Title: From Metrics to Decisions in NBA Analytics: A Critical Integrative Review and Decision-Readiness Framework
Author: Yang Zhou, Tianyu Guan
Institution: TY Sports Analytics Inc. & York University (arXiv preprint, stat.AP)
Publication Date: September 23, 2026
The Problem: Better Predictions, Same Hard Decisions
NBA teams have never had more metrics, yet better predictions do not automatically produce better decisions. A plus-minus rating describes contribution alongside current teammates, not performance in a new role. An injury risk forecast does not reveal what resting a player would actually change. A salary projection does not say what a contract is worth to one specific roster. The authors argue the gap is not data granularity or algorithm complexity. It is a mismatch between the question a model answers and the choice a coach or GM faces on a specific date, under specific cap rules, with limited alternatives.
Methodology: A Decision-First Lens on 58 NBA Studies
This is a critical integrative review, not a new model. The authors build a decision first lens and apply it to 58 NBA reports:
Decision chain (film to final score): Evidence is traced from observation to state estimate, candidate action, decision rule and evaluation. A strong link at one stage does not validate the next.
Six readiness gates (the pregame checklist): Point in time validity, uncertainty, context portability, action feasibility, opportunity set observability and evaluation. Each is a separate way a sound estimate can fail a real choice.
Coupled player value (fit over rating): Ability, role, synergy and availability interact with cost and team context, so the same wing carries different value for a deep contender than for an apron team.
Evidence gap map (the research depth chart): Seven research streams are graded at five claim levels, from measurement through prospective evaluation.
Why it Matters: The Research Stops Short of the Decision
The map shows where public research can back a decision:
Strong on measurement, weak on action: Of 35 stream by claim cells, only 5 rate "developed" (four are measurement), 18 are emerging and 12 are thin.
No proven decision rules: Zero retained studies prospectively tested an implemented NBA decision rule against a credible alternative. All seven streams rate thin there.
Accuracy gains can mislead: A cited regularized plus minus model cut game margin error from 12.01 to 11.47 points, but only given lineups actually played, so it is not a pregame forecast.
Rest is unsettled: League surveillance data showed no significant injury risk difference across load management groups, though rest was not randomized.
Treat this as a diagnostic, not proof. It is a preprint, the search was not systematic, no inter rater agreement is reported, and the framework has not been shown to improve team outcomes. It covers public research only, not private team practice.
ACTIONABLE TAKEAWAYS
Require a decision record: Before acting on a model output, document the decision owner, date, information cutoff, alternatives considered and the conditions that would change the recommendation.
Screen legality before ranking: Run CBA feasibility (apron status, salary matching, pick rights) first. The paper's trade deadline example shows the top rated wing can be off the board for an apron team.
Log forecasts in shadow mode: Freeze recommendations and forecasts before outcomes are known, then grade calibration afterward without redefining the target.