College Football Accountability
Every game's prediction is locked 90 minutes before kickoff — spread, total, win probability and vig-free moneylines, plus the market's opening and closing references — and never rewritten. After the final it is graded against the score. n is shown beside every figure and each model version is pooled separately, so eras never blend.
564 snapshots captured · 561 graded · 3 awaiting a final · model c3-v1.2, c3-v1.5, c3-v1.6, c3-v1.7.
n=561
bias -2.13
vs 0.2500 baseline
Spread bias -1.32 (+ = the projected home margin ran too high) · winner accuracy 77.5%.
By matchup type
Half an early-season college board is an FBS-vs-FCS game and its error
distribution is a different animal, so the two are never pooled into a single
headline. The FCS row is also where the fitted tier calibration
(fcs.spread_offset) has to keep proving itself in live grading.
| Subset | n | Spread MAE | Spread bias | Total MAE | Total bias | Brier |
|---|---|---|---|---|---|---|
| FBS vs FBS | 439 | 12.99 | -0.74 | 12.62 | -1.98 | 0.1907 |
| FBS vs FCS | 122 | 14.83 | -3.42 | 12.03 | -2.66 | 0.0445 |
| Neutral site | 7 | 8.20 | +5.94 | 18.51 | +1.51 | 0.2627 |
By week
| Week | n | Spread MAE | Spread bias | Total MAE | Total bias | Brier | Model |
|---|---|---|---|---|---|---|---|
| 0 | 8 | 16.71 | -2.35 | 12.01 | +6.11 | 0.2429 | c3-v1.2 |
| 1 | 91 | 15.66 | -5.85 | 12.48 | -0.93 | 0.0866 | c3-v1.5 |
| 2 | 86 | 14.17 | -3.41 | 10.94 | -2.62 | 0.1196 | c3-v1.6 |
| 3 | 75 | 11.23 | +0.30 | 13.01 | -4.04 | 0.1087 | c3-v1.6 |
| 4 | 71 | 13.77 | +2.09 | 11.15 | -2.40 | 0.1553 | c3-v1.6 |
| 5 | 62 | 13.90 | -1.64 | 13.93 | -3.01 | 0.1752 | c3-v1.6 |
| 6 | 59 | 12.33 | -0.00 | 12.00 | -0.66 | 0.2367 | c3-v1.6 |
| 7 | 62 | 12.37 | -0.51 | 13.24 | -0.37 | 0.2362 | c3-v1.6, c3-v1.7 |
| 8 | 45 | 11.73 | +0.76 | 14.57 | -4.61 | 0.2140 | c3-v1.7 |
| 9 | 2 | 15.23 | +15.23 | 10.93 | -2.45 | 0.3816 | c3-v1.7 |
By model version
More than one era is in this sample, so the pooled total above spans versions — these rows keep them apart.
| Version | n | Spread MAE | Total MAE | Brier |
|---|---|---|---|---|
| c3-v1.2 | 8 | 16.71 | 12.01 | 0.2429 |
| c3-v1.5 | 91 | 15.66 | 12.48 | 0.0866 |
| c3-v1.6 | 356 | 13.16 | 12.11 | 0.1547 |
| c3-v1.7 | 106 | 11.97 | 13.84 | 0.2289 |
Model vs closing line — paired sample, n=557
| Spread MAE | Total MAE | Brier | |
|---|---|---|---|
| Model | 13.40 | 12.44 | 0.1597 |
| Closing line | 11.53 | 12.25 | 0.1542 |
Head-to-head — spread: model 233, market 324, ties 0 · total: model 264, market 290, ties 2. Spread edge vs close: -1.87 points of MAE.
Closing-line value
spread: n=507 · moved toward the model 36% of the time · mean |open−close| 1.41 · mean |model−close| 6.20.
total: n=410 · moved toward the model 45% of the time · mean |open−close| 1.61 · mean |model−close| 4.20.
winprob: n=386 · moved toward the model 52% of the time · mean |open−close| 0.03 · mean |model−close| 0.12.
Cold-start games whose open and close resolve to the same odds row are excluded (0), never counted as zero movement.
Locked, awaiting kickoff · 3
The immutable record, before the fact. These rows were written 90 minutes before kickoff and cannot change.
| Wk | Game | Kickoff | Model line | Total | Home WP | ML | Market open | Line at lock | Locked |
|---|---|---|---|---|---|---|---|---|---|
| 9 | Tulane @ Charlotte | Fri 7:00 PM ET | -10.6 | 41.3 | 27% | -273 / +273 | -10.0 | -8.5 | 2026-10-30 20:45 UTC |
| 9 | Baylor @ UCF | Fri 7:30 PM ET | +1.4 | 48.6 | 54% | +116 / -116 | +1.5 | -2.5 | 2026-10-30 21:15 UTC |
| 9 | Kent State @ Sacramento State | Fri 10:30 PM ET | -4.1 | 41.0 | 40% | -150 / +150 | -9.5 | -7.5 | 2026-10-31 00:15 UTC |
Market references are recorded for comparison and CLV only — they are never blended into the model (SPEC-CFB §5). Moneylines are vig-free fair prices derived from the model's own win probability.