Advancing NCAA March Madness Forecasts Through Deep Learning and Combinatorial Fusion Analysis

Abstract

This paper presents a novel methodology for accurately predicting game outcomes in the NCAA Men's Basketball Tournament, known as March Madness. The high variance inherent to this tournament poses challenges that often exceed traditional forecasting methods. We implement a rigorous data preprocessing and feature engineering pipeline tailored to this problem using a historical NCAA Men's Tournament dataset. Four diverse neural network architectures, including convolutional, recurrent, feedforward, and residual networks, are developed to model team statistics and sequential game information. We formulate a robust forecasting approach by leveraging combinatorial fusion analysis (CFA) to merge predictions from these heterogeneous models. Our methodology synergistically integrates deep learning, CFA techniques, and domain-specific feature engineering to push the boundaries of state-of-the-art March Madness forecasting. This enables basketball enthusiasts to gain strategic advantages in tournament contests while advancing sports outcome prediction research.