Automating assist identification in football (soccer): a machine learning approach using event and tracking data

Assists—the number of last passes before a goal—have been a predominant metric to evaluate offensive players contribution in association football. This poses two major limitations: first, assist definitions differ across data collection vendors, and second, the focus on passes leading to goals creates a strong dependency on the player taking the shot. We introduce the term potential assists comprising all actions, that prepare a teammate’s shot regardless of conversion. Additionally, we use expert-knowledge to derive an objective definition for potential assists and show that it can be detected with a substantial inter-rater reliability (Fleiss’ $$\kappa =0.63$$ , $$n=5$$ ). Using this definition, we utilize video footage to label 500 assists manually and train multiple supervised machine learning models to classify potential assists based on 29 features derived from positional and event data. We apply our automated detection approach on 105 DFB-Pokal matches and demonstrate a scalable method to enhance automatic event detection and improve the comparability of assist statistics for application in player assessment and scouting.