Comparing the Role of Physics Features in Machine Learning Models Across Sports

Machine learning has become a central tool in sports analytics, yet its performance often depends on how well models reflect the underlying physics of play. This study explores when and where physics-informed features improve predictive performance across three professional sports with distinct dynamical regimes: Major League Baseball (MLB), the National Football League (NFL), and the National Hockey League (NHL). Using freely available datasets—Statcast batted-ball data, NFL Big Data Bowl tracking, and NHL play-by-play shot records—we construct parallel prediction tasks: home-run classification, tackle completion, and goal probability, respectively. Each task is modeled using both standard data-driven methods (gradient boosting, logistic regression) and physics-augmented variants that include features derived from kinematics, geometry, and energy transfer principles. We compare model accuracy, calibration, and interpretability across sports to assess how the structure of each game influences the usefulness of physical priors. Preliminary analyses suggest that the degree to which physics-informed features add value may correlate with how strongly physical laws constrain play dynamics. These findings aim to clarify the interplay between data-driven and physics-based modeling in sports contexts with differing physical complexity.