NCAAB model performance

Stratified by power vs mid-major — testing the small-conference edge hypothesis

Hypothesis under test: mid-major lines may be sloppier (less professional handicapping) and yield more model accuracy (and, if odds are weak, more ROI). Tracking accuracy by tier is one way to spot it; ROI requires forward odds capture via the EV scanner.

All games

—

0 settled

Power vs Power

—

0 settled

Mid-major vs Mid-major

—

0 settled

Model: XGBoost on v2 features, 2-season training window. Backtest: 67.7% ML accuracy / Brier 0.206. Measured against real closing lines (n=935) the market wins: 66.95% model vs 67.70% market, Brier gap CI [0.016, 0.037] excluding zero, ROI −6.2% to −8.9% at every threshold. Verdict: informational only.