Every projection on this site comes from NOVA-STARS™, a rating engine originally built for Nolensville youth football and adapted here for college and professional football. This page explains how to read it. It deliberately does not publish the formulas, constants or fitted coefficients — those are proprietary.
The core idea: grade against expectation, not against results
Most power ratings ask “how much did you win by, and against whom?” NOVA-STARS asks a different question: did you do better than you were expected to?
Every completed game is scored against the number the betting market had set beforehand. Beat that number and your rating rises; fall short and it falls, even in a win. A team that wins by 3 as a 17-point favourite has told us something bad about itself, and this model listens to that. Where no betting market exists — Division II, Division III — the model grades against its own pre-game projection instead, which is a weaker signal and is labelled as such.
What each number on a game card means
- NOVA-STARS: TEAM by 7.4
- The projected final margin, and which side it favours. This is a margin, not a prediction of the score.
- 74.1% win
- The chance that side wins outright. Derived from the gap in Titan Strength between the two teams, so it accounts for how far apart they are rather than just who is ahead.
- PIT 73 B · BUF 84 A-
- Titan Strength and grade for each team. Titan Strength runs 50 to 100 within a division, where 100 is the strongest team in that division. It is not comparable across divisions — a 100 in Division III is not a 100 in the SEC. The letter is a composite grade blending strength with scoring margin, efficiency, dominance and how consistently a team beats expectation.
- 4.4 PT EDGE · BUF · PRESEASON
- The gap between our projection and the sportsbook line, and which side that gap favours. “PRESEASON” means no games have been played yet this season, so the rating rests entirely on last year plus roster information — the least reliable state it will ever be in. In that state we also cap how strong the chip is allowed to look, so a large preseason gap cannot masquerade as a strong signal.
- 2/3 models agree · 3.5+ pts
- Three independent models run where their inputs exist: ours, ESPN’s FPI, and SP+. This says how many landed on the same side of the line, and the smallest margin among those that agreed. Independent agreement matters far more than one model shouting.
- only NOVA disagrees
- The warning case. Our model is far from the line while the others sit near it. When one model stands alone against both the market and its peers, the model is usually the one that is wrong. Hover any chip for the full reasoning and all three projections.
What goes in
- Last season’s results, solved so that beating a strong opponent counts for more than beating a weak one, with blowouts capped so a 70-point win doesn’t distort the picture.
- Recruiting talent, averaged across the classes plausibly still on the roster. Only applied where a division is well covered — it is not applied to Division II or III at all.
- Returning production, as the closest available stand-in for the age curve the original youth model used. Available for FBS only.
- Home field, measured separately for each division from actual results rather than assumed. It is worth noticeably more in college than in the NFL.
- Every completed game, graded against the pre-game number as described above.
Yes, it is self-adjusting
The ratings update every day. Each finished game is compared with what was expected of it, and both teams move accordingly. Early in a season the ratings lean almost entirely on prior-year information; as games accumulate, the current season progressively takes over, and by roughly the sixth week it dominates. The division-level calibration retunes on the same schedule, so the relationship between a rating gap and a projected margin is re-estimated from what has actually happened rather than fixed in advance.
Adaptation speed is itself a fitted quantity. The original youth-league setting moved ratings very slowly, which suited a twelve-game season; measured against college results it was too slow, and it has been retuned.
Where it falls short — measured, not guessed
We publish the model’s scorecard on the model performance page, including the cases where it loses. As of the most recent test the model is less accurate than the closing line and does not beat the market. That is the honest position, and it is on the site because a projection you cannot audit is worth nothing.
Specific known limitations:
- It knows nothing about injuries, suspensions, transfers, weather or depth charts. The market prices all of that. This is the single largest reason it disagrees with a line, and usually the reason it is wrong when it does.
- Preseason ratings are weak. Roster turnover in college football is severe, and last year’s team is often not this year’s team.
- Cross-division games are harder. When a Division I team plays a lower division, blowout caps hide how large the real gap is. This is corrected for explicitly, and the correction is fitted on actual cross-division results, but such games remain the least certain projections on the board.
- Division II and III share one pool. The schedule feed publishes them as a single group, so they are rated together even though the conference data can tell them apart.
- Preseason NFL games are excluded from the ratings entirely, and no agreement signal is shown for them. Starters barely play; the results are noise.
What this is not
It is not betting advice, and it is not a tout service. Nothing here is a recommendation to wager. The numbers exist so you can see what a transparent, measurable model thinks and how often it turns out to be right — including, at present, that it is not beating the closing line.