How the model works
SportEdge uses a tested statistical model, blended with the betting market and scored in public. No model predicts football with high accuracy (the best are right about half the time on match results), so what matters is whether the percentages hold up.
1. Team ratings from goals and expected goals
Each team gets an attack rating and a defence rating. We fit them with the Dixon-Coles method (1997): goals follow a Poisson distribution, with a correction for low scores such as 0-0 and 1-1. Recent matches count more, with a half-life of about one year. Each league has its own scoring rate and home advantage. A second set of ratings uses expected goals, which change less from week to week than goals do. We blend the two sets.
A team that is new to a league starts from the ratings of last season's weakest teams instead of the league average. Promoted teams are usually among the weakest.
2. Every market from one score grid
The ratings give each team's expected goals. The expected goals give the chance of every score up to 10-10. Match result, double chance, draw no bet, both teams to score, totals, team totals, correct score and Asian handicaps all come from that one grid, so they never contradict each other. Half-time markets use the share of goals each team scores before the break. Corners and cards have their own models: negative binomial counts, and a referee factor for cards.
3. Blended with the market
With the bookmakers’ margin removed (using Shin’s method), their price is the best public estimate of a match, so we blend the model with it. We set the blend weight on matches the model hasn’t seen, and in most leagues the market gets the larger share, leaving the model to adjust it a little.
4. Champions League and other European ties
Our ratings compare teams inside one league. A Champions League tie puts teams from two leagues together, and we do not yet have a tested way to put all leagues on one scale. So for European ties we show the market's own view: the sharpest bookmaker's price (Pinnacle or an exchange when available, otherwise the median of all books), with the margin removed. We find the expected goals that give those prices, and the goal markets come from those expected goals.
We record European leans in the ledger like any other prediction, but we claim no edge and show no value picks there. Each club's form and rating come from its own league, when we cover that league.
5. Scored in public
For every league we compare the blend's log loss and ranked probability score with the closing market's, with a bootstrap interval. We show value picks only in leagues where the blend beat the closing line by more than noise (at the moment there are none, so the site shows probabilities without picks). We seal each day's predictions with a hash before kick-off and settle them afterwards on the track record page.
Where the data comes from
- Results, shots, corners, cards and closing prices: football-data.co.uk.
- The 7-day fixture list, European prices and European results: The Odds API.
What this will not do
It will not give you "sure odds". A 70% favourite loses three times in ten, and a run of five losses happens to everyone. A realistic long-run result for a skilled bettor is a few percent of the money staked, and bookmakers limit accounts that win.