Everything is built from public play-by-play and roster data
(nflverse, CC-BY 4.0). For each
player we estimate volume (targets, carries, attempts) and efficiency from
recent games, weighting the recent past most — a game from a month ago counts
roughly half as much as last week's. Opponent adjustments rebuild every week
from the same data, so by mid-season they describe this year's defenses, not
last year's.
The projection is a distribution, not a point. That's what lets us
quote a floor, a ceiling, and the probability of clearing any line you name —
the same numbers serve a lineup decision and a prop question, because the unit
is a reception, a yard, a touchdown probability.
The record
Every projection is locked at the last publish before kickoff and
never revised. After the game, the same page shows the locked number against
what happened. Locks are written once; a missed lock stays missed rather than
being filled in later. The record contains no back-filled history — it starts
at the first number a stranger could actually load, and grows one week at a
time.
What we deliberately don't do
Injuries and inactives don't move the number — yet. We tried the
adjustment everyone assumes: half the vacated share to the next man on the
depth chart, first week of the spell. It made receiving projections
worse. We show you the evidence — what actually happened to each teammate's
usage when a player sat — instead of guessing. If a player is out, his card
says so, and says the number assumes he plays.
No live bias correction. Recent form and opponent data update every
week; we do not apply a live self-correction on Saidmark. Until a correction
proves itself out of sample, the published number stays raw.
Weather is honest but crude. Sustained wind of 15+ mph outdoors
trims passing about 7%. Below that, and for temperature, we currently apply
nothing — and we'd rather tell you that than pretend otherwise.
Some intervals are known to be too wide. On high-volume receivers,
our stated ranges have caught more outcomes than a 50% band should, in every
season we measured. The cards say so. We publish the miss instead of hiding
the flag.
Tested and rejected
Most ideas that sound like improvements aren't. Recent candidates, each fit
on one season and judged on held-out years, each rejected on evidence:
Folding a receiver's target depth (aDOT) into the projected mean — degraded accuracy on the fit year.
A depth-of-target mix adjustment — helped low-volume receivers, hurt high-volume ones. Not shipped.
Retuning the opponent-adjustment strength — a fit-year gain that shrank tenfold out of sample.
Two families of interval-width fixes — both failed a proper scoring rule.
A high-volume mean correction — improved the fit year, failed the held-out years.
The injury next-man-up rule above.
A projection source that never shows you its dead ends is asking you to take
its live ideas on faith. Ours are listed.