the study reports about forecast accuracy

The Verge also reports that the analysis found the method doubled the accuracy of ECMWF’s new AI forecasting system. That comparison is a reported study result; it should be read alongside the particular model and evaluation used, rather than as a claim that every AI weather forecast can be made twice as accurate. The approach depends on the forecasts it receives as inputs.

Nature says the system outperformed the best alternative dynamical, AI and hybrid systems in a real-time forecasting competition. MIT News reports that the team placed first in an international competition. MIT also says the researchers’ paper found forecasts that beat climatological averages between two and six weeks ahead. Together, these are evidence of performance in the researchers’ evaluations and competition, rather than measured benefits to people using the forecasts.

Potential uses for farmers and emergency planners

More useful information weeks ahead could help decisions that cannot wait until a short-range forecast is available. MIT News identifies farmers, energy operators and emergency planners as potential users. The Nature paper similarly points to agricultural planning, energy management and disaster preparedness, including the prospect of better predictions of extreme events. These are proposed applications of the method, not documented outcomes for those groups.

The Verge reports that Microsoft researcher Lester Mackey said the tool can run on CPUs. That could matter to forecasting agencies with limited computing resources: an improvement that requires less specialised hardware may be easier for them to explore. The report describes the software as open source, while the paper presents its forecasts as designed for operational deployment. Neither point establishes that agencies have already adopted it.

The collaboration also matters to how the development is described. MIT News identifies MIT researchers on the team, while The Verge focuses on the Microsoft researchers. The reported advance is the team’s published bias-correction method and its evaluated results, rather than a claim that one company alone has solved subseasonal forecasting. Its practical value will depend on how those gains translate when forecasting organisations use it.