ForecastEx Odds of a Major Hurricane Landfall Compared to Weather Models

By the afternoon of October 26th onward, ForecastEx odds were better than Google DeepMind’s at indicating a major hurricane landfall in Cuba.

Hurricane Melissa’s approach to Cuba was the first substantial test of the accuracy of ForecastEx’s prediction markets on major hurricane landfalls. 

Thus, it is appropriate to assess the probabilities implied by the ForecastEx prediction market against probabilities from alternative sources. 

Hurricane Melissa made landfall in Cuba at approximately 3:10 AM ET on Wednesday, October 29th, 2025, as a major (Category 3) Hurricane.

Below are the probabilities of a major (Category 3+) hurricane landfall for the 5 days leading up to Melissa’s actual landfall in Cuba (in 12-hour increments) from ForecastEx (red) and two weather models. One of the weather models is the ECMWF ENS global model (green), considered the best in the world but not specialized for hurricanes (it is a global model with relatively coarse spatial resolution). The other model is from Google DeepMind (dark blue), which adds machine-learning components to traditional physics-based models and is tuned for hurricane forecasting. The probabilities for each of the weather models are calculated by counting the proportion of their 50-member ensembles that were Category 3 or higher at the closest time to their Cuba landfall. 

Data from Google DeepMind Archive and ForecastEx. Chart made in Excel.

We can see that Google DeepMind substantially outperformed the lower-resolution, physics-only ECMWF ENS model, with probabilities consistently in the 2/3rds to 3/4ths range from the afternoon of October 24th onward. 

Since the ForecastEx market for this event was new, it was relatively unknown, and it didn’t attract much activity until October 26th. Thus, its probability on the 24th and 25th was at its initial seasonal level of 29%. However, traders sensed an opportunity by the 26th, and participation started moving the odds. 

Even with relatively few traders in a new market, the participants demonstrated the skill to discriminate between better and worse sources of information, and odds moved toward the eventual better-performing model and even surpassed it: 

By the afternoon of the 26th onward, ForecastEx odds were higher than Google DeepMind’s and had a much better mean Brier Score. 

This demonstrates some well-known attributes of prediction markets. The financial stakes involved strongly incentivize participants to seek out and use the very best available information. Further, the market aggregated this diverse set of information from various participants and distilled it into a single, useful probability. As such, it’s important to remember that these markets are not substitutes for other means of obtaining information but rather sit on top of them and essentially summarize all the other available information. This is why these markets have consistently demonstrated skill as good or better than the information on which they sit on top of. This is also what makes the information coming out of prediction markets so potentially valuable as a representation of the best consensus.

It is also interesting to note that from the afternoon of the 24th through the afternoon of the 26th, Google DeepMind’s probabilities were higher than ForecastEx’s. And in fact, other hurricane specialty models that are not probabilistic (like the HWRF model) had the location and intensity of Melissa’s Cuba landfall almost exactly correct by October 22nd (though that model was significantly off on the timing).

This is a good lesson in the context of thinking of YES positions as risk-hedging tools (where a YES position functions much like parametric insurance, providing rapid, unconditional payouts when a major hurricane hits). 

There is a considerable advantage to seeking out the best forecasts at the longest range so that YES positions can be obtained at the lowest price (probability), before other forecast methods catch up. This dynamic, of course, will push the odds towards the best long-range forecast methods and make prediction markets particularly useful for longer-range forecasts when there is the most disagreement between other methods.

Overall, this first test of the accuracy of ForecastEx’s prediction markets on major hurricane landfalls indicates that the well-established virtues of prediction markets are applicable in this domain just as much as they are in others.