Methods
Everything shown on a match page comes from the methods below. No hidden inputs, no odds, no human picks.
Goal model (time-decayed Dixon–Coles Poisson)
Each team gets an attack and a defence rating fitted on the last three seasons, with recent matches weighted more (half-life about one season) and a home-advantage term per league. The expected goals for a fixture give a full scoreline grid, adjusted for the known low-score dependence between teams, from which 1X2 and over/under probabilities are read off. Half-time markets use a second fit on half-time goals.
Strength rating (Elo with goal margin)
A second, independent view of team strength. When it disagrees strongly with the goal model, the prediction is labelled low confidence.
Low confidence
Shown when fewer than 200 settled matches back the market for that league, when a team has little history (newly promoted), when the two models disagree, or when fewer than three supporting statistics exist.
Freshness
Probabilities are computed once when kickoff is within 24 hours and refreshed once more within 6 hours. The track record settles the 24-hour version — the earliest, strictest publication. Warm-up: a league is predicted only after 150 matches of history.
Statistics shown on match pages
Each market shows the three to five statistics that best separated outcomes in our back-test for that league. Their meaning:
- Rating gap
- Difference between the two teams’ strength ratings (home minus away). Ratings rise with wins over strong opponents and with the margin of victory; a bigger gap means a clearer favourite.
- Form points gap
- Points collected over the last 10 matches, home team minus away team. Recent results, not reputation.
- Over 2.5 rate
- Share of each team’s last 10 matches that had three or more goals, averaged over both teams. High-scoring habits tend to persist.
- First-half goal rate
- Share of each team’s last 10 matches with at least one goal before half-time, averaged over both teams.
- Home goals scored
- Goals the home team scored per match over its last 10 matches.
- Away goals scored
- Goals the away team scored per match over its last 10 matches.
- Home goals conceded
- Goals the home team conceded per match over its last 10 matches.
- Away goals conceded
- Goals the away team conceded per match over its last 10 matches.
- Home clean sheets
- Share of the home team’s last 10 matches in which it conceded no goal.
- Away clean sheets
- Share of the away team’s last 10 matches in which it conceded no goal.
- Home shots on target
- Shots on target per match by the home team over its last 10 matches with match statistics. Chances created, not just goals scored.
- Away shots on target
- Shots on target per match by the away team over its last 10 matches with match statistics.
How accuracy is measured
Every prediction is settled against the final result: hit rate and Brier score, compared with a naive baseline (the Elo favourite for 1X2; the most frequent outcome for totals). Markets and leagues below 200 settled matches are shown as "building track record" and excluded from headline figures.
How confidence tiers are graded
Stated probabilities are grouped in 5-point buckets. For each league, market and bucket we count how often past predictions with that stated probability actually hit. A new prediction is tier A when that observed rate, shrunk toward the stated probability by 50 pseudo-predictions, is at least what we said; B when it falls short by up to 5 points; C beyond that. Fewer than 100 settled predictions in the bucket means "building". Only history known when the prediction was made counts, so a tier never benefits from hindsight.