This project explores how deep learning can quantify shot quality and decision-making in basketball. We developed a system that predicts whether an NBA three-point shot will be made or missed by analyzing the ten seconds leading up to each attempt. Using player tracking data from the SportVU dataset (2015–2016 season) and player statistics from Basketball-Reference, we engineered a dataset of over 23,000 plays, each containing player distances, movements, and contextual stats.
Our model suite includes Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) architectures—each optimized for time-series data. The RNN achieved the best performance with around 80% accuracy, learning temporal dependencies in player motion and spacing that correlate with shot success.
To visualize our results, we integrated YOLOv8 models for player and court keypoint detection and applied homography transformations to map SportVU coordinates to real broadcast footage. This allows the system to overlay player names and real-time shot probabilities directly on video frames, offering new insights for analysts, coaches, and fans.
Overall, this project demonstrates how combining machine learning, sports analytics, and computer vision can provide a deeper understanding of NBA gameplay, highlight the mechanics of effective shooting, and pave the way for future AI-powered tools in sports analysis.