Model evaluation · 2026
Prediction performance
Recorded publications and reconstructed history, with sample sizes and simple baselines shown separately.
What is measured
Every metric and baseline in a row uses the same completed, non-tied games. Both teams need at least two results in the projection’s prior information set. Accuracy is the fraction of winners selected; at exactly 50%, the designated home team is selected. Tied finals are excluded from accuracy, margin error, and Brier score alike.
Mean absolute margin error is the average absolute difference between the projected and actual home scoring margin. Lower is better. Brier score averages the squared difference between the home win probability and the home win outcome (one for a win, zero for a loss). Lower is better; a constant 50% forecast scores 0.250.
The home baseline always selects the designated home team, including at neutral sites. The prior record baseline selects the team with the better prior win percentage, counting a tie as half a win. Equal percentages select the designated home team. These are winner-selection baselines, so they have no invented margin or probability forecasts.
Recorded publications
These odds were saved after a successful production deployment, with the deployed rankings artifact verified by SHA-256. The selected snapshot is the latest successful capture before the earliest confirmed kickoff seen by the recorder. A later reschedule does not reopen that cutoff.
Publication verification times among graded games: to .
| Sample | Games | Model accuracy | Home baseline | Prior record baseline | Margin error (pts) | Brier score |
|---|---|---|---|---|---|---|
| All eligible games | 305 | 76.4% | 50.8% | 70.5% | 13.3 | 0.164 |
326 candidate projections; excluded: 21 unfinished, 0 tied finals, 0 invalid, 0 without usable prior records. Equal prior win percentages used the designated home team in 30 graded games.
By period
| Sample | Games | Model accuracy | Home baseline | Prior record baseline | Margin error (pts) | Brier score |
|---|---|---|---|---|---|---|
| Week 10 | 10 | 70.0% | 50.0% | 70.0% | 10.9 | 0.156 |
| Week 11 | 295 | 76.6% | 50.8% | 70.5% | 13.4 | 0.164 |
Within each division
| Sample | Games | Model accuracy | Home baseline | Prior record baseline | Margin error (pts) | Brier score |
|---|---|---|---|---|---|---|
| FBS | 72 | 69.4% | 51.4% | 66.7% | 14.0 | 0.210 |
| FCS | 62 | 64.5% | 51.6% | 58.1% | 10.4 | 0.210 |
| Division II | 66 | 75.8% | 54.5% | 69.7% | 13.6 | 0.162 |
| Division III | 105 | 88.6% | 47.6% | 81.0% | 14.3 | 0.106 |
Both teams must belong to the listed division. 0 cross-division and 0 unclassified games remain in the overall totals.
Confidence calibration
| Favorite probability | Games | Mean forecast | Actual win rate |
|---|---|---|---|
| 50% to under 60% | 70 | 55.3% | 64.3% |
| 60% to under 70% | 58 | 64.9% | 63.8% |
| 70% to under 80% | 65 | 75.0% | 72.3% |
| 80% to under 90% | 62 | 84.6% | 87.1% |
| 90% to 100% | 50 | 95.0% | 100.0% |
A calibrated 70% favorite would win about 70% of comparable games over a large sample. Small bins can fluctuate substantially. Empty bins display a dash.
Model versions in this sample
period-ridge-walk-forward-v2: 305 games
Reconstructed walk-forward backtest
These projections are reconstructed from the current source data. For each game, ratings, prior records, and calibration use only earlier periods. They describe a walk-forward backtest; they are not evidence of odds published before kickoff.
| Sample | Games | Model accuracy | Home baseline | Prior record baseline | Margin error (pts) | Brier score |
|---|---|---|---|---|---|---|
| All eligible games | 2,308 | 72.8% | 55.4% | 68.2% | 14.7 | 0.178 |
2,309 candidate projections; excluded: 1 unfinished, 0 tied finals, 0 invalid, 0 without usable prior records. Equal prior win percentages used the designated home team in 344 graded games.
By period
| Sample | Games | Model accuracy | Home baseline | Prior record baseline | Margin error (pts) | Brier score |
|---|---|---|---|---|---|---|
| Week 2 | 46 | 60.9% | 60.9% | 58.7% | 14.1 | 0.211 |
| Week 3 | 245 | 68.2% | 62.9% | 67.8% | 17.4 | 0.195 |
| Week 4 | 256 | 71.5% | 53.9% | 67.6% | 16.7 | 0.184 |
| Week 5 | 293 | 74.4% | 57.7% | 64.8% | 14.9 | 0.187 |
| Week 6 | 266 | 72.2% | 56.4% | 68.0% | 13.7 | 0.178 |
| Week 7 | 295 | 72.9% | 51.5% | 65.4% | 14.5 | 0.183 |
| Week 8 | 296 | 70.9% | 56.8% | 67.9% | 14.0 | 0.178 |
| Week 9 | 306 | 76.1% | 53.6% | 73.9% | 14.3 | 0.157 |
| Week 10 | 305 | 76.7% | 51.1% | 70.8% | 13.0 | 0.161 |
Within each division
| Sample | Games | Model accuracy | Home baseline | Prior record baseline | Margin error (pts) | Brier score |
|---|---|---|---|---|---|---|
| FBS | 456 | 67.3% | 58.1% | 64.3% | 13.3 | 0.213 |
| FCS | 428 | 66.8% | 56.8% | 63.6% | 13.5 | 0.209 |
| Division II | 579 | 72.4% | 53.4% | 69.6% | 14.4 | 0.174 |
| Division III | 790 | 79.0% | 51.9% | 72.3% | 15.6 | 0.145 |
Both teams must belong to the listed division. 55 cross-division and 0 unclassified games remain in the overall totals.
Confidence calibration
| Favorite probability | Games | Mean forecast | Actual win rate |
|---|---|---|---|
| 50% to under 60% | 531 | 55.0% | 55.9% |
| 60% to under 70% | 560 | 64.8% | 64.5% |
| 70% to under 80% | 509 | 74.7% | 74.7% |
| 80% to under 90% | 409 | 84.5% | 86.1% |
| 90% to 100% | 299 | 94.4% | 97.0% |
A calibrated 70% favorite would win about 70% of comparable games over a large sample. Small bins can fluctuate substantially. Empty bins display a dash.
Model versions in this sample
period-ridge-walk-forward-v2: 2,308 games
Coverage and limitations
Early games, unknown kickoff times, missing captures, failed deployments, and unavailable source data limit coverage. A local build never creates publication evidence. A missing ledger starts an empty recorded history; a corrupt ledger stops capture until recovered. The cache retains the current and previous seasons, and this page evaluates only 2026.
Reconstructed history uses today’s corrected source results and team classifications. It cannot reproduce the exact information available at a historical publication time. Source corrections to earlier results or a model revision can change a reconstruction. Adding later periods does not refit earlier projections. Recorded odds retain their captured model version, ratings, calibration, and prior records after kickoff.
Samples are descriptive, with no confidence intervals shown. A high accuracy can reflect easy favorites or a small sample. The model does not include injuries, weather, or lineup changes. These results do not establish a betting advantage or guarantee future accuracy.
Outside the two evaluation cohorts: 199 live or unverified projections, 0 from another or unspecified season, and 0 duplicate entries. Read the projection method or inspect the computed data.