The hurricane AI saw coming

When Hurricane Melissa began its run toward Jamaica last October, it was a Category 1 storm. Within hours it would become the strongest hurricane ever to strike the island, 185 mph, Category 5.

An AI model called WeatherNext predicted the storm five days out with 80% confidence. Meanwhile, a traditional, physics-based model was still hedging.

That’s not a weather story. That’s a risk pricing story.

For years, catastrophe models feeding reinsurance pricing have had a blind spot. It’s called rapid intensification — a storm strengthening by 35 mph or more in 24 hours — and it has been the industry’s most expensive surprise. Hurricane after hurricane has intensified far faster than the models forecast, leaving insurers with losses they hadn’t priced.

AI is now closing that gap. And the 2025 Atlantic hurricane season proved it.

Improved forecast

WeatherNext is Google DeepMind’s AI-powered weather model, trained on decades of global weather data and specialized datasets of extreme tropical cyclones. It runs 50 different, “What if?” scenarios simultaneously, something traditional models, bogged down solving complex atmospheric equations, simply can’t match.

The 2025 Atlantic hurricane season was its first real test at scale — and it passed. The National Hurricane Center’s own verification report found WeatherNext “slightly outperformed” NHC’s official track forecasts between 12 and 72 hours, and matched NHC’s skill on intensity across three Category 5 storms.

According to a new pre-season outlook from Howden, the National Weather Service has moved from experiment to full deployment for 2026. The European Centre for Medium-Range Weather Forecasts (ECMWF) reports its own AI model outperforms its physics-based systems by up to 20% across certain variables.

AI is no longer a research project. It’s in the forecast room. Its implications reach well beyond the Gulf of Mexico.

Also in the news: Wawanesa tells brokers it plans to explore a direct-to-consumer offering

Canadian P&C insurers and reinsurers have their own stake in this. Post-tropical Fiona carved through Nova Scotia and Prince Edward Island in 2022, causing more than $800 million in insured losses — the costliest weather event in Atlantic Canadian history.

The catastrophe models underpinning reinsurance treaty pricing and OSFI capital requirements are fed by the same forecast data that AI is now improving. Better storm track and intensity predictions mean more precise risk assessment, earlier — whether a storm is bearing down on Jamaica or the Maritimes.

The shrinking cone

The concrete impact shows up on every hurricane forecast map. The cone of uncertainty — the visual representation of historical track error — is getting smaller for 2026 because of AI-improved forecasting. A tighter cone means more precise landfall predictions, earlier.

Howden’s analysis confirms the stakes. Accumulated cyclone energy west of 60°W longitude has a far stronger historical relationship with insured losses than storm counts or landfalls alone. Intensity and track drive exposure; AI is now improving predictions of both.

Evan Thompson, principal director of the Meteorological Service Jamaica, put the human stakes plainly after Melissa: “With early evacuation and better preparation, that reduction in harm really does make a difference to our people. It does actually save their lives.”

The insurance industry’s version of that sentence is less dramatic but equally direct: better forecasts mean better-priced risk.

Not a replacement, an upgrade

Howden is measured about where the technology stands: “AI forecasting models are unlikely to replace physics-based models or human forecasters entirely. Instead, they are expected to complement existing approaches.”

That’s the right frame. The question for Canadian brokers, underwriters, and reinsurers isn’t whether AI replaces forecasters. It’s whether the industry is ready to use better data when it arrives, and whether it understands what a shrinking cone of uncertainty means for how catastrophe risk should be priced.

The forecast has gotten better. The premium models should follow.