The Pulse
Microsoft Team Doubles ECMWF AI Weather Skill in Weeks 3 and 4
A Microsoft-led team applied probabilistic bias correction to ECMWF’s AIFS-SUBS and doubled its average forecast skill for weeks three and four. The researchers also report gains for ECMWF’s dynamical forecasts and first-place results in th
AI.info Team · Written with AI ·
A Microsoft-led team reports that a statistical correction layer doubled the average forecast skill of the European Centre for Medium-Range Weather Forecasts’ AI system at weeks three and four. The result, in a paper first posted on April 17, concerns forecasts for days 19–32, a range where errors can overwhelm weather models. The researchers call their method probabilistic bias correction, or PBC.
How PBC targets weeks three and four
Rather than generate weather predictions from scratch, PBC adjusts probability forecasts from existing models. It learns systematic errors from past forecasts and observations, then combines two corrections: one that adjusts forecast probabilities by location, date and quantile, and another that weighs recent observations alongside climatology.
The team tested the approach on the ECMWF’s AIFS-SUBS AI system, its dynamical ensemble model and a hybrid system called PoET. The evaluation focused on average two-meter temperature and total precipitation over land, plus mean sea-level pressure. For weeks three and four, the researchers scored forecasts against a climatological baseline using the ranked probability skill score, or RPSS.
The score behind the doubling
For AIFS-SUBS, PBC doubled the average RPSS across the tested variables and lead times, bringing the mean score to 0.03 in the 2025 evaluation. The paper presents that as a gain in forecast skill, not as a claim that the AI system predicts weather twice as accurately in every location or for every variable.
On the ECMWF dynamical model, the researchers evaluated forecasts from 2016 through 2024. Compared with the model’s operationally debiased forecasts, PBC improved scores in 98% of evaluated grid cells for precipitation, 91–92% for temperature and 89–91% for mean sea-level pressure. The paper also reports that PBC turned the raw model’s negative temperature and pressure skill scores into positive ones against climatology.
A competition result, not a new weather model
The research team also used the correction method in MicroDuet, its entry in the ECMWF’s AI Weather Quest. The paper says MicroDuet placed first for every weather variable and lead time in the September–November 2025 contest period, ahead of six forecasting centers, ECMWF’s AI forecasts and systems from 34 teams worldwide.
The authors describe the method’s role plainly: “PBC targets predictable errors rather than attempting to relearn full atmospheric dynamics.” Hannah Guan et al., “Enhancing AI and Dynamical Subseasonal Forecasts with Probabilistic Bias Correction”
What the results do—and do not—show
The results support a practical distinction: improving a forecast can mean correcting a model’s recurring errors, not replacing the model that produces the forecast. The researchers say PBC is inexpensive compared with generating large dynamical or AI ensembles, and can be retrained as the underlying models change.
The study tests specific forecast systems, variables, lead times and evaluation periods; it does not establish that the same gains will hold for every model or operational setting. PBC depends on errors that recur in historical data, and the paper frames it as a post-processing tool alongside continued work to improve the underlying weather models.