Your Weather App May Soon Know the Sky Six Weeks Out

Open your favorite Weather App and scroll past day ten, and you will usually find a shrug: vague icons, wide ranges, and confidence that fades with every extra day. This week, a research team led by Microsoft said it has a machine learning method that could push dependable forecasts out to six weeks, and it is giving the code away for free.

The method is called probabilistic bias correction, and it was described in the journal Nature on October 14, 2026. Rather than building yet another forecasting model from scratch, the team built a wrapper that sits on top of existing forecasts, studies the systematic mistakes those forecasts have made in the past, and adjusts them. It learns from patterns of error, not from raw observations, which makes it cheaper to run and easier to plug into systems that weather agencies already use, according to Unite.AI's breakdown of the paper.

The forecasting blind spot

Most modern forecasting is excellent for about ten days. Beyond that, tiny uncertainties in today's atmosphere snowball, model biases compound, and skill collapses. The two-to-six-week window is known as the subseasonal range, and it has been a weak point for both physics-based models and newer AI systems. Farmers deciding when to plant, emergency managers positioning supplies, and grid operators planning energy purchases all live in that blind spot.

The problem is partly built in. Models fill gaps where weather stations are sparse or processes are poorly understood with approximations, and those shortcuts create errors that grow with time. One example the researchers cite is the warming planet itself: a model trained on older records can lean colder than the world it is now forecasting into, and the correction layer can adjust for that drift.

Fixing the mistakes instead of rebuilding the model

The results reported in the paper are striking. In tests covering days nineteen through thirty-two ahead, the corrected forecasts improved on Europe's top physics-based system across ninety-eight percent of the rain targets evaluated. Temperature forecasts improved across more than nine in ten locations tested, and sea-level pressure improved across roughly nine in ten, as reported by WebProNews.

The same approach also topped a public forecasting competition. The team's entry ranked first on every variable and lead time tested in the inaugural season of the AI Weather Quest, a contest running from September through November of last year. And the tool appears to work without a supercomputer: senior researcher Lester Mackey noted that it runs on ordinary processors, no graphics-card clusters required, which matters for weather agencies in developing nations that cannot afford precision forecasting at its current price.

The work was led by Hannah Guan, who carried out the research as a Microsoft intern, alongside a team spanning Microsoft, the Massachusetts Institute of Technology, Harvard, the University of Toronto, and the European forecasting agency that runs the system the method was tested against. The paper's corresponding authors include atmospheric scientist Judah Cohen, a longtime specialist in subseasonal prediction. Cohen told MIT News the findings were a breakthrough: until now, he said, nobody had shown a model predicting this range better than long-term historical averages, and most attempts had come out worse.

What a month and a half of foresight buys the rest of us

For ordinary people, a trustworthy month-and-a-half outlook changes the calendar — and eventually, the Weather App on your phone. A festival organizer could weigh rain risk a month and a half out instead of two weeks. A traveler booking a beach trip during hurricane season gets more than a guess. Airlines, insurers, and farmers all make expensive bets on exactly this window, and better information there is worth real money.

The bigger story may be who gets access. Kenyan forecasters tested the method during this year's rainy season, according to coverage of the announcement, and the open-source release means a small national weather service can adopt it without licensing fees or new hardware. That is a sharp contrast with the current race, where Google, Nvidia, and Microsoft are each building flagship AI weather systems that mostly serve those who can afford them. A correction layer that upgrades existing tools instead of replacing them lowers the bar for everyone else.

There is a limit worth stating plainly. The method sharpens forecasts that already have some skill; it cannot create foresight where the underlying model is blind, and a single city or town will still see wide uncertainty five weeks out. The atmosphere does not negotiate. But the direction of travel is clear: the era when a Weather App stopped being useful after a week and a half may be ending, and the next breakthrough looks less like a bigger machine than a smarter editor. For more on how this week's storms are testing the forecasts we already have, see our coverage of the slow-moving system aimed at the East Coast, and for what happens when long-range climate patterns swing the other way, read about the El Niño-driven sugar squeeze.

External sources: the full paper breakdown, reporting on the forecast gains, and the study's journal reference. More from this newsroom: East Coast braces for a slow coastal storm and the El Niño sugar shortage coming for your wallet.