Predictor — learning loop
Aratea is a weather-factor discovery engine. Every named feature here is a hypothesis; every training run measures whether it carries signal. The bench is the same row-set kalshi_mid Brier — beat the market, on its own ground.
What goes into Aratea's number
Aratea runs several forecasting models in parallel. Each card below is an independent estimate of the same question: "what's the probability this happens?". The numbers do not add up to 100% — they're different ways of guessing the same answer. Only one of them — the champion — actually places the bet; the others run in shadow.
Aratea's parallel estimates
Champion = the model whose probability is actually used to place the bet. Challengers run in parallel as shadow forecasts; one only becomes champion after beating the current one on a rolling window of resolved trades. This is why the “champion” probability shown in the public view can differ from the “learned model” probability here.
- Multi-model ensemblechampion25.6%
The mean of four vendor models (ECMWF, GraphCast, GFS, JMA). Useful as a smoothing baseline.
vendor_ensemble - Learned modelchallenger14.4%
A small regression that learns how much weight to give each component based on past resolutions. This is the one Aratea actually bets with.
learned_v2
What the market is paying
What the Kalshi order book is implying right now. The yardstick Aratea must beat.
Shown for comparison. This is the benchmark Aratea is trying to beat — not one of its inputs.
Track record so far
The chart below shows the Brier score of two forecasters across every training pass. Brier scores accuracy: 0 = perfect, 1 = always wrong, lower is better.
● blue line = Aratea's learned model. ● yellow line = the Kalshi market mid on the exact same events. When blue stays under yellow, Aratea has signal the order book doesn't.