Stats.

Amanoki's regime forecast predicts which of four states a JEPX area will be in 60, 120, or 240 minutes from now — low, normal, high, or scarcity volatility. It answers with four probabilities that add up to one rather than with a single pick. This page is how well those probabilities have held up, measured on data the model was never fitted to.

The measure is a Brier score: the mean squared distance between the four probabilities the model gave and the outcome that actually occurred. It rewards being honest rather than being bold — a model that says 30% and is right 30% of the time scores well, and one that says 95% and is right 70% of the time scores badly. 0 is a perfect score and larger is worse, so lower numbers in the table below are better.

"Held out" means the score is computed on data held back from training, so the model cannot have memorised the answers. The baseline column is what you would score by ignoring every input and always predicting each state's long-run frequency; on this data that lands near 0.43, which is the number to read the model's score against. The improvement column is baseline ÷ model on the same held-out data — a comparison against that fixed-frequency baseline specifically. Most of the model's edge at short horizons comes from persistence, since the state a market is in now is a strong guide to the state it will be in an hour from now; methodology gives the feature breakdown.

Train N and Test N are how many labelled bars went into fitting and into scoring; the split is chronological, with the scored bars coming after the training ones and a gap between the two so that neighbouring bars cannot leak across it.

Pulled from api.amanoki.com/v1/admin/model-performance on page load, so these are the scores of the model actually serving requests right now rather than a figure typed into this page. The endpoint is open — no credentials.