Predicting ATP Player Tennis Performance Using Machine Learning

Abstract

Predicting performance in professional tennis has applications for analysts, coaches, and betting strategies. This study investigates whether machine learning models can accurately predict Association of Tennis Professionals (ATP) match outcomes based on historical player data. Statistics such as serve percentage, break point conversion, and win-loss records were compiled from Ultimate Tennis Statistics and ATP Tour databases. Three models were tested: linear regression, logistic regression, and random forest classifiers. Model performance was compared using accuracy, R², and F1 score. Random forest achieved the highest accuracy (93.36%), followed by logistic regression (91.15%), while linear regression produced a moderate correlation (R² = 0.544). Serve consistency and break point conversion emerged as key indicators of success. Results suggest that ensemble-based models are most effective for capturing the non-linear relationships in tennis performance. While the study is limited by reliance on career-level statistics, these findings highlight the potential for machine learning to enhance tennis analytics, strategic planning, and predictive applications.