March Madness prediction: Different machine learning approaches with non‐box score statistics

Abstract

Abstract The popularity of analytical research specializing in forecasting of March Madness saw an increase in the past decades. While the influence of nongame statistics on the game outcome has become a great interest in sports analytics, little research has focused on situational factors in predicting sports tournament outcomes. Therefore, this study is to examine the use of different machine learning algorithms, including artificial neural network (ANN), k‐nearest neighbors (kNN), support vector machine (SVM), logistic regression, and random forest (RF), to forecast the winning in a matchup between any two given teams during the March Madness tournaments. Our data include 1370 observations with 685 tournament games from 2006 to 2007 to 2016 to 2017 seasons. The results show that neural networks outperformed all other classifiers (67% of accuracy), followed by SVM (65%), kNN (63%), logistic regression (63%), and RF (61%).