Fundamentals
Games Played: How a Projection Model Prices Injury Risk Without Guessing
Quick answer
How does a projection model account for injury risk?
Ours turns a per-game rate into a season total using an availability prior: how many recent weeks each player recorded a touch. The median prices expected missed time, and the P10 end of the range carries the season that ends early.
The Second Number Inside Every Season Total
A season projection is one number doing two jobs. It carries a per-game rate, what a player produces in a week he records a touch, multiplied by the number of such weeks the model expects. Drafters argue about the rate constantly, and the count gets inherited without inspection.
Our season model builds that count from availability priors: the weeks each player recorded a carry, a target, or a pass attempt over recent seasons, weighted toward the most recent one. The methodology page places it in the fourth step, where games and a position-specific age curve turn per-game rates into season totals.
Weighting toward the latest season has a practical edge. A player who missed half of two seasons ago and then played a full slate carries mostly the full slate into his count, while a player trending the other way carries mostly the missed time.
The definition matters because a touch is a role event as much as a medical one. A backup who dressed and never saw the ball records the same zero as a player on injured reserve, so a low count can come from a thin role or a fragile body, and the median alone will not say which.
The draft board publishes each player's median and P10-to-P90 range, and the games count stays folded inside those numbers with no column of its own. That makes a second, manual discount easy to stack on top of the first without noticing.
Why the Discount Lands Hardest at the Top
Points lost to absence equal the per-game rate times the weeks removed. Take the same expected weeks from two players and the one with the higher rate gives back more, in exact proportion: a 16-point-per-game player loses twice what an 8-point-per-game player does.
Take four backs off an invented stretch of board, with numbers made up to show the shape and tied to no current run. Back A projects 15.0 points per game, Back B 14.2, Back C 13.0, and Back D 12.4. Trim a game and a half of expected availability from A and from D alike, and A gives back 22.5 points while D gives back 18.6 for the same missing time.
Real counts differ by player, and that is where boards reorder. At a full seventeen games the four totals come out 255.0, 241.4, 221.0, and 210.8, with A 13.6 points clear of B. Attach expected games of 14.5 for A, 16.2 for B, 15.8 for C, and 16.6 for D, and the totals become 217.5, 230.0, 205.4, and 205.8, with every per-game rate left where it was.
B passes A by 12.5 points, a swing of about 26, and D edges past C by 0.4. The largest gap in the stretch moves too: on seventeen games it sits between B and C, a 20.4-point drop, while on expected games it becomes the 12.5 points between B and A, and C and D collapse from 10.2 points apart into a coin flip.
The Band Carries the Season That Ends Early
An expectation is an average, and an average of a healthy season and a torn ligament in September describes neither. The median prices the expected missed time. A season that ends in October lives in the bottom of the range, and because injuries take weeks away in clumps, the P10 end of a projection sits well below its median.
The band is also where the season model has missed its own target. The P10-to-P90 range is built to hold 80 percent of real outcomes, and across four held-out seasons it held 66.7 percent in 2022, 80.7 in 2023, 74.2 in 2024, and 83.2 in 2025. Two of the four seasons fell short, both published on the methodology page, and for a drafter the costly outcomes sit below the range.
The range comes from how real seasons landed for players at the same position with a similar projected workload. A player's own injury history reaches his projection through the games count, and a player you believe is more fragile than his peers is a judgment you apply to his P10 yourself. Read the P10 before the rank, especially across positions, where similar medians can sit on very different low ends, and the variance page makes the full case for drafting on the width of a range.
Drafting and Starting Against It
Double-discounting is the failure this setup invites. A drafter reads a season median, remembers the player missed time last year, and shaves the number again, charging him a second time for weeks the median already priced.
A discounted price can still be a buy. An expected count below seventeen is a price adjustment, and a correctly discounted player at the right pick is value. Backfilling the weeks he is expected to miss is a roster question with its own arithmetic, and the handcuff piece carries it end to end.
Weekly lineups run on a faster clock. The weekly projections take their cue from the current injury report and the weekly model reruns daily, so a late designation reaches the week's numbers long before it touches a season median, which moves only when the season model is rerun by hand. Read the season backtest on the methodology page and the weekly ledger, then decide how hard to lean on each model, including this piece.
FAQ
Does a lower games expectation mean the model predicts an injury?
No. The count only looks backward, so it holds no opinion about next month's hamstring and no way to know which Sunday goes wrong. It prices a pattern and leaves the diagnosis to the injury report.
Why does a newly promoted starter's projection look light?
His games count still carries backup weeks with no carry or target as zeros, and his per-game rate still reflects the old job. Both halves of his median describe the role he left, so the distance to the new role is yours to judge.
What does a Doubtful tag do to the weekly number?
It takes him out of the week. Players listed Out, Doubtful, on IR, on PUP, or suspended sit out the weekly projections entirely, so a Doubtful player who suits up anyway was never priced for that game.
How does age interact with availability in a projection?
They share a step and stay separate inside it. The age curve trims per-game output past an age threshold set for each position and never touches the games count, so identical recent counts can still yield different medians. The backtest never exercised the curve.
Can I compare a rookie's availability with a veteran's?
Only loosely. A rookie has no NFL weeks to count, so his games count and his rate both come from a prior shared by rookies at his position, a limitation the methodology page names. Read a rookie's median as a placeholder with a jersey number attached.