Calibration
Model record
What this model is, how it is measured, and how it compares with the closing market. This page exists so the numbers on the rest of the terminal can be read correctly.
Signal lostCould not reach the prediction service
Check that the API container is running and reachable at the configured address, then reload.
network /api/model/performanceMethod
How the numbers are produced
Dixon-Coles on goalsA bivariate Poisson fitted to recent results, with exponential decay so older matches count for less, and a correction for low scores because 0-0, 1-0 and 1-1 happen more often than independent Poisson margins allow.
One matrix, every marketAll twenty-seven markets on a fixture page are sums over the same score-probability matrix. That is why they can never contradict each other, and why a lineup change moves all of them together.
Fair price, not offered priceEvery price on this terminal is 1 divided by a probability, with no margin added. A bookmaker adds theirs, which is their income, so a real quote is always shorter than the number shown here.
Ratings computed in-houseElo and pi-ratings are replayed from stored results and rebuilt at every backtest fold. Building them once across all history and reusing them for earlier folds would leak future results into past predictions.
No language model in the numbersNothing on the probability or reasoning panels is generated text. The drivers are arithmetic over stored quantities. A language model, when configured, only writes the briefing paragraph and cannot alter a figure.
Interpretation
What this does not mean
A divergence is not an edgeWhere the model disagrees with a price, the more likely explanation is that the model is wrong. Bookmakers price these matches for a living, and the closing line is the sharpest public estimate available.
Results are not evidenceOver a single season, betting outcomes are dominated by variance. The only early signal that a process has genuine edge is consistently taking better prices than the market closes at, which is why closing line value is tracked per flagged bet.
Calibration is not accuracyA well-calibrated model can be less accurate than a confident one and still be more useful, because its probabilities mean what they say. Ranked probability score and log loss reward being right about how uncertain a match is, which is the thing worth being right about.
The measured gap is small but realThe model sits roughly 0.008 ranked probability score behind the closing line over 2,281 fixtures. That is close enough to be interesting and far enough to be decisive.