05 / Transferability Complete
What happens to attacking output when a player changes leagues?
Similarity can describe a player.
Transferability asks whether his output survives the move.
Finding
Most output survives a move on average (97% vs comparable stayers), but one league-difficulty ladder decides the direction. A model of the player's own history, age and clubs beats “he'll do what he did” on unseen seasons; similarity adds nothing to it.
Every summer, clubs pay for output produced somewhere else. Whether that output travels is the question behind every cross-league signing.
Study 4 can say who a player resembles. It can't say what happens after he moves.
How much of a player's output survives a league move — and can we predict it?
The benchmarkWe compare movers against — same club, role, output and age — so that ordinary isn't mistaken for a transfer effect.
Output drops after any strong season, move or no move. Comparing movers with similar players who stayed separates the move from that ordinary drop.
Understat has no transfer records, so moves were inferred from club changes (including deadline-day moves after early matches). Of 2,917 cross-league moves, the primary population is 428 summer moves (2016/17–2025/26) by forwards and midfielders with ≥900 minutes before and after; every exclusion is listed with its reason.
= output after ÷ output before (xG + xA per 90). Dates of birth were linked from the CC0 Transfermarkt-derived dataset (100% of the 428 moves).
Every prediction uses only information available at the decision date; the final model was scored once on .
Movers keep 97% of their output compared with comparable stayers (95% 93–101%): an average move costs about 3%.
Output kept after a move, vs comparable stayers
xG + xA per 90. 100% = no change relative to similar players who stayed. Line is the 95% interval.
Movers keep 97% (95% interval 93–101%): an average move costs about 3%, and the interval just reaches 100%.
Study 05 · 428 summer moves, 2016/17–2025/26
One explains all 20 directions: moving into the Premier League keeps 82%; moving out of it, 117%.
Premier League hardest, Serie A and La Liga in the middle, Bundesliga and Ligue 1 easiest. Pair-specific effects add nothing (p = 0.91): Serie A → Premier League keeps ~80%, Premier League → Serie A ~114%.
Output kept moving into vs out of each league
vs comparable stayers, xG + xA per 90. 100% = no change.
| League | Moving into | Moving out of |
|---|---|---|
| Premier League | 82% (76–87) · n 137 | 117% (107–127) · n 78 |
| La Liga | 100% (92–108) · n 86 | 100% (91–109) · n 86 |
| Serie A | 96% (89–103) · n 97 | 99% (91–108) · n 69 |
| Bundesliga | 119% (107–130) · n 47 | 90% (82–98) · n 75 |
| Ligue 1 | 118% (108–129) · n 61 | 89% (82–95) · n 120 |
Moving into the Premier League costs the most and moving out of it gains the most; the Bundesliga and Ligue 1 are the reverse.
Study 05 · 428 moves; per-league 95% intervals in the table below
Younger players keep more — about −2% per year of age (U21 109%, 30+ 93%). The destination club matters as much as the league: weakest third of destination clubs 79%, strongest third 106%.
Players coming off a jump in output drop more (−9% per SD of last-season change). No effect: role (once regression to the mean is removed), consistency over three seasons, profile distinctiveness, penalty duties, loans.
Output kept vs comparable stayers, by age and destination club
xG + xA per 90. 100% = no change.
Age at the move
Destination club strength
Under-21s keep 109%, players over 30 93%. Joining a club in the weakest third keeps 79%; the strongest third, 106%.
Study 05 · 428 moves
Each step adds information the naive guess ignores. The final model cuts the error from 0.129 to 0.096 and lifts from 0.58 to 0.76.
Prediction error on 142 locked test moves, step by step
First season after the move, xG + xA per 90. Lower MAE is better; higher R² is better.
Each step adds information and lowers the error: 0.129 → 0.117 → 0.102 → 0.096.
Study 05 · locked test, 2023/24–2025/26
The final model was fixed before the test seasons and evaluated once on 142 unseen moves. It beat the naive baseline by 0.033 xG + xA per 90 (CI 0.018–0.049) and did slightly better on the unseen seasons than on earlier ones — no sign of overfitting.
Adding Study 4's similarity — what the most similar destination-league players produce, match quality, distinctiveness, how similar movers fared — changed the error by 0.000. Comparables alone were worse than the player's own history.
Prediction error on 157 league moves
Mean absolute error, xG + xA per 90, first season after the move. Lower is better.
Similar players alone barely beat the naive guess. The player's own history cuts the error by 17%; adding similarity on top moves it by less than 0.001.
Study 05 · critical experiment, transfers with Study 4 profiles
The chance of keeping ≥75% ranks players well ( 0.80) and beats a same-for-everyone base rate ( 0.179 vs 0.224), and is at the top. But individual forecasts are uncertain: a typical is ±0.2 xG + xA per 90.
Lower thresholds (50%, 60%) were not well calibrated and are not used.
Is the chance of keeping ≥75% honest?
Five groups of about 29 test moves. On the diagonal = says what it means.
Where the model says about 82% and 96%, 86% and 97% of players kept ≥75%. Lower down it is less reliable. AUC 0.80; Brier 0.179 vs 0.224 for a same-for-everyone guess.
Study 05 · locked test, calibration by fifths
Known weakness
The league effect is a fixed amount, not a percentage, so low-output central midfielders moving into the Premier League are under-predicted by about 0.06 xG + xA per 90.
Example: Manuel Ugarte
Manchester United case study, summer 2024 — standing at 1 June 2024 with only what was known then.
xG + xA per 90 · Study 05 case study
Ugarte was bought for defensive work this data cannot see. The model primarily sees attacking output. A percentage-based league effect is the next version.
What the raw data suggested
Early on, centre-forwards seemed to drop the most after a move, and a promoted-club effect looked like a data artefact.
What happened after we checked
Against comparable stayers, centre-forwards keep 96% — mostly regression to the mean, no clear role effect. And the promoted-club effect is real: players joining promoted clubs produce more than the club's rating predicts.
Lesson
Every raw transfer pattern needs a control group — and a second look.
SourcesUnderstat · football-data.co.uk · Open-Meteo ERA5 · Wikipedia/Wikidata · OpenStreetMap
Analysis periodMoves 2016/17 — 2025/26 · frozen dataset · Methodology
08The next question
Five studies, one answer: context belongs to the match; level belongs to the player.