RPS (Ranked Probability Score)
Our primary metric. It measures how far the predicted 1X2 probabilities were from the real outcome — random guessing scores about 0.222, a perfect model scores 0. Lower is better.
Brier score
A multi-class accuracy measure for probabilities: the average squared distance between predicted and actual outcomes. Lower is better.
xG (expected goals)
The quality of chances a team creates, expressed in goals. We use per-match xG where coverage exists and estimates where it does not.
λ (expected goals)
The model's expected number of goals for each team in a specific fixture — the input that drives the whole score matrix.
Dixon-Coles
A bivariate Poisson model for football scores with a low-score correction, fitted per league on historical results.
Market λ injection
Our score engine aligns total-goals expectation with Friday market prices and blends the two grids (b = 0.7) — market information as an input, never as an answer.
Elo rating
A running strength rating updated after every match, scaled by goal difference and importance. Higher means stronger.
Confidence gate
A match is flagged high-confidence when its top probability clears a threshold chosen only on validation data. Accuracy claims are always bound to coverage.
Coverage
The share of matches that pass the confidence gate. An accuracy number without coverage is meaningless — we never publish one alone.
Edge (de-vigged)
The difference between our probability and the market's margin-free implied probability. We test for systematic edges with a sign test before claiming anything.
FAQ
Why is your headline accuracy ~53% and not 60%?
Because full-sample 1X2 accuracy has a physical ceiling set by market information (53–55% for the big five). A '60% full-sample' claim is not achievable — which is exactly why we publish gated accuracy paired with coverage.
Are your probabilities betting tips?
No. Everything on this site is a statistical estimate for information and entertainment. We never publish wagering advice and we accept no gambling advertising.
How often does the model update?
Predictions are locked before kickoff every day (02:30 UTC) and reconciled automatically after the final whistle. Nothing is edited or deleted afterwards.