AI football predictions: how the model works

Ratings based on goals and expected goals

In the Premier League, the EFL, the National League and the Scottish Premiership, each club's attack and defence ratings are fitted using every league match played since August 2024. Recent matches matter more: a game from eight months ago has half the weight of one from last weekend. A new club begins at the average for its division and shifts as more games arrive.

One scale across the whole pyramid

When clubs from different divisions meet in a cup tie, the model learns the distance between tiers. At a Championship ground, a League Two side is given a fair number, not a guess.

Turning ratings into a score matrix

Each side's expected goals come from the two ratings. For each scoreline between 0-0 and 10-10, a Poisson model turns them into a chance, with a small correction for low scores. Summing that matrix gives each market, so all the figures agree with one another: both teams to score, match result, every goals line, halves, half time plus full time, and result and BTTS.

Cards and corners

Each club's won and conceded rates for corners and bookings feed into the figures. The referee's own card rate is added once he has enough games. Corners bunch together, so we use a wider spread than Poisson.

When it turns into a tip

We make a selection a tip when it is at least 55% likely for us. The best UK price must beat our fair price by 2% or more. That price is shown when it's posted, and it's graded at the same price whether it wins or loses.