A tool called Probabilistic Bias Correction (PBC), developed by a team from MIT, Microsoft Research, Harvard and ECMWF, doubled the skill of weather forecasts 2 to 6 weeks out and took first place in the international AI Weather Quest competition. Instead of building another weather model, the researchers trained an algorithm to catch and correct systematic errors in existing ones, including the best systems of the European Centre for Medium-Range Weather Forecasts (ECMWF).
- Probabilistic Bias Correction (PBC), built by researchers from MIT, Microsoft Research, Harvard and ECMWF, is a postprocessing layer that corrects systematic errors in existing weather models and, according to the Nature paper, doubled the skill of 2-6 week forecasts.
- Applied to ECMWF's AIFS model, PBC doubled its average subseasonal skill from 0.03 on the competition metric, and applied to ECMWF's IFS model it improved 91% of pressure targets, 92% of temperature targets and 98% of precipitation targets.
- Microsoft Research and MSN Weather's MicroDuet system, which combines PBC with an ECMWF model and the PoET transformer model, placed first in ECMWF's AI Weather Quest for every variable and horizon in the September-November 2025 season and was again first for both horizons in December-February.
The results are described in “Advancing subseasonal forecasting with machine learning,” published in Nature. MIT publicized them in a release dated October 14, 2026.
Keep exploring artificial intelligence
Read the latest news and practical guides on AIPORT.
PBC does not forecast weather, it corrects other models’ forecasts
Probabilistic Bias Correction is a postprocessing layer. It takes a finished forecast from any model and shifts its probabilities based on how that model has been wrong in the past. Its inputs are recent observations, climatological quantiles for a given location, and a deterministic forecast from a physics-based or AI model. The output is a corrected probabilistic forecast. It does not say “it will be 12 degrees.” It gives the probability that temperature will fall into one of five bins, from well below normal to well above.
Inside, two lightweight machine learning models, Debias++ and Persistence++, run in parallel and generate complementary corrections. Judah Cohen, a climatologist and research scientist in MIT’s Department of Civil and Environmental Engineering, says in the MIT release that the idea came from Lester Mackey of Microsoft Research, with whom he has worked for more than a decade.
The model was trained on MIT’s Engaging supercomputing cluster, but the finished tool does not require a supercomputer. It is open source and cheap to run.
PBC doubled the skill of ECMWF’s AI model and improved 98% of precipitation targets
The researchers tested PBC on two flagship systems from ECMWF, regarded as the best forecasting center in the world:
- AIFS (AI Forecasting System), ECMWF’s machine learning model: PBC doubled its average subseasonal skill, from 0.03 on the competition metric, averaged across all variables and horizons.
- ECMWF’s dynamical model (IFS), already operationally debiased by ECMWF itself: PBC improved forecasts for 91% of pressure targets, 92% of temperature targets and 98% of precipitation targets.
- Weaker models from other centers: with PBC applied, they approached the level of ECMWF’s forecasts and in some cases matched them.
According to MIT, this was the first time a model showed a sustained advantage over climatological averages at the 2-6 week horizon. Until now, Cohen says, most forecasts for that window were worse than a simple long-term average.
Cohen said:
“We consistently beat climatology averages with our models.”
The window matters because meteorologists call it the “predictability desert.” Short-range forecasts benefit from precise initial conditions, and seasonal forecasts benefit from slow boundary factors such as El NiΓ±o. The 2-6 week range gets full help from neither. Edward Lorenz, the creator of chaos theory and a longtime MIT professor, set two weeks as the theoretical limit of weather predictability in the 1970s. Cohen says in the release that he never fully believed in that limit.
Tell our editorial team about your project. Include a short description, the relevant links and your contact information.
A PBC-based system took first place in ECMWF’s AI Weather Quest
ECMWF launched the year-long AI Weather Quest in August 2025. Participants submit probabilistic global forecasts every week for weeks 3 and 4. The MicroEnsemble team (Microsoft Research together with the MSN Weather team) entered a system called MicroDuet. It combines PBC applied to an ECMWF model with PBC applied to PoET, a transformer model that improves ensemble forecasts.
