Football/ Ayush Pawar

Case study · Manchester United · 2024

What would the model have told Manchester United?

A leakage-free retrospective of the club's 2024 recruitment, using only information that was available on 1 June 2024.

As-of date

Manchester United are about to enter the summer transfer window. What would this research have said about the players they could consider?

No hindsight.

The analysis uses only information that would have been available as of 1 June 2024.

Available

  • Pre-transfer player history (output up to 2023/24)
  • League context (league effects fitted on earlier moves)
  • Player age at the move
  • Destination context (Manchester United's 2023/24 club strength)
  • A model trained on 334 earlier moves only — none after June 2024

Not available

  • Post-transfer performance
  • Future seasons
  • Actual transfer outcome
  • Any information from after 1 June 2024
  1. Information cutoffEverything after this date is hidden from the model.
  2. Model generates predictionsEvery candidate scored for Manchester United, with an 80% range.
  3. Transfers happenPlayers move; the model is not refitted.
  4. Post-transfer output observedOnly now is the new season looked at.
  5. Prediction vs realityInside or outside the range, judged with at least 450 minutes.

The recruitment board

Manchester United needed attacking and midfield recruitment. The candidate pool: every player in La Liga, Serie A, the Bundesliga and Ligue 1 with ≥ 1,500 minutes in 2023/24 whose Study 4 role is centre-forward or central / defensive midfielder.

Recruitment questionWho would the model have expected to translate successfully?

Strikers · 106 candidates

Target profile: Rasmus Højlund, 2023/24 · ranked by predicted Premier League output at United.

  1. 1Victor BonifaceBayer Leverkusen0.99
  2. 2Kylian Mbappe-LottinParis Saint Germain0.90
  3. 3Serhou GuirassyVfB Stuttgart0.83
  4. 4Artem DovbykGirona0.77
  5. 5Deniz UndavVfB Stuttgart0.77
  6. 48Joshua ZirkzeeBologna · signed0.46

Number = predicted xG + xA per 90 at United in 2024/25.

Midfielders · 206 candidates

Target profile: Casemiro, 2023/24 · ranked by similarity to Casemiro (a holding midfielder's job is mostly outside this data, so Study 4 builds this shortlist; the forecast is the risk column).

  1. 1Bryan CristanteRoma0.05
  2. 2Yangel HerreraGirona0.18
  3. 3Roberto GagliardiniMonza0.13
  4. 4Mikel MerinoReal Sociedad0.21
  5. 5Neil El AynaouiLens0.14
  6. 138Manuel UgarteParis Saint Germain · signed0.02

Number = predicted xG + xA per 90 at United in 2024/25.

Study 05 · Milestone 8 recruitment board, as of 1 June 2024 · the full board is at the bottom of the page

Case 01 · Striker · from Bologna

Joshua Zirkzee

Model prediction
0.46xG + xA / 90 at United
Actual 2024/25
0.471,378 minutes
Board rank
48 / 106by expected output

The actual output landed almost exactly on the model's prediction.

He kept his Bologna output almost exactly (0.49 → 0.47). But June-2024 data did not make him a standout: he finished 48th of 106 strikers by expected output, and his profile was the least like Rasmus Højlund's on the board (closer than only 1% of the pool). The board pointed to higher-output options — availability, price and fit are outside the data.

Zirkzee · prediction → actual

xG + xA per 90 · band = 80% prediction range

Predicted 0.46 (range 0.35–0.68); actual 0.47 — inside the range.

What signal did the model provide?
Output expectationHow much attacking output might survive?
0.46 xG + xA per 90 at United, from 0.49 in 2023/24 (keeps 94%); chance of keeping ≥ 75%: 87%.
UncertaintyHow wide is the prediction range?
80% range 0.35–0.68 (width 0.33) · evidence: 2 seasons of qualifying history.
ContextHow difficult is the destination?
Into the Premier League, players keep 82% of their output on average (76–87%); United's club strength is from 2023/24.
LimitationsWhat is missing?
Link-up play, pressing and dropping deep are not in xG + xA. One season at United, 1,378 minutes: a single outcome, not a verdict on the signing.

Case 02 · Central midfielder · from PSG

Manuel Ugarte

Model prediction
0.02xG + xA / 90 at United
Actual 2024/25
0.141,802 minutes
80% range
0.01–0.05Outside expected range

At first glance, a model failure. That is exactly why it is shown.

Ugarte · prediction → actual

xG + xA per 90 · band = 80% prediction range · note the smaller scale

Predicted 0.02 (range 0.01–0.05); actual 0.14 — outside the range.

Why Ugarte matters

Ugarte exposes two limits of the model, and the research named both before this outcome was known:

  1. A known bias. The Premier League adjustment is a fixed amount, so it wipes out most of a low-output player's forecast. On the locked test, central and defensive midfielders moving into the Premier League were under-predicted by about 0.06 xG + xA per 90.
  2. A blind spot. The model forecasts attacking output only. The data has no tackles, interceptions, completed passes, carries or possession — and Ugarte was bought for ball-winning.

So a player can beat the model's attacking-output expectation without the model being "wrong" about his overall football value — and the forecast was never the right tool for judging a holding midfielder.

What signal did the model provide?
Output expectationHow much attacking output might survive?
0.02 xG + xA per 90 at United, from 0.07 in 2023/24 (keeps 23%); chance of keeping ≥ 75% not shown (2023/24 output under 0.10 — retention of tiny numbers is meaningless).
UncertaintyHow wide is the prediction range?
80% range 0.01–0.05 (width 0.04) · evidence: 1 season of qualifying history.
ContextHow difficult is the destination?
Into the Premier League, players keep 82% of their output on average (76–87%); United's club strength is from 2023/24.
LimitationsWhat is missing?
Ball-winning and build-up work are invisible to xG + xA. Midfield forecasts into the Premier League run about 0.06 low; compare midfielders with each other, not with strikers.

