Six-week AI weather forecasts: “doubled” here means going from 3 to 6 percent

What this is about

A story has been making the rounds since mid-October, and it's now doing another lap through European media: AI can predict the weather six weeks ahead. It's based on a study published in Nature on October 14, 2026. A team led by Judah Cohen (MIT) and Lester Mackey (Microsoft Research), working with researchers at the European Centre for Medium-Range Weather Forecasts (ECMWF), developed a method called Probabilistic Bias Correction (PBC). PBC doesn't forecast the weather itself. It learns from an existing model's past mistakes and corrects that model's output, a bit like resetting a clock you know always runs three minutes slow. Microsoft says PBC “doubled” the accuracy of ECMWF's AI system. Notebookcheck then ran a headline saying researchers had “doubled reliability for 2 to 6 weeks.”

That's true, but it leaves out the scale. We went through the paper: what doubled was a value of about 0.03, which rose to about 0.06. Below we explain what that means, and whether AI can tell you if it will snow at Christmas.

What a skill score measures, and why 0.06 is small

Forecasts for week 3 and beyond don't give you temperatures for individual days. They give you probabilities. ECMWF's forecasting competition AI Weather Quest, which the paper follows, works like this: all past values for a given calendar week are sorted into five equal-sized classes (quintiles), from “very cold” to “very warm.” The forecast spreads 100 percent across those five classes.

The yardstick is RPSS (Ranked Probability Skill Score). It compares the forecast with climatology, the simplest possible forecast: every class gets 20 percent because you know nothing. An RPSS of 0 means “no better than the long-term average,” 1 would be perfect, and negative values are worse than the average. Roughly speaking, an RPSS of 0.06 means the forecast's error is 6 percent smaller than the error of guessing from statistics.

That's why “doubled” sounds so impressive. Going from 0.03 to 0.06 is doubling. So is going from 0.001 to 0.002. Climate consultant Robert Monjo (Climate Research Foundation FIC, Spain) told the Science Media Centre Spain that relative gains are misleading here, a bit like “dividing by zero”:

“Doubling a modest baseline doesn't guarantee a useful forecast.”

What the paper actually shows

The figures come from the preprint (arXiv 2604.16238). We calculated the last column ourselves. Because it isn't clear which lower bound goes with which gain, these are only rough ranges.

The pattern is clear: the lower the baseline, the higher the percentage. Where a model already has some skill, like PoET on temperature, PBC adds 5 to 7 percent. The widely quoted “98 percent for precipitation, 92 for temperature” doesn't describe the size of the improvement either. Those numbers are the share of grid cells where PBC beats the bias-corrected ECMWF model at all, by any margin.

Three more things stand out when you read the paper:

• Not six weeks. According to the methods section, only week 3 (days 19 to 25) and week 4 (days 26 to 32) were evaluated. In the paper, “two to six weeks” is just the definition of the research field, and Microsoft's product delivers weekly values up to six weeks out. The paper has no data on weeks 5 and 6.
• Coarse grid. The evaluation used weekly mean temperature, weekly total precipitation and air pressure on a 1.5-degree grid. Over Germany, one cell is roughly 165 by 105 kilometers. Temperature and precipitation were only scored over land, and snow wasn't scored at all.
• The reference has its own weaknesses. The “ground truth” is ERA5, a reanalysis, meaning a retrospective reconstruction of the weather built from observations and a model. Monjo points out that ERA5 doesn't assimilate precipitation measurements. So for exactly the variable with the biggest gain, the reference is itself a model product.

In fairness, there's another side. PBC wasn't just tested on past data. In the live AI Weather Quest competition, the PBC entry “MicroDuet” took first place in fall 2025 and winter 2025/26. It beat 34 teams plus the models of six weather services (ECMWF, NOAA, JMA, KMA, ECCC and the Russian weather service). In its blog, ECMWF confirms that MicroDuet “significantly outperforms” its own physics-based model IFS “across all variables and lead times.” According to the paper, PBC boosted skill for flooding by 76 to 92 percent. Here too, though, the baseline was only 0.04 to 0.06.

Tropics yes, Europe barely

For readers in Europe, the key criticism comes from geoscientist Froila M. Palmeiro (IGEO, CSIC-UCM): the gains are concentrated in the tropics and are “rather limited” in Europe. The paper only reports regional values in a supplementary figure. Its introduction says dynamical and AI models do worse than climatology at this range, “especially outside the tropics.” As example use cases, the authors cite semi-arid regions in East and Southern Africa. According to MIT, Kenya's weather service is already interested in using it during the rainy season.

