← All research storiesWeather

Peer-reviewedWeather

A machine-learning add-on roughly doubled the skill of forecasts two to six weeks ahead

A low-cost AI method learns the habits of existing forecast models and corrects them, improving predictions in the hard-to-forecast range between next week and next season.

How far alongPeer-reviewedReviewed by independent experts for a journal before publication. Peer review is a quality check, not a grade of evidence strength: confidence still depends on study design, limitations, replication and the wider evidence.
How soon could it reach you?A few years outPromising, but needs more testing or engineering first. Our rough editorial guess is several years, with a lot of uncertainty.
Where it comes fromNatureMicrosoft Research; MIT; Harvard; University of Toronto; IMPA (Brazil); ECMWF; DTN · Oct 14, 2026Read the original →

What the scientists did

Weather forecasts are good for about a week, and seasonal outlooks can say whether a summer will be warmer than usual. The stretch in between, two to six weeks ahead, is so hard that forecasters call it a “predictability desert.” Yet farmers, power companies and emergency planners would love to know what’s coming then.

Every forecast model makes the same kinds of mistakes again and again. The team built a machine-learning method called probabilistic bias correction (PBC) that studies a model’s past forecasts, learns its typical errors, and adjusts new forecasts to cancel them out. It works as an add-on to existing models, and they tested it on forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF), including its AI model.

Words to know

Subseasonal forecast

A forecast for about two to six weeks ahead, between a normal weather forecast and a seasonal outlook.

  • Will the third week of January be colder than normal?
  • Will next month start wetter than usual?

Bias correction

Fixing a model’s repeated mistakes by learning how it tends to be wrong.

  • A model that always forecasts too much rain in one region
  • A thermometer that always reads 2 degrees too high

Forecast variables

The weather measurements being predicted. The study names these three.

  • Temperature
  • Precipitation (rain or snow)
  • Air pressure

What they found

PBC improved ECMWF’s physics-based model for 98% of precipitation targets, 92% of temperature targets and 91% of pressure targets, and it doubled the skill of ECMWF’s AI forecasting system. For rain and snow three to four weeks out, it turned forecasts that were worse than simply using long-term averages into ones that were better.

A version called MicroDuet ranked first for every variable and lead time in ECMWF’s AI Weather Quest competition, ahead of six national weather centers and 34 teams. It anticipated two early-2026 U.S. winter storms 19 to 26 days in advance.

2×skill of ECMWF’s AI forecasting system after correction
98%of precipitation targets improved for ECMWF’s physics model
19–26 daysadvance notice of two early-2026 U.S. winter storms

Why it matters to you

Forecasts weeks ahead could help your town stock salt before a cold snap, help farmers decide when to plant, and give families more time to prepare for dangerous winter storms. The method is open source and cheap to run, so weather services in poorer countries could use it too.

Where this could lead

Earlier storm and heat warnings

Weather services could add the correction to their forecasts to flag risky weeks sooner.

3–10 years

Better planning for farms and power grids

More reliable outlooks a month ahead could guide planting, water use and energy supply.

3–10 years

These are possibilities the research points toward, not promises. Most early findings take years of testing before they reach everyday life, and some never do.

Why scientists care

Most AI weather research tries to replace physics models. This study shows a cheaper route: let AI fix the models we already have. That matters because the two-to-six-week range has resisted improvement for decades.

Keep in mind

Independent experts note the gains were largest in the tropics and fairly small in Europe, and that doubling a low score doesn’t always make a forecast useful. A correction can’t fix a model’s mistakes in rare situations it has never seen, and the method was tested mostly on past data. Some authors work for Microsoft and ECMWF. We read the preprint abstract and news coverage, not the full Nature paper.

How far along is the science?

  • Peer-reviewedReviewed by independent experts for a journal before publication. Peer review is a quality check, not a grade of evidence strength: confidence still depends on study design, limitations, replication and the wider evidence.
  • PreprintA manuscript shared publicly, often before or alongside journal review. Check its separate journal and review status (each story says what we know). Results may change.
  • Conference talkPresented to other scientists at a meeting. Usually short and early; a full paper may come later.
  • Agency reportPublished by a science agency such as NASA, NOAA or NIH. Reviewed internally, but not by an outside journal.
  • Press releaseAn announcement from a university or company. Useful, but always check the research it describes.

How soon could it reach you?

  • Could reach you soonCould show up in products, services or advice soon. Our rough editorial guess is about 1 to 3 years, not a forecast from the research.
  • A few years outPromising, but needs more testing or engineering first. Our rough editorial guess is several years, with a lot of uncertainty.
  • Long-term scienceBasic research that builds knowledge. Real-world uses, if any, are far off.
  • Small stepA small update to earlier work. Useful to researchers, but not a change you would notice yet.

Every story links to its original source: the paper itself, or the conference program when no paper is out yet. Explanations are written with AI, usually from the paper’s abstract (each story says what it was written from), then checked against the source. We only use journals, recognized preprint servers, science agencies, scientific conferences and established science news outlets.