The AP Top 25 backtest: 68-55-1, +6.8 units.
A paper study of the only college football bets we publish: our model on games involving at least one AP Top 25 team. 124 bets graded against the closing line, flat 1 unit per game at -110, walk-forward with no tuning on 2026 games. Sample: every FBS game played through October 13, 2026, closing lines from DraftKings via ESPN.
01The result
| SCOPE | BETS | RECORD | WIN % | UNITS | ROI |
|---|---|---|---|---|---|
| AP Top 25 games (Model B) | 124 | 68-55-1 | 55.3% | +6.8 | +5.5% |
| AP Top 25 games (Model A weighting) | 124 | 67-56-1 | 54.5% | +4.9 | +4.0% |
Win percentage excludes pushes. Exact totals: +6.82 units for Model B and +4.91 units for the Model A weighting.
At a standard -110 price, a bettor has to win 52.4% of spread bets just to break even. The model cleared that bar by about 3 points in ranked games across 124 bets, and at -110 prices that gap is the whole difference between a +5.5% return and a slow bleed. One push in 124 games, flat staking, closing lines, no bet bigger than any other and no selective memory: this is the cleanest test we can run on our own process, and the ranked-game slice passed it.
02What we found
Finding one: the edge is selection, not volume. The same model, run over the rest of the FBS board in the same test, lost heavily:
| SLICE | BETS | RECORD | WIN % | UNITS | ROI |
|---|---|---|---|---|---|
| Games with at least one AP Top 25 team | 124 | 68-55-1 | 55.3% | +6.8 | +5.5% |
| All other FBS games | 325 | 133-187-5 | 41.6% | -66.1 | -20.3% |
Games without a ranked team went 133-187-5 for a -20.3% return, worse than flipping coins and paying the vig. That contrast is why this backtest is scoped to the Top 25 in the first place: the ranked games are where the model's information holds up against the closing market, and everywhere else it does not.
Finding two: the result is not one lucky set of weights. A second weighting of the same model (Model A, in the table above) went 67-56-1 for +4.9 units on the same ranked games. Two builds, the same conclusion.
Finding three: the model lives close to the market. Across the full 449-game test its projected spread landed within 3 points of the closing line in 96% of games. A model like that does not make money by outsmarting the number every Saturday. It makes money, when it makes it, by being right slightly more often where its inputs are richest: ranked teams with deep, stable data.
03Why this site covers ranked games
Our college football coverage follows this result directly. We publish previews for featured games with ranked matchups at the center of the slate, because that is the slice where the model earned its keep on paper. Covering all 449 games would be more content and worse picks.
04Methodology and caveats
Methodology: walk-forward testing with no retraining or tuning on 2026 games; every bet graded against the closing spread at a flat 1 unit and a -110 price; pushes counted as pushes, not wins or losses. The model is score-based, built on prior-season data, and its weights were fixed before the 2026 season.
Sample, stated plainly: this is a historical paper backtest of 124 ranked-game bets from one partial season. It is not this site's live pick record, and a 124-bet sample can swing on a handful of results: roughly four flips would erase the profit. Read the number as evidence about where the model works, not as a promise about next Saturday.
Compromises: closing lines come from a single book (DraftKings, via ESPN) rather than a multi-book median. The model's features are score-based only, with no EPA or play-by-play data, which limits what it can know beyond final scores and margins. No opening lines were available, so the line-move feature was 0 for every game. ESPN week numbering merges the first two weekends into Week 1.
Past performance in a backtest does not guarantee future results. If the ranked-game edge decays, this page gets updated and the coverage changes with it. The study is a snapshot of what the data said in October 2026.