Expected Ball Progression Model

An alternative to goal-based and shot-based models that measures how far the ball advances toward goal during each possession. Every possession has a "high-water mark" — the closest point to goal the ball reaches. With ~200 possessions per game, this produces hundreds of meaningful training signals on a continuous scale, compared to ~2-3 goals. The insight: "when a team gets the ball two-thirds of the way up the field, that means something — they might not score, but it means something." This model captures buildup quality independently of whether it converts to shots.

For each possession, record the maximum ball penetration (closest distance to goal). Each event in the possession is trained against this high-water mark — events in the early phase of a possession that reaches deep into the opponent's third are labeled as higher-value than events in possessions that stall at midfield. Two-phase possessions (build up deep, recycle, build up again) use separate high-water marks for each phase: the first phase trains on its own deepest point, the recycled phase trains on its deepest point.

Key advantage over EPV: turnovers don't necessarily destroy value. If a team chips the ball into the box and the defender heads it out, the ball is still in the attacking third — the high-water mark persists. The ball rarely teleports back to halfway on a turnover. This makes ball progression less sensitive to the possession-definition problem than EPV.

1. Treating high-water mark as a substitute for goal probability: Ball progression measures buildup quality, not scoring likelihood. A team can consistently advance to the final third and still not score. Use it as a complementary signal.
2. Ignoring the pitfall case: "Getting close but no closer is rubbish" — Burnley with 11 behind the ball, the ball reaches the final third but there's no penetration. Pure distance doesn't capture this. Consider weighting by defensive density at the high-water point.