Weather forecasting has become remarkably dependable over short horizons: today's models are highly accurate out to about two weeks, and long-range climate projections are steadily improving as well. Between those two time scales, however, sits a stubborn blind spot. Forecasts in the two-to-six-week window — known as sub-seasonal forecasting — remain unreliable, even though that is precisely the range that matters for deciding when to plant crops, how to allocate water and energy, and how to prepare for wildfires.
Microsoft Research and a network of external collaborators say they have found a way to narrow that gap, not by building yet another forecasting model, but by teaching software to correct the predictable mistakes of the models that already exist. The approach, called Probabilistic Bias Correction (PBC), is detailed in a paper published in the scientific journal Nature and described in a Microsoft blog post published this week. It is one of the more quietly practical entries in the wave of applied AI research covered among the latest AI developments, because it upgrades the systems meteorologists already rely on rather than asking them to switch to something new.
Why the Two-to-Six-Week Window Defies Forecasting
Physics-based models simulate the atmosphere and oceans using scientific equations, while newer AI models train on vast archives of observed weather data. According to Microsoft, both families perform well out to roughly ten days. Beyond that point, small initial errors compound and the complex interactions between the atmosphere, oceans, and land surface overwhelm both approaches, which is why two-to-six-week outlooks have long lagged the rest of forecasting.
Rather than enter a crowded field of competing models, the Microsoft-led team focused on a different target: the systematic, repeatable errors that models make in this range. "Our new pipeline (PBC) takes your favorite model as input and learns in what ways it tends to make errors. It then corrects those errors, yielding a better output forecast for the sub-seasonal range," the company explained in its blog post.
A Lightweight Layer on Top of Existing Models
Technically, PBC is a lightweight machine learning system that studies any given model's archive of historical forecasts to learn where and how it is systematically biased. Once those patterns are learned, the system adjusts the model's probabilistic output to compensate. In evaluations described in the Nature paper, Microsoft reports that PBC improved the performance of the world's leading operationally debiased sub-seasonal model, operated by the European Centre for Medium-Range Weather Forecasts (ECMWF).
The results held up in live competition, not just in retrospective tests. ECMWF launched the AI Weather Quest in August 2025, a real-time competition designed to advance probabilistic sub-seasonal forecasting. A team called MicroEnsemble, made up of Microsoft researchers and external collaborators, took top honors with an entry dubbed MicroDuet, which applied PBC to ECMWF's industry-leading dynamical model.
"The AI Weather Quest provides a unique real-time benchmarking platform for sub-seasonal forecasts," said Dr. Joshua Talib, a scientist in sub-seasonal forecasting at ECMWF. "Contributions from Team MicroEnsemble spotlighted how machine learning can enhance the skill of traditional dynamical forecasting systems through post-processing."
From Benchmark to Real-World Pilots in East Africa
The technique is already moving beyond benchmarks. In Kenya, forecasters are preparing to test PBC as one of the country's two rainy seasons gets underway. The effort involves the Kenya Meteorological Service Authority (KMSA) and scientists at the Climate Hazards Center at the University of California, Santa Barbara, with advice from Genevieve Flaspohler, who runs Rhiza Research, a nonprofit focused on improving AI-powered weather forecasting.
Flaspohler argues that the value of a forecast depends on arriving in time to act on it. "A good weather forecast, delivered at the right time, on the right communication channel, can be a big differentiator for agricultural outcomes," she said, noting that East Africa has been oscillating between the driest years on record and the wettest years ever. In a region where farms largely lack irrigation and planting, fertilizing, and seed-purchase decisions all ride on the outlook, a more skillful sub-seasonal forecast is not an abstract improvement.
"Because we are in the midst of the warmest El Niño ever on record, there is a lot of fear among governments and in the climate science community about how that will manifest," Flaspohler said. "There is a huge need for simple forecasting methods that work robustly."
An Efficiency Argument in an Era of Bigger Models
What makes PBC notable within AI research is its economy. While much of the field races to train ever-larger end-to-end forecasting models, PBC is a comparatively small system that borrows most of its skill from the operational models it corrects. That makes it cheaper to run and easier for meteorological agencies to adopt, since they can keep their existing infrastructure and simply add a correction layer.
It also reflects a broader pattern in applied machine learning: post-processing and calibration often deliver gains that would otherwise require a full model rework. For weather services weighing where to spend limited budgets, a method that demonstrably improves the hardest forecast range — validated both in Nature and in a live ECMWF competition — is likely to draw attention.
What Comes Next for Sub-Seasonal AI
Microsoft's blog post stops short of announcing an operational product, and the company has not said whether PBC will be integrated into commercial forecasting offerings. The near-term test is whether national weather services begin adopting bias-correction layers like PBC for their sub-seasonal outlooks, and whether the approach composes well with the AI forecasting models now entering operational use.
If it does, the two-to-six-week window — the stretch where farmers, grid operators, and disaster planners have historically had the least to work with — may finally start closing. For a field that has spent decades chasing marginal gains in skill, letting a small machine learning system clean up the systematic errors of giant physical models is a pragmatic kind of progress.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
