Next-Generation Weather Forecast Models

Weather Forecasting Models have traditionally been large ensemble based models running on super computers for a long time. These models use massive amounts of data gathered from sensors and satelites, to compute predictions about the weather in the future.

Most of the methods involve various decision tree based prediction algorithms. The NOAA The High-Resolution Rapid Refresh (HRRR) has been one of the highest accuracy and fast reporting models available for many years. While it doesn’t cover the entire world, only the continental USA. It does provide detailed coverage and issues forecasts very quickly.

The European Centre for Medium-Range Weather Forecasts (ECMWF) model is also considered globally the current world standard in accuracy and precision with an ensemble based algorithms. Providing timely forecasts with high accuracy for the entire earth.

These models are based on ensembles of multiple different algorithms using data to compute predictions. Typically statistical based methods, these algorithms fall short at longer time horizons and the compute required means they do not model the earth in anyway so much as predict based on historical outcomes what will happen next in a statistical manner.

What’s been acheived

There’s been a huge amount of development recently in newer artificial intelligence and deep learning based weather forecasting models.

• Google released “GraphCast” in November 2023¹
• NOAA released “HRRR-Cast” in July 2025²
• Google released “WeatherNext”
• ECMWF released AIFS in 2024³
• Google released “WeatherNext 2” in November 2025⁴

These newer neural net based models are trained on largely the same data, but produce significantly higher accuracy especially at longer time horizons than previous ensemble based models.

Google’s WeatherNext 2 being the current SOTA, acheives a reported “99.82%” accuracy in their testing.⁵ The “Joint Probabilistic” model uses a neural net that models the entire earths physics within some probability range producing outstanding results over longer time horizons, and without being computationally intractable. It’s processed on a single TPU versus older super computer based methods that require enourmous amounts of power and take up large rooms compared to a single rack.

What comes next

The ability to model the entire atmosphere’s physics in a simulation based on current data yields highly accurate results for storm prediction.⁶ Expect significantly higher accuracy when it comes to forecasting local temperatures and precipitation times in the coming year. As well as forecasting tropical storms, cyclones, and hurricanes.

Can they be used for prediction markets?

No. These models operate over large spatial ranges, typically 0.25°. This huge range is the size of an entire city and aren’t going to make much difference when it comes to predicting winning temperatures in Kalshi Climate Markets. But their forecasts as available are being incorporated into Climate Sight for subscribers so the increased accuracy of these models can be leveraged.

As well active research is being put into apply these same kinds of neurel net models on estimation of hyper local regions, specific to climate markets. These custom models being developed will yield the highest accuracy forecasting for finding optimal probabilities on climate markets.