Atlas grades its own numbers.
Every college football game gets a card: what the market says, what Atlas projects, what is driving the difference — and a letter saying how much that card's information has historically been worth. Atlas never tells anyone what to do with it.
What Atlas is
A research desk, published
Atlas models 59 college football games a week from a point-in-time database — every figure uses only what was knowable before kickoff, going back to 2018.
The number is Atlas's own. Every team starts a season at a preseason expectation and is updated after every game, opponent-adjusted, and the projection is the mean of a full score distribution, to one decimal. The market is never an input: it sits beside Atlas's number so you can see where they differ and what is driving it.
And then it does the thing nothing else in this category does: it grades itself, in public, on every card, using its own historical record.
What Atlas is not
- Not a selections service. No card names a side. Not as a lean, not as an arrow, not as a highlighted row.
- Not a sportsbook. Nothing here can be acted on from this page, and nothing is sized.
- Not a record of wins and losses. The record Atlas publishes is calibration: what it claimed, against what it delivered.
- Not urgent. No countdowns, no alerts, nothing on any Atlas page moves.
How to read a card
eight things, in the order you meet themEvery card is the same shape. The first screen answers five questions — who is playing, what the market says, what Atlas says, why, and what should make you careful — and everything else is one tap below it.
- 1The game
Both teams with their crests, rank, record and conference, then kickoff, broadcast and venue. You should recognize the game before you read a word.
- 2The market
What the betting market currently says — the spread and the game total. This is the reference everything else is measured against, not a price to act on.
- 3Atlas
What Atlas projects: the score to one decimal, and the total underneath it. The model's own number - the market is not an input to it.
- 4The difference
How far Atlas sits from the market on the spread, with the total beneath. This is the number the grade is mostly about — and a large one is a warning, not an opportunity.
- 5The grade
A letter, a score out of 100, and three lines explaining itself: what the letter says, what Atlas did, and the out-of-sample record behind it.
- 6Why
The three things the model is reading, each with the crest of the team it favors.
- 7Be careful about
Up to three warnings computed from this game — a thin market, a lopsided spread, a missing metric. A card with nothing to flag says nothing.
- 8Everything else
Market detail, the projection's arithmetic, how the grade was computed, all the drivers, how the line has moved, and the reliability record — each one tap away.
Worked example: Virginia at Wake Forest, graded A+.
What the grades mean
computed, never assignedA grade answers one question: how much weight does the information on this card deserve? It is not a rating of the game and it is not a recommendation.
| Grade | What it means | Share of the record |
|---|---|---|
| A+ | Atlas and the market land on the same number. Reliable — and Atlas is adding least here. | 6% |
| A | Closely aligned. Strong calibration. | 14% |
| B | A moderate difference, in the range where the claim and the delivery stay close. | 31% |
| C | A wide difference. The claim starts to run ahead of the delivery. | 25% |
| D | A large difference. Atlas commonly struggles this far out. | 15% |
| F | A very large difference, in the range where Atlas has been least reliable. Atlas marks its own card down. | 10% |
The thresholds are absolute and were set once from the out-of-sample record of results. The same card grades the same on a quiet Tuesday and on championship Saturday, so a screenshot means the same thing whenever it was taken.
The full rubric, the four components and the calibration curve behind them.
Why Atlas exists
Every model in this category publishes its numbers with the same confidence every week. None of them tells you which of those numbers has historically been worth anything.
Atlas measured that, and the answer was uncomfortable: the further its model sits from the market, the worse it does. Across every season it never saw out of sample, cards claiming 77% accuracy delivered 50%. The loudest cards are the weakest ones.
Most products would bury that. Atlas made it the largest element on the card. A grade that can say F is the only kind of grade worth anything, and it is the reason the A means something too.
The out-of-sample record, the methodology, and every number behind this page · questions.
Start here
This week's board · how Atlas works · what grades mean · what is free and what is not
Everything that makes Atlas checkable — the grades, the research, the methodology and the reliability record — is free and always will be.