The capabilities and limitations of machine learning in athletic evaluation and predictive modeling

Abstract

For over a century, evaluating athletic talent relied mostly on human intuition and the "eye test," which is a subjective approach limited by human bias, small sample sizes, and financial mistakes. Because modern sports have started collecting massive amounts of player tracking data, teams are now shifting toward machine learning (ML) to process this complex physical information. Driven by this technological shift, the purpose of this thesis is to investigate exactly what machine learning can and cannot predict compared to traditional human evaluation. To accomplish this investigation, the research compares human baselines against computer models across three key areas: finding undervalued global talent, designing highly structured set-piece plays, and medically clearing injured players to return to the field. Based on these comparisons, the findings show that machine learning greatly reduces human error in highly structured situations like corner kicks and in predicting physical injuries before they happen. Conversely, the analysis also reveals strict limits. Computer models struggle during chaotic open-play scenarios, often fail when there is not enough historical data, and fundamentally cannot measure human psychology or locker-room chemistry. In addition to these limitations, highly accurate models are often uninterpretable "black boxes" that cannot explain their reasoning, forcing teams to use Explainable AI (XAI) tools to translate the math into something coaches can trust. Ultimately, considering both its predictive power and its flaws, this research demonstrates that machine learning works best as a tool to assist and coexist with humans rather than replace them. Therefore, by letting computers handle the data processing, sports teams can free up human experts to combine those mathematical probabilities with the psychological insights and instincts needed for elite athletic success.