Michael Horton, Machine Learning Researcher, Sportlogiq
The introduction of player tracking systems to professional sports has generated large datasets capturing player movement and critical events that occur during games. The availability of such data, along with a growing appreciation of the role of analytics in evaluation and decision-making in sports, has created the conditions where analysis of player and team performance—as realised in the individual and group movement of the players—is now possible.
Concurrently, advances in machine learning methods and systems, in particular deep learning, have demonstrated success in a broad range of tasks. However, there is a fundamental mismatch between player tracking data and the implied input format for standard deep-learning architectures. To date, this issue was dealt with by feature engineering, where raw tracking data is preprocessed into a format suitable for the deep-learning model. However, such feature engineering is time-consuming, requires significant domain knowledge, and inevitably results in information inherent in the tracking data being discarded.
In this paper, we present a flexible neural network framework that accepts raw trajectory data as input, without the need for any feature engineering or preprocessing, and can be used for a wide variety of sports analytic tasks. We show that the framework works well on several football prediction tasks when using the player tracking data from the 2019 NFL Big Data Bowl as input: predicting the success rate, location and air-yards of passing plays; and predicting the tackler, tackle location and yards gained on running and passing plays.