# Applying Deep Learning to Basketball Trajectories

**Authors:** Rajiv Shah, Rob Romijnders  
**arXiv:** [1608.03793](https://arxiv.org/abs/1608.03793)  
**Date:** 2016-08-12  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1608.03793v2/

## Key points

- Shift towards utilizing player and ball tracking data in sports analytics
- Rise of deep learning predictive methodologies with reduced emphasis on feature engineering
- Use of recurrent neural networks for forecasting success of three-point shots in basketball games
- Ability to train models solely based on sequential positional data without prior knowledge of physics
- Comparison between deep learning models and static machine learning model using NBA SportVu data
- Deep learning models outperforming static feature-rich machine learning model in predicting three-point shot outcomes
- Implications of deep learning approaches in enhancing traditional feature-based machine learning techniques in sports analytics

## Summary

In the realm of sports analytics, there is a noticeable shift towards utilizing player and ball tracking data for enhanced insights. This trend is accompanied by the rise of deep learning predictive methodologies that leverage extensive datasets with reduced emphasis on feature engineering. A recent study conducted by Rajiv Shah and Rob Romijnders delves into this intersection by employing recurrent neural networks in the form of sequence modeling to forecast the success of three-point shots in basketball games. What sets this research apart is its ability to train models to comprehend the trajectory of a basketball solely based on sequential positional data, without requiring any prior knowledge of physics. To provide a benchmark for comparison, a static machine learning model incorporating a comprehensive set of features like angle and velocity alongside positional information was also evaluated. The dataset used for analysis comprised over 20,000 three-point attempts sourced from NBA SportVu data. Remarkably, the findings revealed that the deep learning models relying on sequential positional data outperformed the static feature-rich machine learning model in predicting the outcome of three-point shots. This outcome suggests that deep learning approaches may present a promising avenue for enhancing traditional feature-based machine learning techniques when applied to tracking data in sports analytics. The study's implications extend beyond basketball trajectories, hinting at broader applications within sports analytics where deep learning models could potentially offer significant improvements over conventional methodologies. By showcasing the efficacy of recurrent neural networks in capturing complex patterns within sequential data, Shah and Romijnders' work underscores the transformative potential of deep learning in revolutionizing how we analyze and interpret sports performance metrics.