The injury tax is not on his production. It is on his calendar.

What this means for your roster. When a guy comes back, he is basically the guy. Stop shaving your projections for rust, and stop paying somebody else's rust discount as if you were being clever. Price Zach Charbonnet at RB18 on the games he will miss, not on what he will do in the games he plays. Quinshon Judkins at RB11 has an ankle, which is the mildest repeat risk of the five injuries that matter. Where you should genuinely flinch is knees, and only knees: Michael Penix Jr. at QB43 is already marked down hard for it, and Patrick Mahomes at QB8 is not. James Conner at RB72 is the profile to stop worrying about entirely, a foot, which produced zero repeat absences in 42 player-seasons. Malik Nabers at WR50 is where our board and the market sit furthest apart, and that entry below is worth reading before you act on either.

What the consensus says

The injury-return genre got very confident this summer. Fantasy Points tells you quarterbacks back from surgery hold their passing but lose rushing efficiency "for about 12 months," and that Achilles cases keep a dip in explosiveness even once cleared. Yahoo prices a major ankle fracture at "about a 12-month dip in efficiency," peak performance filed under late October. Athlon and FantasyAlarm run the same shape at every position. Fantasy Points on hamstrings puts the highest reinjury risk in the first two weeks back.

The through-line: cleared is not back, expect less per game for months. It is a satisfying story because it sounds like the responsible, grown-up version of "he's healthy, draft him." Nobody I could find had run it against a control group.

The claim, in plain English

Take every week a player with a real role recorded zero offensive snaps while his team played, and attach the body part the official injury report listed for those weeks. That is an injury episode. Measure what he scored in his first game back, his second and third, his fourth through sixth, against his own per-game average before he got hurt.

Then, and this is the whole game, match him to healthy players at the same position, in the same season, putting up the same per-game number before he went down, and measure them over the exact same weeks. If returning players fade and their healthy twins do not, that is the injury. If both fade, that was never injury, that was a guy who was running hot and stopped. Nobody sells you the second version, so I built the control first. The instrument is scripts/validate_injury_prognosis_curves.py.

How I beat on it

1,377 attributed absences across 2018 to 2025, 944 of them ending in a return that same season, scored in half PPR with a tight end premium. 4,241 individual return games against a pool of 1,496 clean player-seasons. Everything public: official NFL injury and practice reports, public snap counts, our own weekly stat file. No competitor injury model touches this, by design. Framework in the research spec. Nothing after 2025 feeds it.

What the data actually said

First swing, no control, and the genre looks fine. First game back, a knee guy loses 1.09 points off his own average, a shoulder guy 1.55, a hamstring guy 0.68. Real dips, pointing exactly where you were promised they would.

Because aggregates lie, and because the guy who got hurt was, on average, in the middle of a good stretch when he got hurt. So I ran the healthy twins, expecting mean reversion to eat the whole thing. It did not. The twins barely moved: four hundredths of a point at hamstring and at shoulder, fifteen hundredths at knee. Ankle was the one real exception, half a point of drift, which is most of why its raw 0.89 dip shrinks to 0.40 once you account for it. So the fade is not a reversion artifact. It is just small.

Four of the five biggest injury types cannot be told apart from zero. Shoulder is the only one that clears, and it barely does, and it does not behave like a recovery curve when you follow it out: 1.59 down in game one, then 1.10, then 0.96, then back to 1.98 at seven-plus games on twenty players. Curves do not do that. Unstable estimates do. The snap counts undercut the story further: at knee, ankle, hamstring and shoulder a returning player is down seven to nine points of snap share in game one and back within a few points by game four. He is not diminished. He is on a pitch count, briefly.

Then the number that settles it. I gave a model perfect knowledge of the hamstring curve, the real one, fitted on the answers, and asked it to predict return games. It beat the naive "he is just himself" assumption by 0.115 points a game, range 0.04 to 0.19. Statistically real. Practically nothing, because a single fantasy game misses by 5.08 points on average no matter what you know. A tenth of a point of signal inside five points of noise is not a projection input. It is a rounding error with a press agent.

So I went looking for where the injury actually costs you, and it was one column over the whole time.

Same season, after he has already come back, knees repeat 18.6 percent of the time and shoulders 17.2. The takesmiths are right that injuries have types. They just put the type on the wrong axis. A knee does not make him worse. It makes him likelier to be gone again.

What the engine already figured out

Our board discounts availability and nothing else, which turns out to be the right shape more by luck than by design. Every durability signal we compute is a function of games missed. The flag actually named repeat soft tissue fires off a range of games played, not off the word hamstring. Body part reaches the player card and has never once reached a rank.

So the thing this study calls worthless, a per-game haircut for returning players, is a thing we never built. Nabers is marked down on availability alone, 11.1 expected games, and his per-game rate is untouched. That was right, and now it is right on purpose.

