On 28 August 2026 our result model gave Crystal Palace a 51% chance of beating Manchester City. The bookmakers made City favourites at 57%. Both sides had played once, the model had almost nothing to go on — and it leaned on the wrong thing. City won 4–1.
Our match-result probabilities come from a goal model: each side's attack and defence, measured against its league, predict how many goals it should score. Early in a season those measures rest on a game or two, so the model pulls them towards a starting value until real evidence arrives. The starting value was the problem. It was the league average — so in the opening weeks Manchester City and a newly promoted side began from the same place.
Once we looked, the symptom was easy to measure. Across 1,138 priced fixtures with no more than two games played, our home-win probabilities had a spread of 5.1 points; the bookmakers' had 15.3. The model was barely telling fixtures apart. Trusting the early data more, or less, made no difference — the anchor itself did not know who the team was.
Each team now starts from its own previous season — its attack and defence relative to that league — and its current results take over as they accumulate. That works for clubs that stayed in the division. But about a third of fixtures involve a side with no record in that league last season: promoted, relegated, or arriving from a league we do not cover. Where should they start?
We measured every club across five seasons: 4,649 team-seasons of sides that stayed put, 465 that moved between divisions we cover, and 565 that arrived from leagues we do not. The figures are relative to the league average — 1.00 is average, and for defence a figure above 1.00 means conceding more.
| Group | Teams | Attack | Defence |
|---|---|---|---|
| Promoted from the division below | 234 | 0.90 | 1.08–1.09 |
| Relegated from the division above | 231 | 1.06 | 0.94–0.96 |
| Arrived from a league we don't cover | 565 | 0.90–0.91 | 1.10–1.11 |
| Stayed in the same division | 4,649 | 1.01 | 0.98–0.99 |
So a promoted side scores about 10% fewer goals than the league average and concedes about 8% more; a relegated side does roughly the reverse. None of that is surprising — the step up is hard, the step down is easier. The surprise came next.
The obvious refinement is to use a promoted team's record from the division it has just left: surely a champion that scored freely is stronger than a side that scraped through the play-offs. The data says no. Across the 465 moves, a club's attack and defence figures in its old division correlated negatively with the same figures in its new one — between −0.18 and −0.28. Carrying the old numbers over is worse than ignoring them.
The reason is mostly about who moves. Pool the two directions together and the clubs with the best numbers in their old division are overwhelmingly the ones going up, into stronger company, while the clubs with the worst are the ones going down. What survives the move is not how good a team looked — it is which way it went. So that is all the model uses: promoted sides start at 0.90 attack and 1.08 defence, relegated sides at 1.06 and 0.94, and sides from uncovered leagues at 0.91 and 1.10. From there, their own results take over.
We tested the change the way we test everything: a walk-forward replay of 92,451 matches, with the settings fitted on 2023-24 and judged on 2025-26 games they never saw. For season openers, prediction error (log-loss) fell from 1.0812 to 1.0437, and the share of games where our most likely outcome actually happened rose from 42.6% to 46.6%. At matchday one, agreement with the bookmakers' favourite rose from 64.5% to 77.7%. It improved every later stage of the season too, all the way to games 19 and beyond, which is why it stays on all year rather than switching off after the opening weeks.
After the fix, the same model rated Palace v City 28% / 27% / 45%. The 51% we published before kick-off was frozen and graded as a miss, as it should be: a better model never gets to rewrite a worse one's record. That is how our record works for every rating.
Method: walk-forward testing — every prediction is made using only information available before kick-off, then graded against what happened, on the site's own dataset. Questions: contact us.