The real test

Candidates who joined Premier League clubs

For every board player who moved to the Premier League that summer, the forecast was re-made for the club he actually joined, using the same June-2024 information. A forecast is judged only with at least 450 minutes.

4 / 4

The striker predictions held up.

All 4 evaluated striker transfers landed within the model's 80% prediction range (mean absolute error 0.052). One retrospective evaluation — not proof of predictive success.

Strikers who joined Premier League clubs: predicted xG + xA per 90 with 80% range, actual 2024/25, and result
PlayerPredicted80% rangeActual2024/25ChartResult
Federico ChiesaJuventus → Liverpool · 95′0.540.41–0.800.78Not judged · under 450′
Joshua Zirkzee · signed by UnitedBologna → Manchester United · 1,378′0.460.35–0.680.47Within range
Niclas FüllkrugBorussia Dortmund / Werder Bremen → West Ham · 775′0.420.32–0.630.49Within range
Jørgen Strand LarsenCelta Vigo → Wolves · 2,624′0.410.31–0.610.52Within range
Iliman NdiayeMarseille → Everton · 2,478′0.310.23–0.460.33Within range

5 / 6

Midfield was harder.

Five of six evaluated midfielders landed within the model's 80% prediction range (mean absolute error 0.090). The forecasts sit visibly low — the known bias — and Ugarte is the exception.

Midfielders who joined Premier League clubs: predicted xG + xA per 90 with 80% range, actual 2024/25, and result
PlayerPredicted80% rangeActual2024/25ChartResult
Ilkay GündoganBarcelona → Manchester City · 2,231′0.360.16–0.850.29Within range
Mikel MerinoReal Sociedad → Arsenal · 1,578′0.280.12–0.650.52Within range
Daichi KamadaLazio → Crystal Palace · 1,527′0.230.10–0.550.17Within range
Manuel Ugarte · signed by UnitedParis Saint Germain → Manchester United · 1,802′0.020.01–0.050.14Outside expected range
Boubakary SoumareSevilla → Leicester · 2,182′0.010.01–0.050.02Within range
Guido RodríguezReal Betis → West Ham · 1,179′Forecast floored at 0.00 by the pipeline, below its own range — the additive league effect pushed it under zero.0.000.01–0.050.03Within range

Ring = prediction, band = 80% range, dot = actual 2024/25 · xG + xA per 90 · Study 05 Milestone 8

The biggest lesson

A recruitment model can be useful without being a complete player model.

The case study shows both sides of the system. It can make remarkably accurate attacking-output predictions for some transfers, while missing important dimensions of players whose value isn't primarily captured by attacking statistics.

How the studies fit together

Finding a player and forecasting him are different jobs.

Study 4 finds statistically similar players; Study 5 estimates whether their output is likely to survive a league move. In this case study the midfield shortlist comes from Study 4 and the forecast from Study 5. Similarity finds candidates — it adds nothing to the forecast itself.

Study 3 explains why raw output is never read on its own. But for forecasting a move, the locked tests found that adjusting for opponents faced or home share did not improve the forecast; what carries the signal is the player's own three-season history, his age and the strength of the two clubs and leagues.

Study 04 · Similarity

“Who looks like him?”

Study 05 · Transferability

“What happens when he moves?”

Forecast inputs

Three-season output · age · both clubs' strength · both leagues

Method

How was this evaluated?

What we knew — and deliberately didn't — as of 1 June 2024

Known

  • Each player's output and minutes up to 2023/24 (three-season level where available)
  • League effects and club-strength effects, fitted on 334 moves whose first destination season ended by June 2024 (2016/17–2023/24)
  • Age at the move; the destination club's 2023/24 strength
  • Prediction ranges from the model's out-of-sample errors in rolling folds 2019/20–2023/24

Deliberately not known

  • Any 2024/25 performance
  • Later seasons
  • Who actually signed whom — used only afterwards, to choose which club to re-score a player for, still with 2023/24 club strength
Is this a new model?

No. The case study is an application of the locked Study 5 model (Model A: context + three-season level), refitted only on moves available by June 2024. Predictions were generated using information available before the transfer window and compared with post-transfer output afterwards. No forecast on this page was changed after the outcome was known.

Locked test · overall model, not this case study

Unseen moves
142
MAE
0.096
R²
0.76
AUC · keeps ≥ 75%
0.80

The model fixed before the test, scored once on moves in 2023/24–2025/26. Its 80% ranges covered 87% of outcomes. These are the model's overall numbers; the Manchester United results above are a separate, much smaller retrospective.

Limitations

What this case study cannot tell us

This is a performance-output model, not a complete recruitment department.

  • Whether United should have bought a player
  • Whether a player was tactically suitable
  • Defensive contribution
  • Injury risk
  • Personality and adaptation
  • Transfer economics
  • Contract situation
  • Squad fit

The full board

Every candidate, as of 1 June 2024.

Brief

Loading the board…

Predicted = expected xG + xA per 90 at Manchester United in 2024/25, with the 80% range. Retention = predicted ÷ 2023/24 output. P(≥75%) is blank when 2023/24 output was below 0.10. Similarity = closer than X% of the candidate pool to the target profile (strikers: Rasmus Højlund 2023/24; midfielders: Casemiro 2023/24). Actual = 2024/25 output wherever the player played; blank if he left the five leagues or didn't play — outcomes outside the Premier League are not a test of the forecast.

Want to run the model yourself?

Try the Transfer Calculator Find Similar Players

Or go back to the research.