The ratings
Every team has an attacking and a defensive rating, updated after each game. They learn from goals and, more heavily, from shots: shots tell you more about how well a team played than the scoreline does. Each game counts for more or less depending on who the opponent was and where it was played, so beating the leaders away moves a rating further than beating the bottom side at home.
All four English divisions share one scale, as do the two Scottish ones, so a Championship side can be compared directly with a Premier League side.
From ratings to chances
The two teams' ratings give an expected scoreline, which becomes a chance of a home win, a draw and an away win. A small recent-form term (momentum) adjusts this for teams getting more or fewer points than their performances suggested over recent weeks. The conversion is checked against every past season in each league.
Line-ups
About 50 minutes before kick-off the confirmed XIs come in. In the Premier League each side is rated on its players' recent match ratings; in the other top flights, on the market value of the XI against the team's usual side. A weakened XI lowers that team's chances. In the cups, where rotation is heavier, the cup calculator measures how much of each side's usual league team is playing.
The shortlist
A game makes the shortlist when the model gives one result 55% or more. It is a statement of confidence, not a tip: a 60% shot loses four times in ten. The record shows whether those numbers have held up.
Flattered and better than it looks
Each team is ranked on how well it has actually played, by the same measure the ratings use, adjusted for opponents and venue. When that rank is five or more places from the real league position, the team is marked Flattered (higher in the table than its play deserves) or Better than it looks (lower).
Points v an average side
How many points an average team would have taken from the same fixtures, home and away, compared with what the team actually got. A side ten points ahead of that has done something real, whatever the table says about its neighbours.
How we test changes
Every idea is written down before it is tested, then checked on seasons the model has not seen. It only goes live if it improves the predictions in more than one stretch of seasons. Most ideas fail, and that's the point: it keeps the model honest.