Most projection systems start from a player’s efficiency: his yards per carry, his catch rate, his touchdown rate, multiplied by an expected workload. We measured how much of each of those rates is actually the player, and the answer shaped everything else on this site.
How much of a rate is the player
The test is the intraclass correlation (ICC): the share of the game-to-game variation in a rate that belongs to the player rather than to the game. An ICC of 1 would mean a player’s rate is the same every week; 0 would mean it is pure noise around the league. From it follows k, the number of games of a player’s own history it takes before his rate deserves as much weight as the league’s.
| Rate | Position | ICC | Games to trust (k) |
|---|---|---|---|
| Yards per carry | RB | 0.097 | 9.3 |
| Yards per target | WR | 0.113 | 7.9 |
| Yards per target | TE | 0.128 | 6.8 |
| Catch rate | WR | 0.160 | 5.3 |
| Rush TD per carry | RB | 0.102 | 8.8 |
| Rec TD per target | WR | 0.112 | 7.9 |
Roughly nine parts in ten of a running back’s yards-per-carry swing from week to week is the game, not the back: the defense he faced, the blocking, one long run.
Own rate against the league’s
The stronger test: give two forecasts the same number of opportunities, one using the player’s own prior rate and one using the league rate for his position, and see which misses by less. In this comparison the league rate wins all four rows.
| Forecast | Own prior rate | League rate |
|---|---|---|
| RB rushing yards | 17.71 | 16.86 |
| WR receiving yards | 21.50 | 20.68 |
| TE receiving yards | 15.78 | 15.07 |
| RB receiving yards | 14.06 | 13.00 |
Forecast error in yards; lower is better.
That is not the whole story. Built over three seasons and shrunk toward his position, a player’s own yards per opportunity did beat the league rate for receiving yards in 2024 (13.61 against 13.65 yards of error) and 2025 (13.40 against 13.46), and for rushing yards in 2024 (11.29 against 11.42); it lost on rushing yards in 2025 (11.80 against 11.68). His own catch rate beat the league’s in both seasons. Own efficiency is not worthless; it is small, and it does not hold every season.
What the engine does with it
So the engine does not project yards from a player’s own yards per target or per carry. It projects them from his role. Every target and carry has an expected value given how deep the throw is, where it goes, the down and the field position. A player’s average expected yards per opportunity says how he is used, and unlike his results it repeats: over three seasons and shrunk toward his position, it beat the flat league rate for receiving and rushing yards in both 2024 and 2025, and beat his own shrunk rate in three of those four. A receiver who is thrown deep projects more per target than one who works underneath, and a big game on five targets still barely moves him. Catch rate keeps a small pull toward a player’s own season (k = 5.3 games), which also beat the flat rate in both seasons. Receiving yards are then scaled to the team’s quarterback’s projected passing (a receiver with a line counts at his mean there, so the yards shown do not add up to it exactly), because the one efficiency that does repeat is the passer’s (see the quarterback engine on the record page). A yards projection is therefore about as good as the volume and the quarterback underneath it, and the site says so rather than dressing a volume forecast up as an efficiency model.
Touchdowns looked like the exception: touchdowns per touch are noise, but red-zone role seemed like a repeatable place to find them. We tested it. Season-to-date red-zone share forecast anytime touchdowns worse than plain touch share in both 2024 (log loss 0.4685 against 0.4553) and 2025 (0.4438 against 0.4370). So a player’s touchdowns are his share of the team’s projected touches times the team’s expected touchdowns, and red-zone share stays on the page as context only.
The practical read: when a receiver has a big yardage game on five targets, the board barely moves. When his targets go from five to nine, it does.