Title: HoopEval: Individual Player Action Evaluation via Deep Reinforcement Learning
Authors: Xing Wang, Yu Fu, Sheng Xu, Konstantinos Pelechrinis, et. al
Institution: MIT Sloan Sports Analytics Conference
Publication Date: March 7, 2026
The Problem:
NBA teams struggle to assign value to individual decisions within a possession, especially off-ball actions. Traditional metrics (points, plus-minus, RAPM) and even advanced frameworks like Expected Possession Value (EPV) evaluate outcomes at the team level, not the player level. EPV assigns the same value to all offensive players at a given moment, making it impossible to isolate who actually created or destroyed advantage. This limitation is most costly when evaluating screeners, spacers, cutters, and connectors whose impact shows up later, not immediately in the box score or shot quality.
Methodology:
HoopEval reframes a basketball possession as a multi-agent decision sequence, similar to how chess engines evaluate move-by-move advantage. Instead of predicting forward like EPV, the model uses offline reinforcement learning to propagate value backward, assigning credit to each player’s action at each moment.
Using NBA SportVU tracking data, the court is broken into a hexagonal grid and represented as a graph: players and the ball are nodes, and their relationships (teammate, defender, ball-handler) are edges. A graph attention network acts like advanced film vision learning who matters to whom spatially while a transformer captures timing and sequence, similar to understanding how a play unfolds step-by-step.
The core innovation is an autoregressive Q-function, which evaluates actions in order (ball action, then each player’s movement), allowing the model to assign marginal value to passes, dribbles, shots, and off-ball movement. Rewards are shaped using Quantified Shot Quality (qSQ), providing dense feedback instead of waiting for makes or misses.
Why it Matters:
HoopEval provides a mechanism to identify "surplus value" that traditional box scores miss. It captures how coordinated actions, such as a pick-and-roll or off ball spacing, create future offensive advantages even when immediate shot quality remains low. This allows for a more rigorous ROI analysis of player archetypes; for instance, the model can penalize a "good" open shot if the shooter is statistically incapable of making, while rewarding the tactical movement that led to it. It transforms subjective "eye-test" scouting into a quantified, decision aware evaluation system that isolates a player's true marginal contribution to winning.
It quantifies how a weak-side relocation, a well-timed screen, or a delayed dribble improves the future quality of a possession even when no assist or shot attempt follows.
Coaches can diagnose whether offensive stagnation is driven by poor reads, poor spacing, or correct decisions that teammates fail to capitalize on.
ACTIONABLE TAKEAWAYS
Optimize Off-Ball Valuation: Use HoopEval to identify and reward players who consistently generate "unseen" value through high IQ spacing and screening, even if they lack high traditional usage rates.
Refine Decision Making Drills: Implement action level Q-value feedback in practice film sessions to show players which specific movements (e.g., dribbling to maintain possession vs. forcing a pass) objectively maximize the team's long-term expected value in specific defensive configurations.
Player Development: Tailor film sessions around negative value habits (late screens, stagnant spacing) rather than outcomes alone.
Player Evaluation: Integrate HoopEval-style action values into internal reports to flag off-ball contributors who outperform their contract or reputation.
Research Paper Link: HoopEval: Individual Player Action Evaluation via Deep Reinforcement Learning