Predictive modeling of UAAP Men's Basketball game outcomes using machine learning

Abstract

Sports analytics is a growing field of computational science where researchers collect and analyze data to derive insights and gain competitive advantages. Despite the unpredictable nature of sports, a popular problem in sports analytics is the prediction of game outcomes across multiple sports. This study tackles the problem of game outcome prediction on UAAP Men's Basketball games using a machine learning (ML) approach. We developed binary classification models namely, Logistic Regression, Random Forest, Naive Bayes, Support Vector Machines (SVM), and Adaptive Boosting (AdaBoost) to determine the winners of UAAP games. For comparison to the performance of the ML classification models, we also defined naive game prediction techniques based on win-loss percentage and point differential as benchmark models. The features used in the ML models are derived from the box score data from UAAP Men's Basketball games for the past four seasons (Seasons 82, 84–86). Game outcomes are predicted via the pre-season forecast approach, where all games for the next season are predicted using only data from past seasons. We also investigate how prediction accuracy is affected by varying the size and recency of training data. Our results show that ML models, specifically LR and SVM, generally outperformed benchmark models. Further work may explore how ML models may be enhanced or combined with more naive yet effective models.