From Stats To Brackets

Abstract

Predicting outcomes in the NCAA Men’s Basketball Tournament is notoriously difficult due to frequent upsets, inconsistent team performance, and the tournament’s single-elimination format. This project develops a data-driven model focused specifically on first-round games by combining historical tournament results (2008–2025) with KenPom efficiency metrics (2002–2025). The modeling process began with Dean Oliver’s “Four Factors” — effective field goal percentage, offensive rebounding rate, turnover rate, and free throw rate — and expanded to include tempo-adjusted statistics, tournament seeding, and engineered interaction features such as efficiency per possession and seed–efficiency interactions. A Random Forest classifier was selected to capture nonlinear relationships among team characteristics. Through systematic feature selection, seven highly predictive variables were identified, including adjusted efficiency margin, efficiency per possession, seed effects, shooting advantages, and possession-based metrics. After sequential hyperparameter tuning, the final model achieved a test accuracy of 79.82%, outperforming key baselines such as random guessing (50%), always selecting the higher seed (71.64%), and prior bracket strategies (75%). Statistical testing confirmed these improvements were significant. To ensure practical application, the model was exported to JavaScript and deployed within a Kotlin web application, allowing users to input team statistics and receive real-time predictions. Overall, this work demonstrates that integrating historical data, advanced metrics, and machine learning can produce robust and interpretable predictions for March Madness.