Darío Redolat (FIC) urges skepticism about big percentage gains. The method needs to be tested in more regions, at higher resolution and on extreme events, he says. Matthew Chantry, head of machine learning at ECMWF, praises the work but calls moving it into operational use “very challenging.” He has a potential conflict of interest, since PBC was applied to an ECMWF model. According to the paper's disclosure, several authors are also Microsoft employees or hold Microsoft stock. Microsoft's claim that MicroDuet spotted two US winter storms in early 2026 some 19 to 26 days in advance rests on individual cases from its own account.

What weather services and apps give you for weeks 3 to 6 today

Since June 2023, ECMWF has run a daily extended-range ensemble forecast with 101 members over 46 days, at roughly 36-kilometer resolution. An ensemble is a set of model runs started from slightly different conditions, and how far they spread shows the uncertainty. Germany's weather service, the DWD, uses these ECMWF runs for its three-to-six-week outlooks. Since July 2026, ECMWF has also had its own AI model for this range, AIFS-SUBS. According to its preprint, it's as accurate as the physics-based model and uses about 200 times less energy. In the competition it edges out IFS, but it hasn't been peer-reviewed yet. ECMWF itself calls the two-to-six-week range “notoriously difficult.” ML models clearly beating the best physics-based systems is still the exception, and those exceptions rely on physics-based input data.

Weather apps, by contrast, often show 14, 16 or even 42 days complete with icons and temperatures. Meteorologists consider concrete forecasts reliable for about five to seven days and rough trends for up to ten to fourteen days. According to futurezone, ORF meteorologist Marcus Wadsak criticizes apps for often passing along individual model runs, outliers and all, without any filtering. Our take: a single-day temperature in week 4 implies a precision that even the best systems, with skill around 0.1, simply don't have.

Does the super El Niño help?

This year has a twist: an unusually strong El Niño is building in the tropical Pacific. According to Carbon Brief, the monthly mean temperature anomaly in the key Niño 3.4 region was 2.84 degrees in September, a record going back to the mid-19th century. An analysis of 14 climate models expects it to peak at around 4 degrees in December. Slow processes like sea surface temperatures are exactly where multi-week forecasts get their predictive power, and they matter most in the tropics. Our take: that fits the finding that PBC helps most there.

For Central Europe, the DWD plays it down. The signals are “much weaker,” it says. Strong El Niño events do bring more frequent disruptions of the polar vortex and a negative North Atlantic Oscillation (NAO). The NAO is the pressure seesaw between the Azores and Iceland, and its negative phase favors cold air outbreaks. But in past events, early winter was often mild and wet. The DWD's conclusion:

“So a strong El Niño alone is no basis for predicting a cold winter in Europe!”

And Christmas?

Counting from November 9, Christmas Eve is day 45, just inside ECMWF's 46-day range. It only falls into weeks 3 and 4, the range the paper tested, for forecasts issued between November 22 and December 5. Even then, you won't get a statement like “snow in Cologne on the 24th.” All you can get is a probability for a weekly average over an area about the size of half a German state, for example: Christmas week has an elevated chance of landing in the coldest fifth. Whether that turns into snow on the ground depends on details nobody knows three weeks out.

So the most honest forecast is still the statistics. According to DWD analyses, the chance of continuous snow cover over Christmas in the German lowlands has been about ten percent, and 30 to 50 percent in the low mountain ranges. It has dropped significantly since the 1960s.

By the way: what's behind the “94.6 percent” wildfire claim

The same pattern shows up in a second headline currently circulating: an AI supposedly predicts how wildfires spread “12 hours in advance” with “94.6 percent accuracy.” According to phys.org, it comes from the NYU Tandon School of Engineering, a prototype called “Digital Fire Twin.” The 94.6 percent refers to a single forecast for the Grasshopper Fire in Oregon, measured with an unspecified “standard measure of spatial overlap.” Another forecast for the same fire, made after a longer gap between satellite images, scored 72.4 percent. Three fires were tested in total. No peer-reviewed publication is mentioned, though a provisional patent application is. The Spanish report by okdiario even says 96.4 instead of 94.6 percent in its intro.

Our assessment

PBC is solid work. The method is lightweight, openly available (GitHub: mouatadid/pbc, MIT license) and won a real competition hosted by ECMWF. Getting week 3 and 4 precipitation forecasts from “worse than guessing” to “slightly better than guessing” is real scientific progress. For weather services in East Africa, it could be genuinely useful. Microsoft is already testing its own product built on it with selected customers in a private preview.

But the paper doesn't back up the headline “AI predicts the weather six weeks ahead.” Weeks 5 and 6 weren't evaluated, skill stays low, and experts say the effect in Europe is small. Our take: whenever you see a percentage, ask what the baseline is and what's being measured, whether it's weather or wildfires. And the question of a white Christmas gets the same answer this year as it would without AI: probably not, with about a one-in-ten chance in the lowlands.