Shooting methodology

Expected Goals model

FootyScout estimates the probability that a non-penalty shot becomes a goal using only information available before its outcome.

← Expected Pass model

Selected model

Xgboost

lowest validation log loss, 0.025238 below logistic_regression. Test metrics used for selection: no.

Modeled shots: 5,545
Grouped folds: 5

Training corpus

233 matches · 5,784 shots · 685 goals

penalties and period-5 shootout shots are preserved but excluded from modeling and player aggregates

Competition 9 · season 281 · 34 matches
Competition 43 · season 106 · 64 matches
Competition 55 · season 282 · 51 matches
Competition 223 · season 282 · 32 matches
Competition 1267 · season 107 · 52 matches

Pre-outcome features

Player/team identity, shot outcome, goalkeeper outcome, end location, future events, and StatsBomb’s provider xG are excluded.

shot_xshot_ydistance_to_goalangle_to_goalunder_pressurefirst_timeone_on_oneopen_goalbody_partshot_typetechniqueplay_pattern

Evaluation performance

Selection used validation log loss. The test split remained untouched until the model and calibration decision were frozen.

Validation · selected form

ROC-AUC: 0.7219
Log loss: 0.2859
Brier score: 0.0806
Accuracy: 0.9042
ECE (10-bin): 0.0376

Untouched test

ROC-AUC: 0.7687
Log loss: 0.2507
Brier score: 0.0693
Accuracy: 0.9136
ECE (10-bin): 0.0192

Grouped out-of-fold

ROC-AUC: 0.7528
Log loss: 0.2703
Brier score: 0.0760
Accuracy: 0.9075
ECE (10-bin): 0.0102

Model feature importance

Calibration

not retained: the conservative material-improvement rule was not satisfied. Temperature: 0.9943; retained: no.

Grouped methodology

match-grouped deterministic approximately 80/10/10 split. Preprocessing is fitted only on training matches, including independently inside each OOF fold.