Quantifying NBA Shot Quality: A Deep Network Approach

Since the introduction of player positional tracking data to the NBA in 2013, the field of basketball analytics has been steadily developing. As such, more and more teams utilize data-driven approaches to maximize the potential for their team to score a basket. In this paper, we explore leveraging recurrent deep-learning architectures for the quantification of the quality of a given basketball possession. To do so, we curate a dataset consisting of player positional and statistical data for basketball shots from the 2015-2016 NBA season, dividing the data into subsets of Mid-Range, 3 Pointer, Paint (Non-Restricted), Restricted Area, Backcourt, and All shots. We then explore the efficacy of three recurrent deep learning architectures: the vanilla Recurrent Neural Network (RNN), Long Short Term Memory (LSTM), and Gated Recurrent Units (GRU) in classifying whether a shot will be made or missed, leveraging the probability of the classification to quantify the quality of the shot. Each of these models incorporates relative player distances with respect to the shooter and the basket, as well as statistical information for both the offensive and defensive players on the court. Our models achieve state-of-the-art accuracy on this task with scores of 81.7%, 81.9%, 82.1%, 81.0%, 96.3%, and 81.8% for the aforementioned data subsets respectively. Furthermore, we validate that our model's probability score is an accurate measure of shot quality by comparing our predictions with experts from the field.