WeatherTiger’s Hurricane Watch is a reader-supported publication. Paid subscribers get Florida-focused daily tropical briefings, plus coverage of every U.S. hurricane threat, our exclusive real-time seasonal forecast model, and the ability to comment and ask questions. And check out our WeatherTiger Pro service for market- and trading-oriented data solutions.
Today’s double-length newsletter features some rare WeatherTiger video content, featured at the top of the post. Please feel free to share this column and video with those who need it.
In 2017, Dippin’ Dots changed their slogan from “The Ice Cream of the Future” to “Taste the Fun,” a tacit acknowledgement that we were, at long last, living in the future. While the intervening nine years might be charitably characterized as a mixed bag, one of the non-disappointing aspects of the alleged future has been steady progress in hurricane forecast accuracy. Over the past 40 years, the National Hurricane Center’s one- to three-day track forecast errors have diminished by around 75%, with error reductions of about 50% in the past two decades at four- and five-day lead times. In 2025, NHC track forecasts narrowly missed the record skill of the 2024 season, probably due to a more challenging set of storms to forecast.
With the Tropical Atlantic currently in its typical mid-summer lull, let’s have a little chat about the computer models that have made those gains possible. The only area to watch in the Atlantic over the upcoming week is some rotation likely to develop along an old frontal boundary in the eastern Gulf over the weekend. As this system lingers near the Southeastern U.S. into next week, development of a tropical depression is a remote possibility, with a 20% chance of formation in Wednesday’s NHC Tropical Weather Outlook. However, mid-level moisture is limited and upper-level winds are marginal, so whether something organizes or not the primary impact will be enhanced rain chances in Florida and the coastal Southeast. Otherwise, the Tropical Atlantic is sheared out, dusty, quiet, and set to stay that way through the end of July.
Meteorologists can make predictions like that with relative confidence thanks to computers’ incredible facility with repetitive calculations. Models use information about current weather conditions plus math, lots of math, to estimate how the atmosphere will change with time. There are three main types of models we care about for hurricane forecasting:
Physics-based global models. Traditional physics-based weather modeling uses simplified versions of the fundamental fluid dynamics equations that govern how air and water move and exchange energy to “solve” the gigantic calculus problem of the atmosphere with time. Global models are capable of simulating hurricanes, though their approximations of complex hurricane core physics are rough and sub-optimal. The GFS and European models are two examples of this type.
AI/statistics-based global models. A new entrant into the model demolition derby in the 2020s, AI guidance takes a sophisticated, statistical approach to hurricane forecasting. Basically, these models use a reliable dataset of historical atmospheric conditions, including hurricane activity, to train a model and then figure out what future weather is most likely from an initial snapshot of current conditions. There are new, AI variants of older models like the GFS that take this approach, and even models like Google Deepmind that are tuned to make optimized hurricane forecasts. More on this later.
Hurricane-specific models. These models are physics-based, but take fewer numerical shortcuts in order to better capture the structure of actual hurricanes, operating on smaller, storm-focused scales at much higher resolution than global models. That can result in more accurate intensity and track forecasts, plus useable guidance for localized wind, rain, and surge predictions.
Importantly, most forecast models are run as ensembles, in which many versions of the same model are made by tweaking initial weather conditions — for example, shifting the starting location of a hurricane by a few miles this way or that to reflect real uncertainty in what’s actually going on in the atmosphere. Small initial differences can have major long-term forecast implications, so the way the ensemble members spread out with time gives a sense of the confidence of our predictions. Averaging many ensemble members together also results in a more accurate forecast.
There are a wide assortment of models available to the public and huge differences in forecast skill between them. So how do you know which to trust? That’s where I come in. As a scientist who works with hurricane model data each day (and is perpetually frustrated how model data is misused by Facebook pages like Uncle Frank’s Xtreme Weather Basement and Gulf Coast Hurricane Alert/Monster Jam HQ), today I’m going to help you separate those you can trust from those you can’t, building on validation data for models that do both track and intensity forecasting from the 2025 NHC Forecast Verification Report. (Click here for an NHC model abbreviation explainer.)
Without further ado, I present the 2026 WeatherTiger-USA TODAY-LaCroix Hurricane Model Power Rankings, in convenient and memetic tier format. (Ratings are subjective, unofficial, and unaffiliated with the J.D. Power & Associates Award for Best Mid-size Luxury Sedan Powertrain Warranty.)
F-tier
The NAVGEM are NAM are useful, in that if someone posts their hurricane output, you can be certain they trying to scare you for the clicks or perhaps just have no clue what they are talking about. The NAM is being retired in August, while NAVGEM, famous for producing square hurricanes, toils on in futility.
D-tier
There’s a lot of models out there, including mesoscale modeling that is not designed for hurricanes, and private-label and/or AI modeling. I assign these and other model miscellany a provisional grade of D+ (the gentleman’s F) and advise that you stick to name brands. Any model worth its salt must prove it over time.
C-tier
C-tier models have some strengths, but also in my opinion don’t have a particular use case for hurricanes. As such, these models often get used to pad out social media posts with an forecast agenda. C is for #content, and that’s not good enough for me.
German (ICON). The ICON is a newer model that occasionally catches trends early. For every success, there are numerous false alarms, which coupled with a poor ability to resolve strong storms makes it unreliable.
