How good are these forecasts?
Every forecast scored on seasons the model never saw while its settings were chosen.
Held-back seasons
| Forecast | Forecasts | Log loss | Favourite won |
|---|---|---|---|
| The model | 10786 | 0.489 | 75% |
| Base rate | 10786 | 0.682 | 57% |
| Recent form | 10786 | 0.561 | 71% |
| Betting market | 2608 | 0.463 | 76% |
Against the betting markets
Each market's price before the start, scored on the same finished games as the model. A negative figure means the market was more accurate.
| Market | Since | Priced | Model log loss | Market log loss | Model vs market |
|---|---|---|---|---|---|
| Bookmakers (Pinnacle, DraftKings, bet365) | 20 Jan 2026 | 1037 | 0.539 | 0.493 | −9.4% |
| Polymarket | 27 Nov 2024 | 2133 | 0.537 | 0.500 | −7.4% |
| Kalshi | 23 Aug 2025 | 2105 | 0.513 | 0.476 | −7.8% |
Every FBS game since 2001 moves the ratings, and so do FBS teams' games against FCS opponents; a team new to the file starts 200 points below average. Only FBS teams appear in the tables and the playoff.
Before each season a team keeps part of last season's rating, then two things known in the summer move it: the recruiting talent on its roster and its returning production. Both weights are fit on the tuning seasons; the track record below shows each one's effect on the held-out seasons.
The polls are shown on the Rankings page, not used by the model. Tested on 2026-10-01 on the tuning seasons (2016-2023): adding the preseason AP poll on top of recruiting and returning production made the tuning seasons worse at every weight tried, and the latest AP poll before each game as a pre-game adjustment was fit to a weight of zero. On the held-out seasons a small poll weight moved the log loss by less than 0.0001 either way. The voters mostly follow results the ratings have already seen.
The playoff simulation takes the best-record team in each conference as its champion (conference title games are not simulated yet), puts the five best of those in, fills seven at-large places, and seeds by simulated wins and then rating. The committee's judgement is not modelled; its real rankings and seeding are on the Rankings page.
The market comparison uses the closing spread, turned into a win probability with a 15.5-point spread of outcomes.
Returning production counts offence only (passing, rushing and receiving yards): ESPN's defensive box scores are too patchy before 2023 to measure the same way every season. Injuries, coaching changes and the depth chart are not known to the model.
Conference title games and the committee's ranking are approximated.
Questions and answers
How accurate are PodiumOracle's College football predictions?
On 10,786 games from held-back seasons (2024 to 2026), which the model never saw while it was tuned, the side it favoured won 75% of the time. Its log loss was 0.489, against 0.682 for the base rate and 0.561 for recent form, and 0.463 for the closing betting odds. Against the betting markets on the same finished games, its log loss was 0.539 against the bookmakers' 0.493 over 1,037 games, 0.537 against Polymarket's 0.500 over 2,133 games and 0.513 against Kalshi's 0.476 over 2,105 games (lower is better).