The record.
Every projection, range and call Fader makes is written down before kickoff and graded against what happened. This page is rebuilt from that ledger at every refresh, misses included, with its sample sizes.
- 26players on this week's ledger, week 3
- 12live start/sit decisions graded so far
- 0 of 29calls that changed side before kickoff
Projections: closer than a player's recent average.
Every week of 2025 projected from only what was known before it, then compared with the simplest real alternative, a player's own last four games. Fader landed closer on average at every position. How the model works.
| Position | Fader | Last four | Player-weeks | Closer that week | |
|---|---|---|---|---|---|
| QB | 6.5 pts | 6.9 pts | 6% closer | 543 | 53% |
| RB | 4.1 pts | 4.3 pts | 3% closer | 1,333 | 48% |
| WR | 3.7 pts | 3.9 pts | 6% closer | 2,133 | 50% |
| TE | 3.1 pts | 3.3 pts | 7% closer | 1,077 | 52% |
Ranges, checked in the open.
Every projection comes with a likely range, aimed at eight weeks in ten. Fitted on 2022 to 2024, then checked on every week of 2025, projected from the weeks before each one. This is how often the real score landed inside, for projections of 3 points or more: QB and RB too narrow, WR and TE too wide. A range passes when the 90% interval contains 80%. The table keeps every band.
- QB 77% 520 player-weeks
- RB 77% 924 player-weeks
- WR 84% 1,448 player-weeks
- TE 83% 625 player-weeks
| Position | All | 0–3 pts | 3–5 pts | 5–10 pts | 10–15 pts | 15+ pts |
|---|---|---|---|---|---|---|
| QB | 75%n 54373–78 | 43%n 23too few | 73%n 11too few | 78%n 5571–86 | 77%n 26473–80 | 77%n 19073–81 |
| RB | 80%n 1,33378–82 | 86%n 40983–89 | 76%n 19270–81 | 77%n 32574–80 | 80%n 32376–83 | 71%n 8464–80 |
| WR | 85%n 2,13384–86 | 87%n 68585–89 | 89%n 39486–92 | 84%n 72982–86 | 78%n 30074–82 | 84%n 25too few |
| TE | 86%n 1,07785–88 | 91%n 45289–93 | 85%n 20081–88 | 82%n 36778–85 | 84%n 5576–90 | 100%n 3too few |
Below 5 projected points the floor is often 0, and a 0 counts as inside, which is why low projections run over 80%.
Live, 2026
Recorded before each player's kickoff since week 3, graded once each week's stats settle.
0 player-weeks graded. Too few to judge: a rate is shown from 100.
This week, on the record.
26 players recorded for week 3
QB 3 RB 6 TE 5 WR 12
In 4 rows so far, the newest written Mon 28 Sep, 7:06 PM ET. A new row is written whenever a projection moves, and each player is graded on the last one before his own kickoff.
| Player | RB D'Andre Swift |
|---|---|
| Projection | 13.0 points, half-PPR |
| 80% range | 4.8 to 23.1 |
| Recorded | Mon 28 Sep, 7:06 PM ET |
| Kickoff | Mon 8:15 PM ET |
| Graded | after week 3's stats settle, on his last entry before kickoff |
Replay, 2025: stated probability vs what happened
- 50–55% 51%326 calls, 90%: 46–55
- 55–60% 56%290 calls, 90%: 51–61
- 60–65% 61%274 calls, 90%: 56–66
- 65–70% 71%192 calls, 90%: 66–76
- 70–75% 74%159 calls, 90%: 67–79
- 75–80% 73%100 calls, 90%: 65–80
- 80–100% 84%82 calls, 90%: 76–90
Bar: how often the favored player won. Notch: the average probability Fader gave. 1,423 calls, 0.5 PPR, Brier 0.2242. Partly in-sample: the projections are walk-forward, but the error shapes behind the probabilities were measured on 2025.
Start or sit: the probabilities, checked.
Replayed across the 2025 season, when Fader said a player would win 60 to 65% of the time, he won about that often, and so on up the scale.
329 replayed calls were too close to call. The side it would have leaned to won 51.1% of them, a coin flip, which is why it did not pick.
Live, 2026
Each decision counts once, on the call that stood at kickoff.
12 decisions graded. Too few to judge: a rate is shown from 100.
They come from 1 league-week, which is closer to one observation than to 12.
0 of 29 calls changed side between their first recording and kickoff.
20 rows were written after every pair had locked. They are not calls and are left out of every rate.
Draft availability, 3.0 times sharper than ADP.
Before each of 90 picks, Fader gave every player a chance of still being there. Most of those 36,994 claims are near-certain and easy, so the ones in doubt are scored on their own: 3.0× less error than ranking by ADP.
Error on the picks in doubt Brier score, lower is better
6,924 claims between 2% and 98%, 90% interval 0.092–0.133 resampling picks.
Recorded, waiting to be graded.
- 400Waiver claims recordedjudged over the four weeks after each claim; first gradeable 14 October 2026
- 1,960Alerts recordedjudged once the news they flagged plays out; first gradeable 5 October 2026
Where it was wrong
- The ranges, until 24 September 2026. One fixed width per position made them too wide for starters (90 to 96% inside for backs projected 10 or more) and too narrow for fringe players (62 to 65% for backs and receivers projected 3 to 5). They are now fitted by projection level; the numbers above are the result, checked on a season the fit did not see.
- An earlier claim on this site. It said the ranges held "79.6 to 81.2% at every position". That was only true for projections of 3 points or more, averaged over two opposite errors.
What this does not tell you
- Whether Fader beats your platform's projections. There is no head-to-head test, so no claim is made.
- Anything about a league before it was connected. Calls are recorded from connection on.
- Player-specific risk. Two players at one position with the same projection get the same range; the range is the honest width of a projection, not a read on the player. We tested whether a player's touchdown dependence, usage volatility, snap share or games of history predict when his range misses (2022 to 2024, 10,625 player-weeks, tests fixed in advance). None did in the direction predicted, so ranges stay by position and projection level.
- Kickers and defenses. They have projections but no tested range, so none is shown here.
How a call is graded.
Recorded before kickoff
Every projection, range and open start/sit call goes into an append-only ledger. A player's entry is the last one before his game.
Checked against the clock
Start/sit calls are audited against kickoff times from the NFL schedule, and the projection grader refuses anything written after kickoff.
Graded when the numbers settle
A weekly call is scored once the week is played and the stat corrections window has closed, the Thursday after.
Counted once
A player-week, pick or decision counts once, on what stood at kickoff, however often it was re-recorded.
Published, misses included
Below 100 graded claims a section shows the count and no rate.
- Graders
- abstain-v1, brier-availability-v2, brier-startsit-v1
- Range and accuracy check
- ffa coverage --season 2025, ranges 9a86fdced759
- Code
- commit bf1f43118739
- Data
- newest fetch 28 Sep 2026, 23:02 UTC; built 28 Sep 2026, 23:08 UTC