Abstract
Recently the increase in volume of data related to sport- ing events has led to the development and increased use of machine learning (ML) as a tool to extract information. Taking sports into consideration, the utilization of ML has appealed to the field of sports betters who desire to exploit the betting markets. In this work, we implemented a prob- abilistic betting algorithm which utilizes model ensembles that combine the predictions from multiple tuned binary classification models to accurately predict the winner of a National Football League (NFL) game. We evaluated four common betting classifiers – Decision Tree, Logistic Re- gression, XGBoost, and Random Forest – and analyzed the accuracy of predicting NFL games for each model. Our results demonstrated that of the four models, the Decision Tree model performed the worse with a cross-validation ac- curacy of 69.5%. Using our three best models, we utilized a calibrated classifier model with a voting classifier as the base to build a probabilistic betting algorithm. Overall, on unseen data from the 2019 NFL season, our betting algo- rithm placed 207 bets out of the possible 251 games and won 170 of those bets. This resulted in a win percentage of 82.13%.