Predicting NCAA Tournament Outcomes: A Gradient Boosting Approach Integrating Possession Efficiency and Expert Consensus

The NCAA Division I Men's Basketball Tournament is inherently characterized by extreme variance, stochasticity, and the high volatility of a single-elimination format. Historically, predictive analytics in this domain have bifurcated into two disparate philosophies: models relying exclusively on historical box-score statistics, which often fail to capture qualitative externalities, and subjective human rankings, which are inherently biased and mathematically uncalibrated. This comprehensive research report presents a highly specialized hybrid machine learning pipeline designed to resolve this analytical dichotomy. Utilizing a Light Gradient Boosting Machine (LightGBM) configured for symmetrical matchup evaluation, the predictive architecture forecasts highly calibrated probabilistic outcomes rather than traditional binary classifications. The modeling framework engineers possession-based efficiency metrics—specifically pace-adjusted Offensive and Defensive Ratings—alongside a dynamic Strength of Schedule (SOS) derivative synthesized from granular play-by-play data. These objective mathematical features are subsequently hybridized with a "wisdom of the crowds" consensus metric, computationally aggregated from historical Massey Ordinals.