Be precise about what that 11.1 is, though, because it reads like a measurement of Nabers and it is not one. It is 17 games times 0.65, and 0.65 is our floor for proven recovery from a single acute event, the lowest availability we will assign anyone in that class. Rashee Rice, Jayden Reed and James Conner all price to the same 11.1. The floor is doing the work there, not a read on any particular knee.

The gap is the other column. We treat a knee and a foot identically once games missed match, and the repeat rates say 4.6x against zero-in-42. That is a measured, class-shaped miss on the availability side, which is where we already do the work. It is also a change to live ranking math, so it earns its own pre-registered test before it moves a single player. Not today, and not in a blog post.

The 2026 names

Real board ranks, pulled fresh off our 10-team superflex dynasty board, engine v1.6.134, generated 2026-07-31.

Malik Nabers, WR50, age 23.1, knee. Hold, and understand why. This is the disagreement, and the most interesting name here. Dynasty superflex ADP has him WR5. We have him WR50. That is a 45-rank gap, and almost all of it is the recovery floor described above rather than a judgment about his knee. So read this one carefully: there is no version of this where you buy him at WR50, because nobody is selling him there. If you think a blanket floor is too blunt for a 23-year-old a year removed, the market agrees with you and our board does not. Our own arbitrage column flags him RECOVERY_WATCH and declines to call him a value.

Zach Charbonnet, RB18, age 25.6, knee. Hold, comfortably. He clears our durability screen at 16.2 expected games despite the knee history, and his per-game output is not owed anything back.

Quinshon Judkins, RB11, age 22.8, ankle. Buy or hold. Ankle is the mildest of the five that matter, 1.7 times the base repeat rate.

Michael Penix Jr., QB43, age 26.3, knee. Fade, and the board already agrees hard at 2.5 expected games. Worst repeat class, on a player whose availability is the entire question.

Kendre Miller, RB53, age 24.2, knee. Buy, in small size. Same knee class, priced at 6.8 expected games, and the one name where our board is more aggressive than the market: ADP has him RB78, we have him RB53, and our arbitrage column calls that a strong buy. This study is the argument for taking it small rather than not at all. Do not talk yourself into it because "he looked fine when he played." He did look fine. That is the finding, and it is not the risk.

Patrick Mahomes, QB8, age 31.0, knee. Hold at QB8, but this is where the finding bites hardest: his knee is not priced for repeat risk. We have him at 14.9 expected games on an average durability grade. He is the one name here where our board and this study disagree, and the study is the newer evidence.

De'Von Achane, RB3, age 24.9, shoulder. Hold at RB3, and be a little patient in week one. Shoulder is the one class whose first-game-back dip survives the control, about a point and a half, though the wobble in its later windows says treat that as a soft warning and not a projection. Our 15.9 expected games carries the rest.

James Conner, RB72, age 31.3, foot. Sell, but not for the reason you think. This is the stop-worrying case, narrowly. Feet produced zero repeat absences in 42 player-seasons, so the foot is not the thing. His 11.1 expected games are that same recovery floor, not a foot-specific read. But be clear on direction: at 31.3 our board already sits below the market on him and the arbitrage column reads sell. The finding is that the foot should not be what scares you off. Age still should.

George Kittle, TE21, age 32.9, Achilles. Hold. Achilles is the class this study cannot speak to, too few cases to measure, so nothing here overrides the age-driven read that already has him TE21. The honest answer is that we do not know, and I am not going to invent a number for him.

What to do about it

No engine change. Nothing on the production side clears the bar, and the recurrence finding is a live-ranking change that owes a real harness before it earns a knob.

What changes is what you pay for. Two moves, both executable this week. One: stop applying a mental haircut to a returning player's per-game projection, and start taking the other side when your league-mate applies his. Two: when you do discount an injury, discount the calendar, and let the body part set how hard. Knees, shoulders and concussions repeat at three to five times the base rate. Ankles barely. Feet, on the sample I have, not at all.

One honest caveat: the public injury report gives a body part and never a diagnosis. There is no ACL against MCL in it, no high ankle against low. Everything above is body part plus how long he was out, which is the most any of us can honestly claim from public data, and less than the confident twelve-month timelines you have been reading.

Receipts

• The instrument: scripts/validate_injury_prognosis_curves.py (run --selftest for the hermetic check)
• Measured curve set, every class and window: docs/RESULT_G223_injury_prognosis_curves_2026-07-31.md
• Hypothesis record: data/hypothesis_backlog.json (H-81)
• Sources: official NFL injury and practice reports, public snap counts, weekly stats. No competitor injury model, ever.
• Board snapshot: 10-team superflex dynasty, half PPR with a tight end premium, generated locally 2026-07-31

← Back to all posts