A machine learning method that corrects the repeated mistakes of existing weather models has substantially improved forecasts three to four weeks ahead, a stretch of time meteorologists have long treated as a blind spot. The research, published in Nature in October, reports that the technique doubled the skill of a leading AI forecasting system and improved Europe's flagship physics-based model across most of the globe.

The gains are modest in absolute terms, but they land in the window that matters most for decisions that are costly to reverse. Farmers deciding when to plant, utilities preparing for a cold snap and emergency managers staging supplies all work on that timeline.

Readers should know who built it. The 13-author team includes researchers from Microsoft Research and Microsoft, and the paper's competing-interests statement discloses that several authors are Microsoft employees. Some also hold company stock, according to Earth.com.

Inside Forecasting's "Predictability Desert"

Weather models are highly accurate out to roughly two weeks. Beyond that, small errors compound, and by the third week they can overwhelm the useful signal. Researchers call the period from about two to six weeks the "predictability desert," where many models fail to beat the simple long-term average for a given date.

That gap has real consequences. Earth.com pointed to the February 2021 Texas winter storms, which caused the state's worst energy failure and killed at least 246 people, as the kind of event where earlier warning in weeks 3 and 4 could matter.

Part of the challenge is timing. Short-range forecasts lean on the current state of the atmosphere, while seasonal outlooks lean on slow-moving drivers such as ocean temperatures and soil moisture. Weeks 3 and 4 fall between the two — too far out for the first to hold up, and too soon for the second to dominate.

How Learning From Mistakes Works

The method, called probabilistic bias correction, does not forecast weather on its own. It acts as an add-on to existing models.

One half of the system compares about 20 years of past forecasts with what actually happened, by location and season, and subtracts the recurring error. The other half blends the forecast with recent observed weather and the long-term average. The two corrections are then averaged. Because the tool retrains daily and does not need a supercomputer, it could be run by smaller national weather services. The authors have posted the code publicly on GitHub.

MIT climate scientist Judah Cohen, a corresponding author, described the starting idea simply: since AI models make many errors, "maybe we can quantify those errors and remove them."

What the Results Show, and What They Do Not

The study reports that the correction doubled the subseasonal skill of the European Centre for Medium-Range Weather Forecasts' AI system and improved its operationally corrected physics-based ensemble for 91 percent of pressure targets, 92 percent of temperature targets and 98 percent of precipitation targets. It tripled the raw ensemble's small rainfall skill.

For extremes, the team reported a 51 percent gain in heat-wave forecast skill over ECMWF's standard correction, and a 76 to 92 percent gain in rainfall skill across thousands of catalogued flood events, where the standard correction left skill near zero.

The method also performed well in real time. A version entered in ECMWF's AI Weather Quest competition placed first for every variable and both lead times in the fall 2025 season, ahead of operational models from six forecasting centers, including NOAA, and 34 other teams.

The limits are important. The tool cannot recover a signal the underlying model never captured. Skill scores, while positive, remain low on an absolute scale, and evidence that better scores lead to better real-world decisions comes mostly from case studies.

An independent researcher also noted geographic limits. Froila M. Palmeiro of Spain's Institute of Geosciences told the Science Media Centre Spain that the improvement "is concentrated primarily in the tropical belt" and remains fairly limited in Europe. That suggests gains over the mid-latitude United States may be smaller than the global headline figures.

From Grid Operators to Farm Fields

If tools like this are adopted, the benefits would reach households indirectly. Grid operators could firm up fuel and staffing plans earlier ahead of a cold outbreak, which matters for preventing the kind of supply crunch that drives up winter bills. Water managers could adjust reservoir releases sooner, and growers could time planting, irrigation and harvest with more confidence.

Emergency managers might also gain extra lead time to pre-position generators, water and shelter supplies before a heat wave or flood-producing pattern. Kenya's meteorological department has already asked about using the tool for monsoon-season forecasts, according to Earth.com.

None of this changes what the public should rely on today. In the U.S., official weeks 3 to 4 outlooks come from NOAA's Climate Prediction Center and are expressed as probabilities, not day-by-day forecasts. A research advance in a journal is not the same as an operational upgrade.

What Comes Next

Researchers will watch whether national weather services test the method in their own systems and whether the gains hold up in more seasons and regions. Independent verification over the mid-latitudes, including North America, will be the key test.

Another open question is how the method performs during unusual years. This winter's very strong El Niño, for example, is the kind of slow-moving climate driver that can lend weeks 3 to 4 forecasts more predictability — or expose weaknesses in a correction trained on more typical years.

For now, the study suggests the blind spot between a two-week forecast and a seasonal outlook is narrowing, though slowly. Readers making weather-sensitive plans weeks ahead should keep using official outlooks and treat them as odds rather than guarantees.

What Readers Want to Know

What is the subseasonal gap? It is the period roughly two to six weeks ahead, especially weeks 3 and 4, where weather models often struggle to beat the long-term average.

How does the new AI tool work? It learns the recurring errors of existing models from about 20 years of past forecasts and blends in recent observations, then corrects new forecasts accordingly.

Who developed it? A 13-author team from Microsoft Research, Microsoft, MIT, Harvard College, the University of Toronto, ECMWF, and other institutions. Several authors are Microsoft employees.

Does this mean accurate month-ahead forecasts are here? No. Skill improved substantially, but remains modest in absolute terms, and an independent expert said the gains are concentrated in the tropics.

Will NOAA use it? The study does not say NOAA has adopted it. NOAA's Climate Prediction Center continues to issue official weeks 3 to 4 outlooks.

Who could benefit most? Farmers, water managers, power grid operators, and emergency planners who need to act weeks before a weather event.