What a possession-value model measures and what it leaves out

The basic idea

A possession-value model estimates how likely a scoring event is from any given situation and then values each action by how much it changed that likelihood. This converts a match into a sequence of small credits and debits, which allows contributions away from the final action to be recognised. The approach solves a real problem, since traditional counting rewards only the last touch and ignores everything that created the situation.

It also produces a single currency in which very different actions can be compared, which is convenient and quietly consequential. The convenience is the source of most subsequent misunderstanding, because a single number invites comparisons the model was not built to support.

Value is defined narrowly

The value in these models is defined entirely in terms of proximity to a scoring event, so anything not affecting that proximity is worth nothing. A defensive action that prevents a dangerous situation is credited only if the model represents the counterfactual, which many implementations handle poorly. Actions that reduce risk without advancing the ball tend to be valued negatively or not at all, despite being correct decisions in context.

This produces the familiar complaint that such models undervalue players whose contribution is stability rather than progression. The complaint is accurate, and it is a consequence of the definition rather than a flaw in the fitting of the model.

Credit assignment is genuinely hard

When a sequence produces a good situation, the model must distribute credit across everybody who contributed to it, and there is no objectively correct split. Common approaches credit each action by the change in likelihood it produced, which concentrates value in the actions closest to the final situation. That concentration understates the work of players who created space or moved an opponent without ever touching the ball.

Some models attempt to include off-ball contribution, which requires tracking data and a further set of assumptions about what off-ball value means. Each additional assumption improves realism and reduces transparency, which is a trade every analytics department negotiates differently.

The road not taken is invisible

A model values the action that occurred and knows nothing about the alternatives that were available at that moment. A pass valued highly may have been the third-best option available, and a pass valued poorly may have been the only sensible choice. Evaluating decisions rather than actions requires knowing the option set, which is a substantially harder problem and requires far richer data.

Without it, a possession-value figure describes outcomes rather than judgement, and players who take low-percentage options are penalised only when they fail. This is why such models are more reliable in aggregate across many actions than they are as an assessment of any individual decision.

Reading published figures sensibly

A possession-value number is most useful as a description of what a player did contribute rather than of how good the player is. Comparing across roles is the most common misuse, since roles differ in how much opportunity they provide to accumulate the kind of value being measured. Comparing across competitions adds another distortion, because the underlying likelihood estimates were fitted on data from particular contexts.

Used within a role, within a competition and across a reasonable number of matches, the figures are informative and reasonably stable. Used as a general ranking of ability, they reproduce the assumptions of their own definition and present them as a conclusion.