This Hurricane Season, Two Forecast Models Stand Out, but for Very Different Reasons
Google’s newest DeepMind hurricane model was the star of the 2025 hurricane season while America’s flagship weather model stumbled to new lows
Author’s Note: This will be the last routine issuance of the daily Eye on the Tropics newsletter for the 2025 hurricane season unless conditions warrant otherwise. I extend my deepest gratitude to all our regular readers and new subscribers. I’ve written over 500 daily newsletters, and I’m honored you’ve chosen to put your trust in the no-hype forecast information and specialized expertise I provide each day. Eye on the Tropics is reader-supported, and without your support, readership, and shares, it wouldn’t exist. We’ll resume the daily tropical newsletter in June 2026.
The 2025 Atlantic hurricane season appears to be winding down.
Though the season runs through the end of November, typically about 95% of basinwide tropical activity happens before November 3rd and nearly all U.S. hurricane landfalls have occurred before the first week of November (with rare exception like Nicole in 2022 or Kate in 1985, the latest U.S. hurricane landfall on record).
With long-range computer models signaling a barren tropical Atlantic for the foreseeable future, it’s unlikely we see any additional appreciable activity before the season officially ends in 27 days.
Though the season will fall just short of the 30-year average for named storms and hurricanes, the hurricanes that did form were exceptionally strong and long-lived, pushing overall hurricane season activity – as measured by the Accumulated Cyclone Energy or ACE – to above-average levels and on-par with seasons like 2018 and 2019 which saw higher counts in the named storm and hurricane departments.
The 2025 hurricane season recorded three Category 5 hurricanes, the second-most on record for a single season, trailing only the 2005 hurricane season that churned out four Category 5s.
On a positive note, the 2025 hurricane season is the first in a decade without a U.S. hurricane landfall and the first hurricane season since 2019 without a tropical system directly hitting Florida.
A banner year for Google DeepMind
We’ve discussed in detail this year the stunning performance of Google’s DeepMind AI-based hurricane model. The hurricane-specific model was introduced back in June and is trained on historical global weather data as well as information from global tropical cyclone records over the past 45 years.
The beauty of DeepMind and other similar data-driven, AI-based weather models is how much more quickly they produce a forecast compared to their traditional physics-based counterparts that require some of the most expensive and advanced supercomputers in the world. Beyond that, these “smart” models with their neural network architectures have the ability to learn from their mistakes and correct on-the-fly.
In analyses shared by retired National Hurricane Center chief forecaster James Franklin last week, Google’s DeepMind has been the top performing model for track and intensity in both the Atlantic and the Pacific basins this season.
Not only that, DeepMind outpaced even the gold-standard corrected consensus models as well as the National Hurricane Center’s official track forecasts this year.
The model is being evaluated this season in partnership with the National Hurricane Center and gave NHC forecasters extra confidence to issue uncharacteristically aggressive forecasts in the days leading up to Melissa, sounding the alarm well in advance of the season’s strongest and most impactful hurricane.
America’s flagship weather model runs aground this hurricane season
On the other side of the ledger is the U.S. GFS, America’s flagship weather model.
According to preliminary verification compiled by University of Miami hurricane expert and researcher Brian McNoldy, the GFS was head-and-shoulders the worst-performing public forecast model of the 2025 Atlantic hurricane season (note: he did not evaluate the European model and its offshoots since they are not yet publicly available, but the Euro assuredly outperformed the GFS this season).
The GFS was especially awful in its forecast for Melissa, with an average 5-day track error ballooning to over 500 miles, insisting on a turn out to sea that never transpired.
Brian’s preliminary season-to-date verification finds the 5-day track error from the interpolated version of the GFS – the version available to forecasters when they are making their predictions – at its highest point in 20 years.
Since the GFS is usually one of the main forecast models consulted by NHC when making its forecasts, official forecast errors are typically tied to the performance of the GFS. In other words, if the GFS has a bad season, the NHC official forecast usually suffers and if the GFS has a good season, the official forecast usually benefits.
In the case of the 2025 Atlantic hurricane season, the official forecast from NHC was below the 5-year average for the center despite the American GFS errors reaching 20 year highs. The reason NHC zigged while the GFS zagged was likely because it had Google’s DeepMind in the mix this season to counteract the American model’s negative influence.
Forecasters, including myself, recognized quickly how poorly the American GFS was performing early on in 2025 while how exceptional AI models like Google’s DeepMind were proving. Downplaying the GFS in forecasts while building a real-time testbed to leverage next-generation models like DeepMind and the European Centre’s suite of probabilistic guidance and AI models cushioned what could have been a serious threat from the GFS’s spectacularly bad performance to the accuracy of official hurricane forecasts.
The experts at the National Hurricane Center should be credited for being forward-thinking and working to evaluate and implement new powerhouse AI models in real-time this season.
What’s next for the GFS?
It’s not immediately clear why the GFS performed so poorly this hurricane season. Some have speculated the lapse in data collection from DOGE-related government cuts this year could have been a contributing factor, but presumably such a factor would have affected other global physics-based models as well, not just the American GFS.
It will take some time for NOAA’s Environmental Modeling Center – the group responsible for developing and maintaining the GFS – to diagnose and unpack where the model went astray.
We’ll be sure to revisit this topic in future newsletters.
Thank you for covering this hurricane season with such clarity. It has certainly been one that for me, will be remembered as having nothing come ashore near me, and that is always welcome. On the other hand, so many have lost so much this season, while we appear to have a government that would rather inflict cruelty than try to help anyone.
Unfortunately, the sole developer for TC model verification at EMC was laid off at the end of August due to a ~20% cut in the contract budget. It will likely be awhile before any of this is further investigated by the understaffed center.