Penetrative Pass Probability Model (P3)

A deep learning model that predicts, at any passing moment, the probability that a penetrative pass is available — regardless of whether the player actually plays it. A penetrative pass is defined as one that ends inside the opponent's convex hull (the polygon formed by all outfield defending players, excluding the goalkeeper), with at least one attacking teammate inside to receive. The model uses CNN (ResNet34 with transfer learning from ImageNet) on custom pitch images encoding player positions, Voronoi space control, and the opponent's convex hull from StatsBomb 360 data. Event data alone is insufficient (AUC ~0.6); 360 data achieves AUC-ROC = 0.73.

(1) Define penetrative passes: a pass whose end location falls inside the opponent's convex hull, with a receiving teammate present inside. Filter to passes originating between the first and last thirds of the pitch (exclude defensive buildup and final-third finishing). (2) Generate pitch images: for each passing moment, create an image encoding: on-ball team (blue), off-ball team (red), passer (black), Voronoi diagram (space control per team), convex hull (green). Normalize orientation so the attacking team always plays bottom-to-top. (3) Train CNN: use ResNet34 pretrained on ImageNet, fine-tune on the penetrative vs. non-penetrative classification task. 8 epochs on GPU (~2 hours). Training loss ~0.3, validation AUC-ROC = 0.73. (4) Set probability threshold: use minimum Euclidean distance from ROC curve to (0,1) point — yields threshold ~0.38. Above threshold = potential penetrative pass available.

Key design decisions: exclude pass end location from features (that would be data leakage — you only know the end location after the pass is made); exclude passer identity and foot (model should be team/player-agnostic); train on Bundesliga, validate on La Liga to prevent data leakage.

1. Including pass end location as a feature: This is data leakage — you only know where the pass went after it was played. The model must predict from the pre-pass state.
2. Using event data alone: Without player positions (360 data), the model can't see the convex hull or space control. Event-only AUC is ~0.6 (near useless).
3. Blaming players for not playing penetrative passes without checking body orientation: A player facing their own goal physically can't see or play a forward penetrative pass, even if the model says the opportunity existed.