Skip to main navigation menu Skip to main content Skip to site footer

NFLNet: A hard hitting evaluation of deep learning approaches to tackle prediction

Abstract

This paper presents four architectures for predicting tackle probability in an NFL football game. Accu- racy, precision, recall, loss, and F1 scores are com- pared to identify the best classification model for the 2024 NFL Big Data Bowl. The models leverage NFL tacking data, including player position, speed, direction, and location relative to key field mark- ers. Tracking information was processed to extract meaningful plays, determine which features should be used in the solution, and identify plays with suc- cessful and unsuccessful tackle outcomes. A feed- forward network is presented as a baseline, and the performance of a convolutional neural network, transformer, and graph transformer are compared. The feed-forward network yielded an accuracy of 75%, which establishes the minimum accuracy of a simple architecture that uses minimal features. The convolutional network outperformed the base- line with an accuracy of 85%, but performed worse than the transformer and graph transformer, which achieved accuracy results of 90% and 92%, respec- tively. Ultimately, the graph transformer is found to be most effective at predicting the tackle probability for a league-average player.
PDF