The pass was available before the receiver looked open. By the time the closeout appeared on the broadcast view, the useful window was already narrowing. Decision analysis lives inside that interval.
Tracking makes it possible to describe the state of a possession as it changes: player locations, movement, ball position and, when available, game context. A model can use those observations to estimate possible outcomes. It cannot simply read a player's mind or replay an unchosen action as if the alternative had actually happened.
Separate the choice from the result
A contested shot can go in. A well-timed pass can be dropped. If we grade every decision from the final score, we reward some mistakes and punish some useful actions. A more careful review asks what options were reasonably available at the moment of choice, then considers execution and outcome separately.
This distinction matters for off-ball players too. A cutter may create an opening without receiving a pass. A corner player may hold their position and prevent a defender from helping. Neither contribution is captured by describing only the final shot. Tracking can draw attention to these changes, but their interpretation still depends on the coverage and the team's intended action.
What expected value actually means
An expected possession value model estimates an average future outcome conditional on the information it receives. It may account for player and ball locations, time remaining, identities and other features. Its estimate is not a promise about this particular possession. Nor does a difference between two model estimates establish a certain tactical truth.
Dean Oliver's Basketball on Paper and Kirk Goldsberry's SprawlBall offer useful wider context for thinking about basketball efficiency and space. They do not provide a universal decision rule for every team. A model trained on NBA possessions may not transfer cleanly to a local youth team, where passing range, shooting ability and defensive behaviour differ substantially.
Rewind before comparing the options
- Before the help commits.In this illustrative possession, a ball handler turns the corner while a teammate stays available on the opposite wing.
- At the commitment.A help defender steps toward the lane. Pause before the ball handler gathers, and identify which passes can still be delivered.
- After the choice.Replay the pass, shot or continuation. Review the decision timing separately from the quality of execution.
A static frame can exaggerate an option. The receiver might be moving behind a defender; the passing hand might be occupied; the ball handler might have already committed to a gather. Body orientation, reach and ball control are not always captured accurately by a basic position feed. The question is whether the option was usable then, not merely drawable later.
The unchosen action is uncertain
To compare a pass with a drive that never occurred, a model must estimate a counterfactual. That requires assumptions about how defenders and teammates would respond. Defenders do not remain frozen while an imaginary pass travels. A credible analysis should acknowledge this uncertainty rather than presenting a single alternative outcome with the confidence of an observed event.
Evaluation should also avoid using future information to judge an earlier moment. If a system's features accidentally reveal what happened after the decision, its apparent predictive performance can become misleading. Testing on genuinely unseen possessions, separating relevant datasets and inspecting failures are practical safeguards. Coaches should be told what information the model used.
A useful review asks, “What could you see here?” before asking, “Why did you not pass?”
Make the feedback playable
Instead of handing a player a list of supposedly missed opportunities, choose a small number of clear clips. Show the action at normal speed, ask what the player noticed, then revisit the moment before the window closed. A cue such as “read the low defender before the gather” gives the player something to practise. “Choose the highest value option” rarely does.
Keep disagreements visible. A model may favour an option that conflicts with a play call or a player's current skill. That disagreement can expose a missing feature, a coaching preference or an area for development. The aim is not to make the estimate the final authority. It is to improve recognition while preserving room for context, conversation and basketball judgement.