How the app's projections did this week, next to ESPN's, against what your players actually scored — every player on a roster who played, starters and bench, 9 this week. One week is noise; this page gets truer as the season goes on.
ESPN was closer this week: off by 8.1 points a player against 8.9. Within 5 points of the real score: model 33%, ESPN 33%.
This week's number — the app's own weekly projection, graded on the 9 players it had one for: off by 9.1 points a player, within 5 on 33%.
Average error: ESPN 8.1 · model 8.9. A typical projection is off by about 8 points either way — a 12-point projection means 4 to 20.
Starters (5): ESPN 8.5 · model 9.3 — ESPN closer
Bench (4): ESPN 7.5 · model 8.4 — ESPN closer
Week 3 of 17 · season record starts after Week 3.
2 weeks of evidence, plus Week 3 so far (9 players). A rule appears only when the numbers clear a floor; under it the line reads "even".
The average was off by 6.1 points a player against 6.1 for the best single number — not a rule yet.
26 such players; they scored 14.7 on average against ESPN's 15.3 and the model's 7.2. Player by player: ESPN closer on 16, the model on 7, even on 3 — but the level was ESPN's.
1 such player so far.
The model ran 1.9 points low per player (QB 0.9, RB 2.3, WR 1.6, TE 2.8). Shifting every model number by that made it worse: off by 6.5 against 6.3, within 5 on 43% against 52%. The low number pays for itself on busts.
Nothing cleared the floor this week.
Both lists are the players where the two numbers were 3+ points apart, sorted by how much closer the winner was.
4 players where ESPN and the model were 2+ points apart: ESPN closer on 2, the model on 2. ESPN was the higher number on 3 of them, the model on 1. Of the 2 gaps of 5+ points, 2 were ESPN-high — those players scored 23.5 on average against ESPN's 13 and the model's 6.1. When the two disagreed on which of two players to start, the model was right 0% of the time (0 of 2).
4 players scored 20+ this week, averaging 26.2; the projections had them at 14.6 (ESPN) and 14.5 (the model).
Start/sit: when two players at one position finished 3+ points apart, how often the source had the right one higher.
Each team's starters: ESPN's projected total · the model's · what they scored. Finished games only; K and D/ST count ESPN's number on both sides. Tap a team to see its read.
Starters scored 6. ESPN had them at 9.6, the model at 10.1 — both high. 8 starters still to play.
Starters scored 40. ESPN had them at 28.9, the model at 29.6 — both low. Beat both numbers: Christian Watson (+7.2 over the model), Jordan Love (+3.2 over the model). 7 starters still to play.
Starters scored 49. ESPN had them at 30.2, the model at 32.2 — both low. Beat both numbers: Bijan Robinson (+12.9 over the model). 7 starters still to play.
Starters scored 2. ESPN had them at 13.4, the model at 13.4 — both high. 8 starters still to play.
Starters scored 28. ESPN had them at 14.9, the model at 8.8 — both low. Beat both numbers: Drake London (+19.2 over the model). 8 starters still to play.
2 players ESPN had at 0 before kickoff; the model carried 14.1 points on them. That number is not a projection miss — the model is a per-week rate and knows nothing about Sunday; the availability haircut is where this is handled.
Everyone ESPN expected to play recorded a stat line.
ESPN is ESPN's projection for this week. Model is the app's season number — projected points a week over roughly the next month, the figure Rankings, the trade analyzer and Ask Claude use — so it is not tuned to this week's matchup. This week's number is the app's own weekly projection, graded only on the players it had one for.
Off by N points is the average miss over every graded player, either direction. A typical miss of 8 means a 12-point projection landed anywhere from 4 to 20.
Within 5 is how often the number was close enough to trust. Start/sit calls: for two players at one position who finished 3+ points apart, did the source have the right one higher?
Closer and even: a source is called closer only when the gap clears a floor — 0.3 points of average error, 5 points of within-5, or 5 points of start/sit calls. Under that it is even, whatever the decimals say.
Starters and bench are different questions. The headline grades every rostered player who played; By team is starters only, so the two can point opposite ways in the same week.
K and D/ST are left out: the model copies ESPN's number for them. 1 name the model doesn't carry is left out too. Only finished games count; a player who finished with no stat line is in Out this week instead.
Model numbers are the snapshot taken 2026-09-22, before this week's games.