The closing line is the exam
A model that never meets the market is a story. The closing line, the last price before kick-off with the margin removed, is the best public forecast there is, and the only honest test of a model is a walk-forward comparison against it: refit on everything before the week, price the week, score it, move on.
The scores
- Brier: the squared distance between the forecast and the outcome, summed over home, draw and away. Lower is better; a coin scores 0.667.
- Log loss: minus the log probability of what happened. Punishes confident mistakes hard.
- Skill against the market: 1 − Brier_model / Brier_market. Negative means the market is better.
- Calibration (ECE): are the stated 60%s right 60% of the time?
What our numbers say
Across thirteen leagues the model sits 1.5–5% behind the closing line in Brier skill. That is not a failure; it is what a public-data model should expect against a market with team news and money. What the model adds is a price for every market, every match, before the market opens, with the band attached. The Trust tab prints every league's numbers, and the ROI column shows flat stakes at the close losing money, which is the honest reading.