Selected model
Xgboost Reg Pseudohubererror
lowest validation RMSE among the global mean, direct XGBoost objectives, and justified hurdle candidate
Objective: reg:pseudohubererror · selected using Validation Rmse
Possession-value methodology
Attacking Impact model
Estimate how much a pass or carry changes the future expected-goal value of the current possession.
← Expected Pass model·Expected Goals model
Target and state
sum of later FootyScout OOF xG in the same possession; an expected-goal quantity, not a scoring probability
state immediately before an on-ball event; the current event has not occurred and belongs to the future
Training corpus
233 matches · 823,553 events · 667,062 states · 38,419 possessions
- Germany — 1. Bundesliga 2023/24 · 34 matches · 137,765 events
- FIFA World Cup 2022 · 64 matches · 234,637 events
- UEFA Euro 2024 · 51 matches · 187,924 events
- Copa América 2024 · 32 matches · 100,324 events
- African Cup of Nations 2023 · 52 matches · 162,903 events
Grouped methodology
reused frozen V2.1 match-grouped 80/10/10 allocation
Nested upstream labels: inner grouped xG OOF on fixed train matches; held-out validation xG predicted from fixed train only. Held-out outcomes used for training: no.
- Train
- 534,490
- Validation
- 63,763
- Test
- 68,809
reg:pseudohubererror · 300 boosting rounds · 38 encoded features
Validation comparison
The selected model minimizes validation RMSE; the untouched test set is reported only after selection.
Global Mean
- RMSE
- 0.0589
- MAE
- 0.0298
- R²
- -0.0009
- Spearman
- 0.0000
- Positive RMSE
- 0.1440
Xgboost Reg Squarederror
- RMSE
- 0.0568
- MAE
- 0.0271
- R²
- 0.0707
- Spearman
- 0.2419
- Positive RMSE
- 0.1377
Xgboost Reg Pseudohubererror
- RMSE
- 0.0567
- MAE
- 0.0268
- R²
- 0.0735
- Spearman
- 0.2446
- Positive RMSE
- 0.1376
Xgboost Reg Tweedie
Frozen evaluation
Validation · selected
- RMSE
- 0.0567
- MAE
- 0.0268
- R²
- 0.0735
- Spearman
- 0.2446
- Positive RMSE
- 0.1376
Untouched test
- RMSE
- 0.0506
- MAE
- 0.0238
- R²
- 0.0708
- Spearman
- 0.2179
- Positive RMSE
- 0.1247
Grouped OOF
- RMSE
- 0.0516
- MAE
- 0.0249
- R²
- 0.0787
- Spearman
- 0.2322
- Positive RMSE
- 0.1237
Leakage-safe features
OOF xG only for labels; no future fields, outcomes, provider xG, identities, or next actions in model inputs
ball_xball_ydistance_to_goalangle_to_goalperiodminutescore_differencepossession_action_numberpossession_elapsed_timepossession_start_xpossession_start_ydistance_progressed_from_possession_startprevious_action_successunder_pressureprevious_action_typecurrent_play_patternDefinitions
- State value
- Future FootyScout OOF xG remaining in the same possession.
- Action value
- V(after) − V(before); not the probability that an action succeeds.
Limitations
- observational/model-derived, not causal
- on-ball attacking possession value only
- off-ball movement is weakly represented
- defensive value is outside this model
- FootyScout xG has documented upper-tail compression
- sparse rare states are less certain
- product profiles cover the available Bundesliga 2023/24 cohort