DAxT: A Deep Learning Approach to Valuing Defensive Actions in Football

A Novel Metric Inspired by the Original xT Model that Leverages Deep Learning for Better Valuation of Defensive Actions in Football

Introduction

The researchers introduce us to the problem of valuing actions in real-world domains and highlight the specific focus on valuing the actions of defenders in Football. The authors explain that while there has been significant progress in data science research and football analysis, valuing defensive actions remains a challenge due to the nature of these actions that often prevent events from occurring.

The paper emphasizes the importance of measuring player performance in team sports and discusses the unique challenges in football due to its low-scoring and dynamic nature. It mentions the widely used metric “Expected Goals (xG)” that quantifies the probability of a shot resulting in a goal, but notes the lack of similar metrics for defensive actions.

To address this gap, the authors propose a novel data-driven model called DAxT (Defensive Action Expected Threat) that focuses on valuing defensive and out-of-possession actions in football. Unlike traditional metrics that study actions leading to events, DAxT values defensive actions based on what they prevent from happening. The model combines research in football with deep learning techniques to accurately assess the impact of…