In the September-November 2025 season, MicroDuet placed first for every variable (temperature, sea level pressure, precipitation) and every horizon. It beat the debiased dynamical models of six government agencies: ECMWF, NOAA, Canada’s ECCC, Russia’s HMCR, Korea’s KMA and Japan’s JMA. It also beat their joint multimodel, three of ECMWF’s own AI models and the systems of 34 teams worldwide. In the December-February season, MicroDuet was again first for both horizons. It was the only entry among 37 teams to beat the multimodel dynamical ensemble on all variables.
The researchers disclosed a conflict of interest themselves. Two coauthors, Jakob Schloer and Joshua Talib, work at ECMWF, which organizes the competition and develops AIFS.
Analysis
This result says something about where the advantage in applied AI lies today. I read it as not in the next, bigger model but in a cheap tool that learns from the mistakes of existing ones. MIT and Microsoft did not outrun ECMWF. They took its model and made it twice as good. Any organization sitting on piles of historical forecasts, decisions or valuations should pay attention, because there are systematic errors nobody has counted there too.
I have two questions. First, doubling skill from 0.03 still leaves a forecast that is only slightly better than a coin flip. Would a farmer or a grid operator make a decision worth millions on so thin an edge?
Second, PBC learns from a model’s historical errors. What happens when the climate changes faster than historical data can reflect? The authors say the model adapts its corrections as new data arrives. I would like to see how that works in practice over several more seasons, not just one competition.
PBC improved forecasts of extreme temperatures and rainfall, where conventional correction struggled
For emergency management, the most important result concerns extreme events. The researchers recorded a statistically significant improvement in predicting extreme temperatures and precipitation after applying PBC.
Cohen points out that ECMWF’s own non-AI error correction made forecasts of extreme events worse. His interpretation is that classic bias correction always pulls the result toward the mean, while an extreme by definition sits at the edge of the distribution. In an interview with Gizmodo, he added that the tool does not depend on the type of model:
“It’s model agnostic.”
MIT’s release opens with the February 2021 disaster in Texas, where three winter storms led to the largest energy infrastructure failure in the state’s history. At least 246 people died. MIT argues that a two-week lead time would have given operators time to prepare gas supplies, which a two-day warning does not. Cohen says Kenya’s meteorological service has already approached the team about deploying PBC to forecast the monsoon season.
The work was funded in part by the US National Science Foundation, Brazil’s Carlos Chagas Filho Foundation and the European Union. The lead author is Hannah Guan, a Harvard College student. Coauthors include Soukayna Mouatadid of the University of Toronto, Paulo Orenstein of Brazil’s IMPA, Jeremy Berman of DTN, and Alex Lu, Haiyu Dong, Jonathan Weyn, Zekun Ni and Lester Mackey of Microsoft.
Sources and methodology
- Nature, “Advancing subseasonal forecasting with machine learning”: https://www.nature.com/articles/s41586-026-11126-x
- MIT News press release, October 14, 2026: https://news.mit.edu/2026/researchers-improve-weather-forecasts-looking-two-six-weeks-out-1014
- arXiv preprint: https://arxiv.org/html/2604.16238
- Microsoft Research project page: https://www.microsoft.com/en-us/research/project/subseasonal-climate-forecasting/
- Gizmodo (interview with Judah Cohen): https://gizmodo.com/forget-the-10-day-forecast-ai-just-improved-extreme-weather-warnings-up-to-6-weeks-2000825560
Quotes from Judah Cohen, a research scientist at MIT, were verified against the original English-language source material.
Analysis and editorial responsibility: Piotr Wolniewicz, Editor-in-Chief, AIPORT. Original Polish reporting: Piotr Wolniewicz (aiport.pl).
This article was originally published in Polish on aiport.pl on October 20, 2026.
https://aiport.pl/co-nowego-w-ai/mit-i-microsoft-podwoily-trafnosc-prognoz-pogody-na-2-6-tygodni-narzedzie-pbc-wygralo-konkurs-ecmwf/
