What FPS measures, how it is solved, and — the part most pages of this kind leave out — how accurate it has actually been.
Modeled · Measured · Published
FPS is a team's expected margin against an average FBS team on a neutral field, placed on a fixed 0–10 scale. It is a predictive rating, not a résumé — it estimates how good a team is now, not what it has banked.
The anchors are frozen. They were computed once, from the distribution of published ratings across every week since 2022, and they do not move. That means an FPS only changes when that team plays. A team can never slide down the scale because somebody else had a good Saturday.
Beside FPS the site prints Points — the same rating in its raw units, running from -24.9 to +26.9 this season. Points is the working number: subtract two of them, add home field, and you have a projected game. FPS is that value made readable.
The scale spans all 138 FBS teams across 11 conferences, never a partial field, which is what makes a gap in FPS mean the same thing wherever you read it.
Fathom is not a blend of other people's ratings. It is one calculation on one input: the scores of games that have been played.
Every result is a statement about two teams at once — beating a good team by 10 and beating a bad team by 10 are different facts. The solver looks for the set of team values that best explains every margin simultaneously, adjusting each team for the quality of the opponents it actually faced, and iterates until the numbers stop moving. That fixed point is the rating.
Because it is solved from margins rather than assembled from components, there are no weights to tune and no judgement calls about how much any signal should count.
Every FBS game since 2022 — who played whom, where, and the final score.
Each team's numbers read against the teams it actually faced, solved jointly rather than applied as a correction afterwards.
Last year's own rating, entered as two virtual games, fading as real games accumulate.
No market number touches the team or player ratings. They appear only when grading a forecast after the fact.
In September there is not enough football played to let three games speak for a team. Rather than discarding what was known in August, each team is anchored to its own rating from last season, weighted as if it were two extra games.
The anchor is not a separate preseason model or a recruiting estimate. It is the same Fathom rating, computed the same way, from the previous year's results. As real games accumulate it is outweighed automatically — by week six the current season is doing nearly all the work, with no scheduled handover and no weekly weights to publish, because there are none.
This was measured rather than assumed: the anchor is worth 1.2 to 2.5 points of accuracy in the opening weeks and fades to nothing by week six.
The rating splits in two, so the site can show where a number comes from instead of publishing one figure and asking you to trust it.
Off is points generated above an average offence. Def is points prevented against an average offence, so a positive Def is a good defence — the site signs it so that higher is always better on both sides.
The two are held exactly consistent with the top line. Across all 138 teams, the largest disagreement between Off + Def and Points is 0.000. The units can never contradict the rating they came from.
Everything above is description. This section is the evidence, and it is the reason this page exists.
Every figure below comes from walk-forward validation: the model is rebuilt using only games played before the week being predicted, then scored on games it has never seen. Nothing is fitted and reported on the same data. 1,659 held-out games, 2023–2025, weeks 5 onward, both teams FBS.
| Predictor | MAE |
|---|---|
| Market closing spread | 12.02 |
| Fathom projection | 12.54 |
| Season average point differential | 13.88 |
| Home team by home-field edge | 15.27 |
| Every game a tie | 15.52 |
| Predictor | Record | Accuracy |
|---|---|---|
| Fathom projection | 1194–465 | 72.0% |
| Season average point differential | 1150–509 | 69.3% |
| Better win percentage | 1122–537 | 67.6% |
| Home team always | 949–710 | 57.2% |
The comparisons are to rules that cost nothing, because that is the only bar that matters: a forecast is worth something if it beats what you could do without it. Fathom is better than the best free rule by 1.34 points of margin error and 2.7 points of winner accuracy, and sits 0.52 points off a closing line set by a market with real money behind it. Every one of those gaps is statistically significant.
FPP — Fathom Player Points — is what a player's production was worth in points per game, after subtracting what the defences he faced give up to everyone else. The same solver that rates teams against margin is run against yards and touchdowns, then each player's line is adjusted in proportion to his share of his team's production that day.
Yards and touchdowns are converted to points by weights fit from four seasons of team results, not borrowed from a fantasy table. FPS places that value on the same 0–10 scale, read within a position rather than across all of FBS.
Measured the same way, on 38,570 held-out player-weeks: the raw per-game average misses by 3.939 points, the opponent adjustment cuts that to 3.851, and the prior anchor to 3.798. Both improvements hold in every season and at every position taken separately. Players with no prior season start at replacement level rather than average, which beats a position-average start 3.374 to 3.445 across 8,085 debut weeks.
Compare quarterbacks to quarterbacks. A completed pass is credited in full to the thrower and in full to the catcher, so FPP is honest within a position and meaningless added across a roster. 1,247 players are currently rated across 4 positions.
This model was previously recorded as 2.07 points worse than the market on margin. The real figure is 0.52.
The evaluation had been running over every game in the source data — roughly 3,700 a season, of which only about 750 are FBS against FBS. The solver holds ratings for FBS teams only, so on the rest both sides fell back to a placeholder and the projection collapsed to a constant. One such game projected a team to score −5.1 points.
The reason it went unnoticed is worth stating: the comparison required each game to have a posted betting line, and that requirement had been quietly acting as the filter deciding which games were graded. Removing the market as the benchmark forced the population to be stated explicitly instead of inherited. Every validation on this site now names the games it was run on.
It is a measurement of team and player strength, solved from results, graded against what you could do without it, and published with the misses attached.