HWRF. The HWRF used to be the state-of-the-art American hurricane model, but then, dollar stores used to sell things for $1. Known for a comical bias towards over-strengthening storms, HWRF has been in decline for years and is slated for an overdue retirement in 2026. Be skeptical of Cat 8 HWRF forecasts in the meantime.
Canadian (CMC). The Canadian model is strong prognosticator of mid-latitude jet stream patterns, even outperforming the GFS globally by this measure. However, it doesn’t handle convection particularly well, limiting its value in hurricane contexts.
AI-GFS. A not ready for primetime player; will hopefully contribute down the road.
B-tier
HMON. B-tier models are respectable but hit-or-miss, a duality exemplified by the HMON. Another older generation American hurricane model, the HMON, like HWRF, can erroneously intensify or just plain whiff on storms. However, HMON clings to relevance like a Garfield to the window of a Dodge Grand Caravan, posting surprisingly good forecast skill scores over the past three years, particularly in two- to four-day intensity predictions.
British (UKMET). The UKMET ranks second-best in overall global weather pattern forecast errors. For tropical weather, it is an independent opinion worth considering; for instance, it took Ian into Southwest Florida when others swung it farther west into the Gulf. However, the UKMET’s delayed ensembles and limited availability make it a little less operationally useful to forecasters.
American (GFS). The main American global forecast model, the GFS and GFS Ensembles have been riding the hurricane struggle bus in the 2020s. The GFS took Ian, Idalia, Beryl, Debby, and Helene too far west, then mixed things up by embarrassing itself to the north and east with Melissa in 2025. Still, it outperforms non-hurricane-specific models in intensity forecasting, and hopefully a long-delayed model upgrade coming this fall will address stagnation in the GFS’ forecast skill.
Euro-AI. The AI-aided of the European Center’s global model is routinely producing slightly more skillful forecasts for global temperatures and upper-air patterns than the flagship Euro, which in turn is more accurate than the GFS. Lower resolution, a tendency to develop marginal disturbances, and a short operational history are drawbacks for hurricane forecasting, but the Euro-AI has earned its spot on the list.
A-tier
HAFS-A and HAFS-B. The few, the proud: the A-tier models. The HAFS bros are the current American hurricane-specific models, and have been operational for the last three years. HAFS leverages the strengths of hurricane-specific modeling by being designed to work best with real-time radar observations from Hurricane Hunters. For that reason, HAFS was hit or miss with 2025’s open ocean storms, it tends to come through in the clutch when storms are under reconnaissance coverage.
European (ECMWF). The predictive power of the European model for global weather patterns remains unequalled. For hurricanes, “King” Euro has the best average track skill of any individual model over 2023-2025; it also did quite well with high-impact Hurricanes Ian, Idalia, Helene, Milton, and Melissa. Still, kings are human, and the Euro has feet of clay in intensity forecasting, though in a predictable way.
HCCA and TVCN. I lied earlier. There’s actually a bonus fourth type of hurricane model: consensus models like the HCCA and TVCN, which blend output and correct for predictable biases of other reliable models. In doing so, these aids manage higher average accuracy and greater consistency than any individual model, and are an essential part of the forecast workflow at NHC.
Google Deepmind. Last June, Google’s Deepmind division announced a partnership with the NHC to develop and validate hurricane-specific AI modeling. The Google x NHC collab bore immediate fruit to an astonishing degree; in both track and intensity, Google Deepmind had slightly lower average forecast errors than all other models and even the NHC’s official forecasts at most lead times in the 2025 hurricane season. The NHC takes incorporating Deepmind output into their forecast process seriously, indicating this rookie performance is unlikely to be a fluke. New for 2026, Deepmind now comes in a 1,000-member ensemble, helping forecasters get a truly exhaustive handle on possible track and intensity outcomes for a storm. Dippin’ Dots are right: The Future is now.
S-tier
NHC official forecast. AI wunderkinds notwithstanding, there’s only one GOAT. Over the past three years, official National Hurricane Center track and intensity forecasts have, as usual, bested all individual models and equaled the best model blends at all lead times. Even better, NHC forecasts are far more consistent than any individual or consensus model, with five-day forecasts shifting less than 60 miles on average between advisories, against over 100 miles for most model guidance. That makes them more accurate than models, and more dependable.
So, are models useful after all? The answer is a clear yes for the experts and professionals who are equipped to properly assess their output against observations and experience. They are not very helpful for the public, particularly as the scariest maps from any model tier tend to find the most eyeballs. This was true in 2006, it’s truer now as models proliferate, and it’ll likely be even truer in 2046 when your microwave can run its own billion-member AI hurricane ensemble and Dippin’ Dots is marketed as “The Ice Cream of the Past.”
The good news is, unless you are a professional forecaster, you really don’t need to look at models. An individual model may outperform the NHC for any given storm, but over time the NHC’s meticulous approach wins by incorporating skillful models into a forecast framework blending computational power with human expertise. Use as directed, and as always, keep watching the skies.
Hi: Great article as usual. However...if you're going to reference model names in the text, pls make sure you identify their corresponding acronyms in the legends on the graphs so it is not a chore to correlate them. Thx.
Excellent article and video, going into my Save file. Thank you for sharing.