Chuck TackletonKnowledge Base
What Chuck currently believes about football and fantasy football, with his confidence in each claim. Claims are about the game, never about people in the league. Click a claim for the full picture: what would falsify it, the checks behind the number, and its sources.
37 claims, 1 contested.
Draft (5)
| ADP versus value (using market price to time picks) | 60% | inferred | |
| Positional scarcity (where the value cliffs are) | 60% | inferred | |
| Replacement theory in drafting (VORP and opportunity cost) | 60% | inferred | |
| Roster building via drafts | 60% | inferred | |
| Snake draft strategy by slot (Hero RB, Zero RB, Robust RB, late QB/TE, tiers) | 55% | inferred | contested |
In-Season (9)
| Processing injury news and practice reports | 62% | measured | |
| Bye weeks and fantasy playoff schedule planning | 60% | inferred | |
| Weekly lineup optimization | 60% | inferred | |
| No-prior PPG shrink (what a rookie or a player without a qualifying prior season shrinks his season-to-date points toward, and how many pseudo-games the draft-capital prior is worth) | 60% | measured | |
| Window length of the in-season PPG shrink (the whole season to date beats the last four games; recency weighting, a week term and a points-scale fit add nothing) | 60% | measured | |
| PPG-level in-season shrink (how much of a trailing-4-game points ratio carries into the next game, and whether it beats the raw average and the share blend for projecting points) | 60% | measured | |
| How many games of usage to trust (the in-season window) | 60% | measured | |
| In-season efficiency as a points trigger (a weeks 1-6 YPC or yards-per-target fall adds nothing to the trailing points average; measured null) | 55% | measured | |
| Pseudo-game blend of the in-season PPG shrink (the prior season is worth about 4-5 games, but blending by games played adds nothing to the fixed season-to-date line; measured null) | 55% | measured |
Outcomes (8)
| Rookie-year finish as the year-2 and year-3 breakout signal (WR and RB, with target share, TE and QB) | 68% | measured | |
| Draft capital as a breakout prior (early-career top-24 rates by round for WR and RB, top-12 for TE and QB) | 65% | measured | |
| Player range of outcomes (distributions, injury base rates, age curves) | 61% | measured | |
| Scoring distribution shape (weekly spread around expectation, next-season PPG spread, by position and tier) | 60% | measured | |
| Team range of outcomes (season simulation, playoff and title odds) | 60% | inferred | |
| Signals inside a draft-round tier (pick slot within round 1 and age at NFL entry add nothing to round; measured null) | 50% | measured | |
| Soft breakout signals (camp reports, preseason usage, college breakout age, age at breakout) | 45% | inferred | |
| Veteran decline signals (age by position, early-season snap-share drift, efficiency proxies) | 25% | measured |
Sources (2)
| Schools of thought in fantasy football analysis | 60% | inferred | |
| Projection and ranking source accuracy (what public projections explain, how much source accuracy persists, and how to weight sources) | 55% | inferred |
Theory (10)
| Usage stickiness (how well shares of targets, carries and snaps persist, by position and horizon) | 90% | measured | |
| Situational factors (wind, cold, dome, rest, Vegas implied total and spread) | 68% | measured | |
| Touchdown regression (how much of TD rate is noise, expected TDs from opportunity location, QB included) | 65% | measured | |
| Ball-share shifts (what moves when a teammate is absent or the QB changes, how fast, how long it sticks) | 64% | measured | |
| Team-change discount (how much prior-season usage share survives a move, and why a QB change alone is not a discount) | 62% | measured | |
| Correlation and stacking between roster players | 60% | inferred | |
| Football probability base rates (variance, TD regression, sticky stats, game environment) | 60% | inferred | |
| Quarterback interception regression (season INT rate is noise; volume and depth mix are what carry) | 60% | inferred | |
| QB quality and receiver points (how much a passer's quality moves his receivers' points, and why a new passer's prior season predicts none of it) | 50% | measured | |
| Coaching and play-caller change (how much of last season's share survives a new offense) | 45% | inferred |
Trades (1)
| What makes a trade improve championship odds | 60% | inferred |
Waivers (2)
| Streaming DST, K, and TE | 64% | measured | |
| Waiver wire strategy (FAAB, priority, speculative adds) | 60% | measured |
Claim
ADP is a price, not a value. Value comes from projections and replacement theory; ADP tells us when the room will take a player. The two together let us wait on players the room underprices and reach only when the room will otherwise take our target before our next pick. Platform-specific ADP (ESPN for an ESPN league) matters more than cross-platform ADP because the room drafts off the platform's default ranks.
Evidence for
- ECR and ADP diverge systematically by platform; ESPN default rankings drive ESPN room behavior, so ESPN ADP is the relevant price in an ESPN draft (consensus-adp).
- A gap of 15+ picks between ECR and ADP is a repeatable "value window" signal (consensus-adp).
- Averaging many rankers reduces individual error (wisdom of crowds), so ECR is a strong value baseline even though it cannot identify sleepers by construction.
Evidence against
- ADP reflects the crowd's information too; a player going well below ECR sometimes reflects news the rankers have not processed.
- In a live 10-12 person room, individual drafter tendencies swamp ADP: one RB-hungry drafter changes the pick math.
