How the NBA and AWS built an AI system to measure what actually wins basketball games

This article describes how the NBA and AWS developed Leverage Score, an AI system that measures which players and possessions drive basketball game outcomes using counterfactual modeling.

• Leverage Score analyzes alternative scenarios for each possession to allocate credit across all ten players on court
• System operates on −10 to +10 scale, identifying high-leverage moments beyond traditional box score statistics
• Architecture uses Amazon EKS, Apache Flink processing 2,500 events per second, S3, DocumentDB, and Direct Connect
• LightGBM model trained on three seasons of NBA data predicts win probability based on seven engineered features
• System calculates actual outcome, counterfactual scenarios, and net leverage within seconds of possession ending
• Credit attribution integrates Shot Difficulty (xFG%) to distribute credit between shot creators and finishers
• Blue-green deployment with shadow mode testing validates new models against live game data before production
• Architecture represents reusable template for real-time impact analysis beyond sports analytics

Leverage Score combines basketball expertise with AI infrastructure to quantify player impact and answer which actions truly change game outcomes.