Selected model
Xgboost
lower validation log loss by 0.008164 versus pytorch_mlp
- Training passes
- 39,214
- Grouped OOF folds
- 5
Passing methodology
Expected Pass model
FootyScout predicts the probability that a pass will be completed using information available about the pass before its outcome.
Expected Goals model →Attacking Impact model →
Architecture
- 01XGBClassifier (315 boosting rounds)
- 02max_depth=4
- 03learning_rate=0.05
- 04objective=binary:logistic · tree_method=hist
Pre-outcome features
Identifiers, player/team names, pass outcomes, recipients, and future events are excluded from model inputs.
Pitch position
start_xstart_yend_xend_ystart_zoneend_zonedistance_to_goal_beforedistance_to_goal_afterdistance_toward_goalPass geometry
pass_lengthpass_angleforward_distancelateral_distancePressure / context
under_pressureprogressivePass characteristics
pass_heightbody_partpass_typeEvaluation performance
OOF probabilities drive player analytics and are highlighted below. The untouched test split was evaluated only after validation selected the model.
Validation · selected form
- ROC-AUC
- 0.9210
- Log loss
- 0.2615
- Brier score
- 0.0810
- Accuracy
- 0.8840
- ECE (10-bin)
- 0.0102
Untouched test
- ROC-AUC
- 0.9158
- Log loss
- 0.2354
- Brier score
- 0.0691
- Accuracy
- 0.9065
- ECE (10-bin)
- 0.0073
Grouped out-of-fold
- ROC-AUC
- 0.9221
- Log loss
- 0.2338
- Brier score
- 0.0706
- Accuracy
- 0.9023
- ECE (10-bin)
- 0.0041
Calibration
Temperature scaling was fit on the validation split only. It was not retained: not retained because validation log loss, Brier score, and ECE did not jointly satisfy the conservative improvement rule
- Temperature
- 0.9849
- Fit split
- validation only
Calibration can improve probability quality without improving ranking; FootyScout does not claim it improves every held-out metric.
Evaluation sequence
- 1. Match-grouped split. grouped by match_id; whole matches; approximately 80/10/10
- 2. Validation selection. Model selection was performed using validation data.
- 3. Untouched test. The test data was not used for model selection.
- 4. Grouped OOF predictions. Player profiles use grouped out-of-fold predictions.
- 5. Production fit. The production full-data model is for future inference and is not used for evaluation claims.
Player Intelligence
Player-style similarity
FootyScout compares eligible outfield players within the same broad position using six style-only passing and carrying tendencies. Each feature is converted to a population z-score within DEF, MID, or FWD, then profiles are ranked by root-mean-square Euclidean distance. The 0–100 index is cohort-calibrated: 100 is identical and 50 is approximately the median same-position pair distance.
Sample support
Support uses the lower observed-match count of the two players. Pairs below three matches are marked limited. Coverage is uneven, so low-sample neighbor ranks may move as more matches are observed; support never changes the score.
What it is not
Similarity describes observed playing style. It is not player quality, future performance, probability of success, tactical fit, or transfer success. Performance metrics, outcomes, team identity, and archetype labels are excluded.
Team Intelligence
Team Intelligence and Role Fit
Team style and positional roles. Team style describes Leverkusen's 34 observed Bundesliga matches. DEF, MID and FWD roles pool the observed events and actions of players assigned to each frozen broad position.
Role Fit. Eligible players and roles share the same six-feature, position-relative playing-style coordinate system. Root-mean-square distance measures observed style resemblance; lower is closer.
Current players. A Leverkusen player is compared with a leave-self-out role so their own actions do not contribute to the target profile.
Scouting Recommendations. External same-position players are ranked only by Role Fit. This is not transfer-success probability, a player-quality score, causal tactical compatibility, or a forecast.