Modeling Football Player Trajectories During Passes Using Graph-Structured Recurrent Networks

Abstract

Accurately forecasting player trajectories during the ballin-flight phase of American football requires models that capture multi-agent interactions while remaining consistent with underlying physical constraints. This work evaluates how input normalization and target parameterization influence the performance of a Spatio-Temporal Graph LSTM model. Three normalization strategies are considered, along with two prediction targets, and the results show that RMS scaling combined with velocity prediction provides the highest accuracy, with an RMSE of 0.684. The findings indicate that normalization choices have a major effect on forecasting quality and that carefully selecting both the input scaling method and the target representation is essential for reliable trajectory prediction.