Promoted survival
67%
stay up in year one
Cross-league sports analytics · 1993/94–2025/26
Every season, three clubs win promotion to one of Europe's big leagues. They were the best of the division below; within weeks they are the smallest names in the one above. This study asks whether we can tell, before a ball is kicked, which of them will survive. And whether a single model reads promoted clubs as well as it reads everyone else.
One model, two worlds. The gap between these two numbers is what this study is about.
Going up is one of the hardest jumps in sport. A club that finishes near the top of the second tier suddenly has to make its way in a richer, faster league where it is one of the smaller names. The Premier League made the point twice in a row. In 2023/24 all three promoted clubs went straight back down, and the next season it happened again. People began to talk, once more, about a gap between the two divisions that keeps widening.
Whether promoted clubs struggle is not really the question; they plainly do. What we want to know is whether the struggle shows up beforehand in the sort of data anyone can find, and whether a model built on the whole league reads this small group as well as it reads the rest. Promotion turns out to be a clean natural experiment in what statisticians call domain shift, and we look at all five of the big European leagues. The dataset is set out under Data.
A league has about twenty clubs, and only three of them are newly promoted in any season. So a model fitted to the whole league mostly learns from the established majority, and its headline accuracy says very little about the three newcomers it scores with the same rule. Promoted clubs are an unusually tidy example of the problem. They are clearly labelled, they come round every year, and they sit inside a league that is otherwise the same.
The reasoning and the models are laid out under Methods.
Trained on the whole of Europe and tested only on seasons it had not seen, the model looks good on the average. Then you split it by group, and a gap opens. For established clubs the predicted chances match what actually happened, and the calibration line sits on the diagonal. For promoted clubs the line goes flat. Whatever number the model gives, the truth stays near the base rate, so the number is telling you almost nothing.
Calibration asks whether the percentages are honest: of all the clubs given, say, a 70% chance, do about seven in ten really survive? AUROC asks whether the model can tell clubs apart. Take one club that stayed up and one that went down, at random; it is the chance the model gave the survivor the higher score. 1.0 means it gets every such pair right, and 0.50 is a coin toss. The first chart below is calibration, the second is AUROC.
Dots are sized by how many clubs fall in each band. The teal line tracks the diagonal. The clay line is nearly flat, which means a promoted club's predicted chance barely moves the odds of it actually staying up.
The same model scores 0.79 overall and 0.80 on established clubs, then drops close to a coin toss on promoted ones.
This is not a quirk of the survival measure. Points and final position tell the same story: plenty of skill on established clubs, next to none on promoted ones.
| Target | Established | Promoted |
|---|
"Skill" means how much better the model does than simply guessing the group average. For promoted clubs the points model lands at about minus one per cent, which is to say no better than giving every promoted club the same average score.
Train on four leagues, predict the fifth, and the picture holds. Survival is fairly easy to call overall, yet for promoted clubs it sits near a coin toss in every country. So this is not an English habit; it is how the thing works across Europe.
In each league the dark bar, for all clubs, clears 0.74. The clay bar, for promoted clubs only, hugs the 0.50 line. England is also the hardest place to stay up, with barely half of its newcomers surviving.
For the league as a whole, survival rests almost entirely on one thing: how a club's strength rating compares with the rest of the division. Strong clubs stay up, weak ones go down. The trouble is that promoted clubs are all bunched at the bottom, with nearly nine in ten below the middle of their new league. The one feature that separates everyone else has hardly any room to work inside this little group, so which particular promoted club survives comes down mostly to luck.
Promoted clubs, in clay, pile up below the established crowd. On average they sit −76 rating points below the middle of their league, where established clubs sit +38 above it.
For the pooled model, the strength gap towers over everything else. For a model built on promoted clubs alone, every feature flattens to zero. Adding transfer spend, squad make-up and expected goals does not help. The wall is real, not a gap in the data.
We learned a picture of playing style from three million on-the-ball events. The main axis, which captures almost half of what sets clubs apart, runs from a possession game, with lots of accurate passing high up the pitch and plenty of shots, to a direct, deep, long-ball one. Promoted clubs gather at the long-ball end. So the change at the boundary is not only that these sides are weaker; they play the game differently.
Hover over any club. The promoted sides, labelled in clay, sit low and to the left: fewer passes, less accuracy, fewer shots, more route one.
We point the same model at next season's promoted clubs. Since it can barely tell them apart, the chances all land near the base rate, each with a wide margin around it. Hull come up with the weakest rating of the three English newcomers, yet the model gives them much the same chance as the other two. The forecast page sets out every club with its margin, ranks the whole league by relegation risk, explains the method and checks it against last season.