How the College Football model is tested
How a pick is made
Team strength, built from the games themselves. It calls who wins.
- Every game's result
- The margin
- The opponent
- Home field
- How the team started the year
The exact inputs and how they are weighted stay private.
When books price a game, the win chance is built on their no-vig price, so picks usually match the favorite. The model makes no spread or over/under call.
Tuned on old seasons, frozen, then scored on games it had never seen.
A change ships only if it holds up there. Most ideas don't, and they stay out.
Every pick is stored before kickoff.
It can update until then. The version at kickoff is the one graded, and it can't be edited after.
Graded after the final, misses included, nothing deleted.
Winners are the record: no spread or over/under records. Every pick is in the ledger →
122 college football picks locked for the games ahead. Every one grades here for free after the final. Seeing them before the start is Pro.
See the picks →Start 7-day free trialHow it has held up
College Football: when it said 70%, did it win 70%?Each lane is a confidence band. Calibrated means the dot lands in its gray band; the bar shows luck's range.
Yes so far: all 5 bands with enough picks landed where they said, within luck's range.
- Said50–60%60%81 of 136
- Said60–70%65%90 of 139
- Said70–80%75%102 of 136
- Said80–90%88%104 of 118
- Said90–100%97%98 of 101
Live picks, locked before the start. Backtest: walk-forward backtest of the model alone, 2024–26 seasons, not in the record. Lanes with fewer than 10 graded picks are faded.
Expected vs actual winnersThe dashed line is the wins its own probabilities expected, shaded for luck's 90% band. Solid is what happened.
Dotted line, pick 273: the method changed. Before it the calls were the model alone; after it a priced game’s number follows the market’s price, so the stretches are not the same product.
The solid line ends inside the shaded band: the record is where honest probabilities would put it.
Show as a table
| After pick | Date | Won | Expected | 90% band |
|---|---|---|---|---|
| 79 | Sep 12 | 65 | 58.9 | 52.9–64.9 |
| 158 | Sep 19 | 134 | 124.3 | 116.4–132.3 |
| 236 | Sep 26 | 198 | 182.6 | 172.7–192.4 |
| 315 | Oct 3 | 257 | 240.8 | 229.2–252.4 |
| 394 | Oct 14 | 309 | 296.2 | 282.9–309.5 |
| 473 | Oct 24 | 363 | 353.2 | 338.5–367.9 |
| 551 | Oct 31 | 422 | 407.6 | 391.6–423.7 |
| 630 | Nov 10 | 475 | 462.5 | 445.2–479.8 |
College Football backtest · not in the recordEvery game priced only from what was known before it, then scored once. Kept apart from the record.
Walk-forward backtest of the model alone, 2024–26 seasons. Priced before the market-anchored winner call, so it measures the model on its own.
A model read is information, not advice to bet. Past results do not guarantee future ones. Data and analytics only, not betting advice: any bet is your own decision, and we are not responsible for losses. 21+ only. Gambling problem? Call or text 1-800-GAMBLER.