Predictium's college football model is live for the 2026 season. Every game involving one of the 138 FBS programs gets a projected score, spread, total, and win probability, shown next to live Bovada and FanDuel lines wherever a line is posted. This week's board has 99 games on it. The ratings board covers all 138 teams. (We call it the Predictium 136. The name predates the two new programs and it stuck.)
That's the what. The more unusual part of this post is the list of things the model isn't.
It isn't a black box with a hot streak
Our backtest is walk-forward: for each season from 2017 through 2025, the model trains only on the seasons before it and predicts the one in front of it. Across 7,606 games, it is published in full on the backtest page, and it shows the market winning.
| Model | Closing line | |
|---|---|---|
| Spread error (mean absolute, pts) | 12.92 | 12.22 |
| Total error (mean absolute, pts) | 13.50 | 12.77 |
The closing line beats us by about 0.7 points on spreads. In July that gap was 1.04. We would rather you know the number than take our word that a model exists.
The same goes for the live record. Every pick we flag settles in public, and the current tally is 15-17 on 32 settled picks. Split it and the shape gets clearer: the model's leans, where it simply disagrees with the market by a couple of points, are 15-12. The five moneyline longshots it flagged as price plays are 0-5. All of that is on the games page and stays there.
We also track closing-line value on every lean from the first snapshot of a game to the last one before kickoff. It's negative so far: spreads have moved against us by an average of 0.46 points across 108 leans, with the market moving our way only 24% of the time. Totals are flat. That's the expected result for the first week of a season, when preseason lines move on roster news the model is only starting to ingest, but expected or not, it's published.
A note on words, because they matter here. A lean is a model-versus-market disagreement that we settle publicly. It is not a claim of an edge. The only picks that carry an expected-value claim are the price plays, and those are sized to win one unit, which means a +1400 play risks about 0.07 units. Nobody should read a 15-12 lean record as a reason to bet anything.
What it's actually built from
The ratings come from how football is played rather than from who won. Each team gets opponent-adjusted efficiency ratings from EPA per play, success rate, and explosiveness, split by rush and pass, with garbage time excluded so a 21-point fourth quarter against backups doesn't move anyone's number.
Before a season has real games in it, the ratings lean on priors: returning production, recruiting talent, transfer portal movement (over 4,400 moves ingested this offseason), and coaching changes, including the incoming coach's career record rather than just the fact of a change. The prior's weight fades as actual games arrive, and the team pages show that weight for every team. Pace identity carries across seasons, because a team that snapped the ball every 22 seconds last year usually still does.
Margins in college football don't follow a bell curve. They pile up on 3, 7, 10, and 14, so the model uses an empirical score distribution fitted to real historical results, and the probability of covering a 6.5 versus a 7.5 reflects that.
Inputs are public: College Football Data, ESPN, Bovada and FanDuel lines, and Open-Meteo for weather.
We audited ourselves before kickoff and found real bugs
In August we went looking for lazy "league average" assumptions in the model and found two that had been there every season.
The first: ratings were being pooled toward a single national average, which quietly under-rated power-conference teams against Group of Five opponents by about 6.6 points a year. The second: totals were running roughly 3 points hot, because college football scoring has declined since 2017 and the model hadn't noticed. Both are fixed, both are behind gates that would catch a regression, and both are written up on the about page.
The most fun one was North Dakota State. Their FBS debut initially put them third in our ratings, on the strength of a rating earned in FCS. When we checked every program that has made that move in our data, all of them landed below average in their first FBS season. The fix moved NDSU to 81st. The board is more honest for it, if less entertaining.
Humans research every game. The model keeps them honest.
New this week: a research pass on every game on the board, roughly 90 of them, starting with quarterback availability from the conference availability reports. Findings are filed as cited, categorical proposals. The model, not the researcher, converts each one into a point adjustment, and the adjustments are capped: a starting quarterback out is worth at most 4.5 points. The base number and the adjusted number are both recorded, so nothing gets quietly overwritten.
Week 1 produced 8 adjustments across 7 games, including the quarterback losses at Miami (Ohio) and New Mexico. The rule for the whole layer is the same rule the model lives under. If the research overlay hasn't outperformed the pure model after about 40 games, it gets pulled.
How we got here, and what's next
The repo was empty on July 12. The first production publish went out that night. August was a model sprint, the pipeline moved to a dedicated machine on July 28, Week 0 went live on August 29, and the research overlay shipped on September 3. Since the openers posted we've recorded around 8,000 line snapshots, on a cadence of a daily run at 10:00 UTC, weeknight runs, and four pre-kickoff windows on Saturdays.
Next is player props, built shadow-first. We'll run a props board internally through weeks four to six and publish it only if it beats the sportsbook lines on calibration. No date until it passes.
Where to look
The ratings board, every game against the market, the backtest, and how the model works. Top five on the board this morning: Ohio State, Oregon, Notre Dame, Georgia, Miami.
None of this is betting advice. The numbers are model output, the record includes the losses, and if you bet, do it where it's legal, at 21 or older, and with money you can afford to lose.