Contextual models can reveal value hidden by traditional statistics, but model definitions still encode a theory of what winning basketball looks like.
What tracking adds
Player and ball tracking can describe spacing, speed, defensive pressure and the sequence of actions within a possession. That helps analysts study screens, rotations and gravity that a traditional event log misses.
Tracking is measurement, not meaning. Analysts still decide which patterns matter for a team's scheme.
From win probability to credit
The NBA and AWS have described an architecture that separates win-probability prediction from credit attribution. This allows the system to estimate how actions changed the expected outcome and then distribute credit without retraining the predictive model for every explanation.
That distinction is useful: predicting the possession outcome and explaining player contribution are related but separate problems.
The hard cases
A pass can create an advantage two actions before the shot. A defender can force the preferred option away without recording a steal. Models need a time horizon and a definition of contribution; changing either changes the answer.
Coaches should use impact metrics as prompts for film review, not as automatic lineup instructions.
A transparent workflow
Start with a basketball question, retrieve the relevant possessions, compare model output with video, and document disagreements. The most useful metric is one a coach can connect to repeatable behavior.
PRIMARY & AUTHORITATIVE SOURCES
Source notes
Sources are selected for direct relevance. Vendor sources establish what a product or partnership claims; they do not independently prove performance. We state that distinction in the analysis.