Model Performance Methodology

The homepage shows rounded historical backtest results for two separate models: 57.2% for spread selections evaluated from 2022–2025, and 56.4% for totals selections evaluated from 2023–2025. The underlying results are 57.18% and 56.41%, respectively, as summarized August 24, 2026.

These rates apply only to the highest-confidence 20% of eligible games, not to every game the models evaluated. The evaluation produces about 100 selections per model per season.

StatCFB uses weekly walk-forward evaluation. For each test week, the model is trained using only information available before that week and then evaluated on that week's completed games. Results from a test week are not available to that week's model.

Selections are evaluated against recorded opening market lines. Spread and totals results are calculated separately because the models use different targets and evaluation windows.

The reported hit rates summarize the highest-confidence 20% of eligible model predictions. This confidence gate is intended to represent the subset the model would identify as its strongest opportunities, rather than implying a wager on every matchup.

Games without the inputs or market line required by an evaluation are not eligible for that evaluation. Pushes are not counted as wins.

The current research database contains 8,121 historical game-result records as of August 24, 2026. The homepage rounds this to 8,000+.

The number of model-ready rows varies by model because a matchup may be excluded when a required point-in-time feature is unavailable. The historical-record count is therefore a description of the research dataset, not the number of qualifying backtest selections.

Backtested results are historical simulations, not live customer results. They can be affected by data availability, market-line timing, selection rules, model changes, and other assumptions.

The displayed percentages are rounded summaries and do not promise a particular result. Historical and backtested performance does not guarantee future performance.

StatCFB intends to keep forward model results separate from historical backtests. A forward record should include each published prediction, its pregame timestamp and line, the applicable model version, and its final grade before it is described publicly as a tracked record.