From real results to a probability

We rate how strongly every team attacks and defends, from its real results. For any game we turn those ratings into a chance for every possible scoreline. Add up the right scorelines and you have the chance of a home win, of over 2.5 goals, of both teams scoring, and every other market.

Where there isn't enough history to trust the numbers, we say so and publish no pick. Where there is, we show the selection the model would stand on and the price it needs to be worth taking.

Each team gets two numbers: attack (how many goals it scores compared with the league average) and defence (how few it concedes). They are fitted together, so scoring three against the league leaders counts for more than three against the bottom side.

• Recent games count more. A result loses half its weight every six months or so, because squads and form change.
• Home advantage is measured for each competition, not assumed, and switched off at neutral-venue tournaments.
• Chances count, not just goals. Where the data records shots on target, half of each rating comes from the chances a team creates and allows, because goals alone are a noisy count of which chances went in.
• Small samples are pulled to the middle. A newly promoted side with three games doesn't get an extreme rating from one lucky win, and starts a little below average, where promoted sides usually are.
• National teams play too few games in one competition, so they are rated on every international in their confederation plus the World Cup and friendlies.

The ratings give each side an expected number of goals for this game. From those we work out the chance of every scoreline from 0-0 upwards, with a correction for low scores (the Dixon-Coles adjustment): real football has slightly more 0-0s and 1-1s than a plain goals model predicts.

Every market is read off that one grid. That is why our 1X2, over/under, GG and correct-score numbers always agree with each other, and why Forge can price two picks in the same game exactly instead of guessing.

We only publish a pick when the history behind it is deep enough:

• at least 200 finished games in the competition, and
• at least 3 games for each team in that sample.

Below that the game is still listed, marked “not enough history”, with no pick. When the sample clears the bar but is under 400 games, the pick is labelled as a guide. We tested this: on 40 games of history, picks landed 58% of the time against the 72% they claimed; on 200 they landed 71%, and on 400 they matched their claims. We would rather show nothing than a confident number built on a handful of games.

For each game we look at every market and rank the selections by how far the model's chance sits above what usually happens in that competition. A 75% home win in a league where home sides win 45% of the time is a strong read; a 75% over 1.5 goals in a league where that lands 76% of the time says nothing new.

The top of that list is the pick you see on the card and the match page.

When the model is most sure of a pick (a high chance, plenty of history and a clear favourite), it is marked ★ Strong. That is decided before kick-off and never changed after. Tested on two past seasons the model never saw, Strong picks were about one in six and landed about 80% of the time, against about 76% for the rest.

Strong picks have their own record on the track record page, beside the record for every pick.

Every headline pick is written to a log before kick-off with its probability, and graded automatically against the final score. Nothing is edited or deleted afterwards. The track record shows all of it, wins and losses, by league and by market.

It learns from results only. It doesn't know about an injury announced this morning, a rotated team before a cup final, a manager sacked yesterday or a waterlogged pitch. Use the numbers as a starting point, not the last word.

Probabilities are not promises: a 70% pick loses three times in ten, by design. Bet only what you can afford to lose. Responsible gambling.