How Atlas works
Five phases of research sit behind every card. This page is the short version, and it is free because the evidence for a claim should never cost more than the claim.
The short version
Atlas measures college football with an opponent-adjusted, point-in-time database: every figure attached to a game uses only information that existed before kickoff. The model is fitted on completed seasons and scored on seasons it has never seen.
The projection is the model's own. Each team opens the season at a preseason expectation built from last season's ratings, talent, recruiting, returning production and the coaching situation; after every game a filter updates its offense and defense, opponent-adjusted; the total adds pace and wind; and the score is the mean of a full distribution over every possible final, to one decimal.
Then it does something most models do not. It compares itself to the market, and it publishes what that comparison found. Walked forward over five seasons it never saw, Atlas's number is closer to the final margin than a rating system or a preseason ranking and not as close as the closing line. The market is not an input to the model; it is the reference the grade is measured against.
What a grade means
A grade is not a recommendation. It answers one question: how much weight does the information on this card deserve?
It is computed, never assigned. Four components, one hundred points:
| Component | Points | What it measures |
|---|---|---|
| Calibration | 45 | how far short of its claim a card this far from the market has historically fallen, read from a curve fitted to every completed season out of sample |
| Market agreement | 25 | how far Atlas's number sits from the market on the spread |
| Card conditions | 15 | how mature the season is, and how settled the market has been since the number opened |
| Data completeness | 15 | how many of the required inputs are present |
| Score | Grade | Share of the record |
|---|---|---|
| 96–100 | A+ | 6% |
| 90–95 | A | 14% |
| 79–89 | B | 31% |
| 66–78 | C | 25% |
| 51–65 | D | 15% |
| below 51 | F | 10% |
The six thresholds were chosen once, from the distribution of scores across the out-of-sample record, and then fixed. They are absolute: a card's letter depends on that card and on nothing else on the board, so the same card grades the same on a quiet Tuesday and on championship Saturday. Atlas does not grade on a curve, because a curve would make a screenshot mean something different depending on the week it was taken.
A top grade is not a signal to read that card first. Cards where Atlas and the market agree to within a point have realized 50.8% against a 51.3% claim across 760 games — statistically a coin flip. They grade highest because they are the most reliable, and they are the most reliable because Atlas has added nothing to them. The grade tells a reader what to discount, not what to look at.
Why large disagreements lower confidence
This is the least intuitive thing Atlas publishes, and the most important. A model that disagrees violently with the market looks exciting. Measured across 3,864 out-of-sample games, it is the opposite.
| Disagreement | Games | Claimed | Realized | Gap | Seasons above 50% |
|---|---|---|---|---|---|
| 0-1 | 853 | 52.1% | 49.5% | −2.7 pts | 1 of 5 |
| 1-2 | 739 | 53.8% | 49.9% | −3.9 pts | 3 of 5 |
| 10+ | 116 | 79.0% | 49.1% | −29.8 pts | 1 of 2 |
| 2-4 | 1,123 | 57.1% | 48.6% | −8.5 pts | 1 of 5 |
| 4-6 | 634 | 61.9% | 48.3% | −13.5 pts | 1 of 5 |
| 6-8 | 267 | 66.4% | 48.7% | −17.8 pts | 1 of 5 |
| 8-10 | 132 | 71.0% | 53.4% | −17.6 pts | 2 of 4 |
Read the gap column. Where Atlas barely disagrees with the market, its claimed accuracy and its realized accuracy nearly match. Where it disagrees by ten points or more, it claimed 79% and delivered 49%.
So the loudest cards get the lowest grades. Every other product in this category shouts hardest exactly where its model is weakest, because a big disagreement makes a compelling post. Atlas measured that and does the opposite, in public, on every card.
Calibration
Calibration asks whether a claimed 65% is a real 65%. Against the closing number Atlas is overconfident, and gets more so as it gets more confident: its top confidence bucket claims 79% and realizes about 49%.
The grade corrects for this on every card. It is built from the calibration record itself, so a card in a badly-calibrated band cannot grade well no matter how interesting it looks.
The claimed number is still shown, beside the realized one, because a correction you cannot see is a correction you cannot check.
The second finding, kept separate
Atlas measured disagreement twice, against two different reference points, and got two different answers.
Distance from the closing number predicts the model's own failure — that is the calibration table above, and it drives the grade. Distance from the opening number predicts which way the line will move, which is a fact about the market rather than about the football game.
A model far from the close is wrong. A model far from the open is early. Atlas tracks the second finding in an open, pre-registered record with its own kill criteria, and deliberately keeps it out of the grade.
What Atlas does not do
Atlas does not publish selections. Not as a lean, not as an arrow, not as a highlighted row. Five phases of research established that Atlas has no edge worth acting on, and a product that implied otherwise would be saying something its own reliability record contradicts.
59 cards are published this week. None of them tells you what to do with the information.