Estimates the probability that the defending team will regain possession within the next K actions given the current game state, using action features (excluding action type/result to avoid tautology) plus contextual features like score line. This is the "Pre-Recovery" baseline — how likely is the team to win the ball back based on the situation alone?
For each defensive frame: compute features from the game state (ball location, time, score, pressure state). Exclude features that directly encode the outcome (action type, action result) to prevent tautology. Train to predict "ball recovered within next 5 actions." The output is Pre-Recovery probability.