How the predicted lineups are built

We tested lineup methods over all 740 team-matches of the 2025/26 Premier League season, scoring each by how many of the actual eleven starters it named. Repeating the previous match’s starting eleven got 8.81 of 11 right on average.

SquadCheck’s earlier model picked the players who had started most often across the season, weighted by how important they were to the side. It got 7.95 — nearly one starter a match worse than doing nothing clever at all. Giving every player the same weight changed almost nothing (7.95 either way), so the problem was not the weights; it was that season-long frequency forgets what the manager did last week. The current method is built around that result.

1. Start from the last match. The predicted eleven begins as the club’s starters from its most recent match.
2. Remove who cannot play. Any player the official feed lists as injured, doubtful or suspended is taken out. Absent players are listed separately, flagged when they would otherwise have started.
3. Fill the gaps. Each gap goes to the best-scoring available player. The score mixes how often he has started recently (35%), how often he has started this season (20%) and his overall weight in the squad (45%). Recent matches count most: the last five in full, the five before at 70%, the five before that at 40%.
4. Choose the shape. The formation is the club’s most common one across its last ten matches, ties going to the more recent. A club with no history defaults to 4-4-2.
5. Put players in positions. Every pairing of a player with a position is scored at once — 50% how well his usual position fits the slot, 35% the selection score above, 15% recent starts — and the best pairings are made first. Any slot still empty is filled starting with the position that has the fewest candidates.

The weight in step 3 is the same one behind the Power Loss figure — see how Power Loss is calculated.

Placing players one slot at a time goes wrong in a specific way: the first slot filled takes the best available player, even when he belongs in the next one. On the pitch graphic this showed up as a pair of centre-backs on the wrong sides — the left-centre-back slot, filled first, took the player whose own history said right-centre-back. Scoring every pairing together fixed it.

A second, separate bug had a similar symptom. The Premier League’s own lineup data numbers the players in each row of a formation from right to left, while our templates number them from left to right, so wide players appeared on the wrong flank. Rows are now mirrored as they are read, which also corrected every stored lineup from earlier matches.

Each predicted starter shows a percentage. It is not a calibrated probability. It is 60% plus 35% of the player’s recent start rate, kept between 5% and 99%. A player who has started every recent match shows 95%; a player who has started none shows 60%, because the method picked him anyway — usually to fill a gap left by an absence.

Read it as a ranking of how settled each pick is, not as odds. Turning it into a real probability would need it to be checked against thousands of actual team sheets, and that has not been done yet.

SquadCheck sees only public, official data. It does not know about a knock picked up in training that has not reached the feed, rotation before a European tie, or a manager’s plan for a particular opponent.

We measured how much better last week’s team would do if we knew in advance exactly who was unavailable — an upper limit, because it uses information nobody has before kick-off. It gains about half a starter per match (+0.45 in 2025/26, and +0.49 in each of the two seasons before). That is the most that better injury information could ever add. The rest of the gap between 8.8 and 11 is the manager’s choice, which no public data can see.

The predicted eleven appears on every match page beside both teams’ absences, and on the matchup page for the next round of fixtures.