Brier Score Explained
Measure the distance between probabilities and outcomes.
Measure the distance between probabilities and outcomes.
The multiclass definition
For each match, take the squared difference between every predicted probability and its outcome indicator, then sum across home, draw and away. Average these match scores over the evaluation cohort. The NQ convention uses the unscaled three-class sum, ranging from 0 to 2; lower is better.
An illustrative calculation
For a forecast of 50% home, 30% draw and 20% away, followed by a home win, the score is (0.50 − 1)² + (0.30 − 0)² + (0.20 − 0)² = 0.38. The calculation evaluates the full distribution, not just the selected winner.
Compare like with like
Some implementations scale Brier scores differently. Always publish the convention, outcome classes, match count and forecast horizon. A model evaluated on one league cannot be fairly ranked against another evaluated on an unrelated set of fixtures.
What the score leaves open
A single average is not a complete diagnosis. Pair it with reliability diagrams and cohort breakdowns. In v0.1 the scoring utility is implemented for future evaluation, while all public historical score fields remain unavailable.
Sources & further reading
Prepared by NinetyQuant. Numerical examples are illustrative. References explain general concepts and do not validate NQ’s demo forecasts.