A Multi-Model Ensemble Framework for March Machine Learning Mania: Leveraging Gradient Boosting and Statistical Efficiency Metrics

Abstract

Trinh Quang Minh, Ngo Thi Lan, Nguyen Minh Hieu, Tran Minh Tan, Do Chi Tam, Lam Tan Phuong, Nguyen Chi Cuong and Bui Xuan Tung

Forecasting outcomes in collegiate basketball tournaments is inherently difficult due to the volatility and non-linear dynamics of team performance. This paper introduces a predictive framework designed for the March Machine Learning Mania 2026 competition, addressing both Men's and Women's NCAA tournaments. By applying advanced statistical feature engineering, the study moves beyond simple win-loss records to derive refined metrics such as possession- based offensive and defensive efficiency, effective field goal percentage (eFG), and pace. At the core of the framework is a weighted ensemble model that integrates three complementary machine learning algorithms: Logistic Regression, XGBoost, and Random Forest. Optimized weights—20% for Logistic Regression, 50% for XGBoost, and 30% for Random Forest—enable the ensemble to combine the interpretability of regression with the strong non-linear learning capacity of boosting and bagging methods. Validation experiments on 2025 season data confirm the effectiveness of this approach, with all individual models achieving Area Under the Curve (AUC) scores above 0.8420. The ensemble model reached a peak AUC of 0.8424, demonstrating superior predictive accuracy and stability. These findings highlight the value of combining domain-specific feature engineering with ensemble learning strategies to advance forecasting performance in sports analytics. The algorithm and processing procedure of the code predicts the outcome of March Madness 2026. This code implements a complete Machine Learning process from Feature Engineering to Ensemble Learning.

PDF