Our Methodology

Our prediction system uses machine learning models trained on comprehensive college football data spanning multiple seasons. We employ a consensus approach that combines multiple models to produce more accurate and reliable predictions.

Team Performance Metrics

• Team Efficiency Ratings - Advanced offensive and defensive metrics
• Power Rankings - Statistical team strength indices
• Historical Performance - Season and multi-year trend analysis
• Talent Metrics - Roster composition and quality indicators

Game-Level Features

• Home field advantage adjustments
• Rest days and scheduling factors
• Historical matchup data
• Conference strength metrics

Consensus Approach

A single ridge regression predicting game margin, built from 184 features. Every feature is computed from games played before kickoff, so no prediction can see its own result:

• Season-to-date form - rolling per-game box score and efficiency averages, blended with the prior season early in the year
• Opponent-adjusted ratings - ridge-solved offense/defense ratings recomputed each week from prior games only
• Advanced efficiency - success rate, explosiveness, line yards, with garbage-time snaps excluded
• Preseason priors - returning production and recruiting, which carry the most weight in early weeks

Training Process

• Ridge regression (α = 10)
• Trained on 9,306 FBS games with closing lines, 2013–2025
• Walk-forward validated: each season predicted using only prior seasons
• Win probability fitted to measured outcomes, not assumed

Score Calibration

We apply calibration factors to ensure our predictions match real-world scoring distributions:

• Training data average: 55.2 total points per game
• Predictions are scaled to match this distribution
• Prevents systematic OVER/UNDER bias

Team Name Mapping

Critical quality control step to ensure accurate data alignment:

• Canonical team names mapped across all data sources
• Verified mappings for all teams before generating predictions
• Prevents feature misalignment that could cause incorrect predictions

Model Accuracy

• R² Score: 0.607 (explains 60.7% of variance in game scores)
• Mean Absolute Error: ~10 points per score prediction
• ATS Target: 55%+ win rate (breakeven is 52.4%)

Edge Detection

We calculate the "edge" as the difference between our prediction and Vegas lines:

• Small Edge: < 3 points - Lower confidence
• Medium Edge: 3-7 points - Moderate disagreement with Vegas
• Large Edge: 7+ points - Significant value opportunity

What We Can't Predict

• Injuries to key players (unless reflected in pre-game lines)
• Weather conditions (extreme weather can significantly impact scoring)
• Motivational factors (rivalry games, playoff implications)
• Coaching changes or in-season staff turnover