What happens to a player’s statistics when he changes leagues?
A preprint by Mohammad Arshan Shaikh attempts to answer this using machine learning and conformal prediction.
The question is simple:
If a player produces X per 90 in Ligue 1, what can we expect from him in the Premier League?
The answer shouldn’t be a single multiplier. We also need to understand the uncertainty range. 👇
Aug 14, 2026 · 5:55 PM UTC
3
6
58
4,562
A key pass per 90 in Ligue 1 doesn’t necessarily mean exactly the same thing as one in the Premier League.
The pace, pressure, teammates, playing style and the player’s role all change.
That’s why the paper doesn’t try to determine whether a player is “good” or “bad”.
It attempts to estimate how his statistics might change when he moves into a different competitive context.
1
3
382
To study this, the authors track players who played at least 450 minutes in one league and, the following season, at least 450 minutes in another.
This creates “bridges” between competitions: they observe what actually happened after the move.
They analyse 1,040 cases between 2017/18 and 2024/25, separating forwards, midfielders and defenders.
They use 12 metrics, includingquw key passes, progressive passes and carries, shot-creating actions, tackles, interceptions, clearances and blocks.
1
3
267
The model uses the player’s original statistic per 90, age, minutes played, league of origin, destination league and the difference in UEFA level.
It then compares several methods: CatBoost, Random Forest, XGBoost, linear models and a neural network.
But the most important part comes next: it doesn’t provide only a central prediction. It also provides an interval of plausible outcomes.
1
1
203
Example:
A player records 4.2 key passes per 90 in his original league.
The model estimates that he will produce 85% of that figure in the new league:
→ Prediction: 3.6 key passes/90
→ 90% interval: approximately 2.2–5.8/90
This interval does not mean that the player has exactly a 90% probability of falling within it.
It means that, across comparable cases, the method aims to cover approximately 90% of the actual outcomes.
1
1
214
The results are promising:
• The intervals achieve at least 90% coverage for 9 of the 12 statistics.
• Average coverage is 93.2%.
• Defensive metrics translate better across leagues than key passes or progressive carries.
However, some league routes have very few cases, roles are grouped together, and the model does not include the coach, team or tactical system.
The scouting lesson:
Don’t ask only, “How much will he produce?”
Also ask, “What range is reasonable, and how much evidence do we have?”
This is a concept that could be highly relevant for FBPlot: showing current performance, projection and uncertainty in the same chart.
Full article: sportrxiv.org/index.php/serv…
2
5
230
