injury return pathway
how the model prices players returning from IR
the standard projection model has a blind spot: players who were on IR for most of 2025 have no recent production signal. a model that reads recent weekly stats and finds nothing will either produce a garbage projection from stale pre-injury data or skip the player entirely.
both outcomes are wrong. we built the injury-return pathway to fix them.
the problem
a player missing 12+ games in 2025 looks like this to the base model:
- small sample: 4 games, maybe 3
- high variance: early-season production is noisy even for healthy players
- confidence score: 0.25-0.40
the base model would produce a low-confidence, low-stability projection and rank the player 30 spots lower than their actual expected value. this is the systematic error the pathway corrects.
what the pathway does
for each affected player, the pathway:
- classifies the injury by type and duration (games missed on IR)
- applies a confidence tier based on historical recovery certainty for that injury class
- applies a recovery factor that scales the pre-injury Elo signal up or down based on median return-game performance
- blends the recovery projection with whatever recent signal exists, weighted by confidence
the output is a projected FP/game and confidence score that honestly reflects what we know and what we don't.
confidence tiers
| injury class | confidence | notes |
|---|---|---|
| soft tissue (hamstring, quad, calf) | 0.40 | most cases recover to full production within 3-4 games |
| knee (non-ACL) | 0.35 | return rates good; production dip averages 15% first 4 games |
| shoulder / upper body | 0.35 | minimal long-term production impact in most cases |
| ACL | 0.25 | 14-case sample. production at 6-month mark averages 78% of pre-injury level |
| achilles | 0.20 | 2-case sample. recovery trajectory highly variable; one outlier inflates the RF |
confidence is the primary signal telling you how much to trust the number. a 0.20 confidence score is the model saying "we have 2 data points and one of them was Cam Akers." treat it as a wide range.
recovery factors
the recovery factor (RF) is a multiplier on the pre-injury Elo signal. it is derived from the median ratio of return-game FP/game to pre-injury FP/game across all historical cases in each class.
examples from the calibration:
- soft tissue: RF = 0.92. most players return slightly below pre-injury production in weeks 1-4, then normalize.
- ACL: RF = 0.75. significant production discount persists for half a season on average.
- achilles: RF = 0.75 with high variance. the confidence score (0.20) reflects the small sample more than the RF does.
blending
the final projected FP/game is a weighted blend:
final_proj = (confidence × recovery_proj) + ((1 - confidence) × position_median)
where position_median is the median projected FP/game for healthy players at the same position. at low confidence, the model leans on the position median as a safe anchor. at high confidence, the recovery projection dominates.
this is conservative by design. overconfidence on injury return projections has a known failure mode: the first return game goes badly (re-aggravation, load management, rust), the projection was way off, and you've been rostering a liability.
the 58 players
the 2026 version of the pathway covers 58 players. the breakdown:
- 15 ACL (estimated return: early 2026 season)
- 12 achilles (estimated return: week 3-8 range)
- 18 soft tissue (returning from IR, pre-season healthy)
- 13 other (shoulder, knee, misc)
every player in this cohort is flagged with data_source = "injury_return" in the API response. you can filter on this field to see the full list.
what it does not do
- it does not predict whether a player will be re-injured. nobody does that reliably.
- it does not account for load management decisions (coaches reducing snaps on returning players). these depress real return-game performance but are unknowable in advance.
- it does not have large enough samples for all injury classes to be considered final. the achilles class in particular (2 historical cases) should be treated as directional, not precise.
the pathway is honest about its uncertainty. that's the point: a 0.20 confidence score means "this is a range with high variance," not "skip this player."
frequently asked questions
can I filter the API to see only injury-return players?
yes. query /v1/rankings/preseason and filter client-side on data_source == "injury_return". or request a specific player by ID.
why is my favorite player's confidence so low?
low confidence on a returning player reflects the injury class sample size, not a judgment about the player's ability. Davante Adams at 0.35 confidence is not a commentary on Adams. it's the model saying "we have 12 historical knee cases and they came back at roughly 85% of pre-injury production on average."
how is this different from just projecting the pre-injury number?
projecting the raw pre-injury Elo signal without adjustment would systematically overstate return-game production. the historical data shows that even soft-tissue returns average about 8% below pre-injury production in the first 4 games. the pathway applies that haircut.