How our predictions work

We collect match result prices from GB-licensed bookmakers. The odds add up to more than 100%, because the bookmaker builds in a margin, and they shade long shots more than favourites. We take the margin out with Shin's method, which allows for that shading, to get what the market really thinks each result's chance is.

The market is a strong guide: it reflects a lot of money and information. But it is not the whole story, so it is one source of three.

Our own model rates every team's attack and defence from past scores, using a Dixon-Coles model, a well-tested approach for football. It uses the last two years of results, with older games counting for less: a game a year ago counts half as much as one played today. Teams with few games are pulled towards average until there is enough evidence.

From those ratings it works out how likely each scoreline is, and from the scorelines the chance of a home win, a draw and an away win. It knows nothing about the odds, so it gives the blend a view that is independent of the bookmakers.

The last five games for each side, home and away, catch what longer-run ratings can miss: a new manager, an injury crisis, a team on a run.

The three sources are combined as a weighted average. Every week we re-fit the weights on past fixtures and test them on games held back from the fit, scoring each version with the Brier score, the standard measure of how close forecasts are to what happened. A new version only goes live if it beats the current one on those held-back games.

We also calibrate the result, so the numbers mean what they say: across a season, picks given 60% should win about 60% of the time. A 60% pick still loses about two times in five.

In the app, Pro picks are re-checked when confirmed lineups land, about an hour before kick-off. If a key player is missing and a pick weakens, Pro members hear first.

Every pick shows a fair price: the odds that match our chance exactly, one divided by the chance. A 50% pick has a fair price of 2.00, a 40% pick 2.50. When the best bookmaker price is longer than the fair price, the bet has value by our numbers. When it is shorter, the market is more confident than we are.

Picks update as prices move until kick-off, then they are frozen. The frozen pick is the one we settle on the final score and count in our record. Nothing is deleted or rewritten afterwards: every past prediction keeps its own page with the result.

See every week's picks and how they did:

Our chances are estimates, never guarantees, and even strong picks lose. Only bet what you can afford to lose, and never chase losses. You must be 18 or over to bet.

Free, confidential help is available from GamCare, and you can block gambling sites with GAMSTOP.