Valuing Actions by Estimating Probabilities
VAEP°An estimate from a model we built. Useful, but not a fact — every modeled stat's page says where it's weak.
Definition
VAEP values every on-ball action, not just shots. Two models estimate, from the game state before and after each action, the probability that the team scores next and the probability that it concedes next. An action's value is the change in scoring probability minus the change in conceding probability. A player's VAEP is the sum across their actions.
How it is computed
Actions are converted to a uniform SPADL representation, then scored by paired scoring/conceding models over the resulting game states.
How much is enough
VAEP per 90 is usable from roughly 600 minutes. Season totals are dominated by playing time before that.
Where it is weak
Systematically generous to players who touch the ball often in the middle third, and thin on off-ball contribution it cannot see. Goalkeeper VAEP is not comparable to outfield VAEP.
What it refuses to measure
VAEP does not measure defensive positioning, off-ball running, or anything that does not involve touching the ball. A player can be enormously valuable and score near zero.
What would prove this wrong
If ordering players by raw touch count predicts next-season VAEP as well as VAEP itself does across a full release, the model adds nothing over volume and should be retired.