How we use it
- Before the draft, build a board with three columns: our VORP rank, ECR, and ESPN ADP.
- Wait rule: if a target's ESPN ADP is more than (picks until our next pick + 4) later than the current pick, do not take him now; take the best VORP whose ADP is inside that window.
- Reach rule: reach at most 8 picks ahead of ADP, and only for a player who is the last of his tier.
- Value window flag: any player with ECR at least 15 picks better than ESPN ADP is flagged; take him one round before his ADP, not at ADP.
- During the draft, update the ADP column with the live room: if a position has gone 4+ picks faster than ADP in the last round, shift its ADP up one round for the rest of the draft.
- Log every pick's (VORP rank, ECR, ADP, pick number) so REFLECT can score whether we captured value.
Open questions
- Does this league's room historically draft off ESPN default ranks or off outside sources? Answer from draft history in
league/.
Claim
For a 10-12 team managed H2H league in 2026: anchor with at least one RB in the first two rounds (Hero RB or Robust RB), take WRs in volume in rounds 3-7, wait on QB until a rushing QB with top-5 upside remains (typically round 7+), and draft TE either in the elite tier (rounds 2-3) or as a late streamer, never in the middle. Draft by tiers, not ranks: within a tier, choose by need and by the drop-off to the next tier at our next pick.
Evidence for
- Best Ball Mania VI: teams with an RB in Round 1 advanced ~24% vs ~12% without; two RBs in Rounds 1-2 advanced ~30% vs ~8% for no RB in either (market-efficiency). 2026 analysts describe Hero RB as the prevailing approach.
- 2026 is a year with a dozen or so workhorse RBs and a weak rookie RB class, which most sources say makes Zero RB less attractive than in prior years.
- Late-round QB: ~13 confident starters and ~20 QBs with QB1 upside in 2026; the overall QB1 has run for 350+ yards and 4+ rushing TDs every year since 2019, so rushing upside is the selection filter (consensus-adp, contrarian-upside agree here).
- Tiers over ranks is near-universal across schools.
Evidence against
- Best-ball advance rates are for a 12-team, 18-man, no-waiver format with tournament payouts. H2H with waivers rewards floor and repairability more; the Round 1 RB edge may shrink.
- Zero RB proponents: RBs are the most injured position and the RB replacement pool refreshes weekly via waivers, so paying early-round capital for RBs is paying for fragility.
- Advance-rate structure findings shift year to year with the player pool; they are descriptive of last year, not a law.
How we use it
- Default structure: Hero RB (one RB in rounds 1-2) unless our slot is 1-3 and two elite RBs are on the board (then Robust RB). Zero RB only if the top two RB tiers are gone by our second pick.
- Rounds 3-7: WR-heavy. Take an RB in this window only if he is the last of a tier and the VORP gap to the next tier exceeds the WR gap.
- QB: do not draft before round 6. From round 7, take the highest-ranked QB with rushing upside (>=400 projected rushing yards) when he is the best VORP on the board; otherwise wait. Never take a pocket QB before round 10.
- TE: take an elite-tier TE if available at or below his ADP in rounds 2-4; otherwise wait until round 9+ and take two upside TEs late, or stream.
- Tiers: use the aggregate projection tiers; within a tier pick the player whose position has the larger drop-off before our next pick.
- Record which structure was executed and why in the decision record so REFLECT can score it.
Open questions
- Does this league's history show early-RB or early-WR champions? Resolve via
league/once draft history is pulled.
Resolved 2026-09-11 (Q-0001, Q-0002): this league is standard H2H points, PPR 1.0/rec, 10 teams, single QB and single TE (no superflex, no TE premium) — rules 3 and 4 stand as written. 2026 was an ESPN autodraft (owner, "in the interest of time"; all draft work dropped for this season), so these rules were not exercised; they remain the plan for the 2027 draft.
Claim
Scarcity is the steepness of the projected-points curve at a position between our current pick and our next pick. In 12-team 1QB PPR, the RB1-to-RB24 gap (~100+ points) is far larger than the QB1-to-QB12 gap (~60 points), so RB is scarce early and QB is not. Elite TE is scarce (a two- or three-player tier, then a cliff). WR is deep through round 7 in 2026. Scarcity is format-dependent: superflex makes QB scarce; TE-premium makes TE scarce.
Evidence for
- VORP tables from multiple projection sets show the same shape: RB cliff after ~RB12 and again after ~RB24; QB flat; TE a 2-3 player elite tier (analytics-projection).
- Best-ball structural data rewards early RB, consistent with RB scarcity being underpriced by the room (market-efficiency).
Evidence against
- The RB cliff is partly an artifact of injury: the RB curve is steep because many mid-round RBs miss games, not because healthy talent is scarce. Replacement RBs appear on waivers weekly.
- 2026 has more rushing QBs than ever, which flattens QB even further and moves the QB cliff later; the exact round shifts yearly.
How we use it
- Compute the projected gap at each position between the best available now and the best expected available at our next pick (using ADP). This is the cliff score for that position at this pick.
- Among candidates within one VORP tier, take the position with the highest cliff score.
- Elite TE rule: if an elite-tier TE is available and the cliff score for TE exceeds every other position's, take him; otherwise never draft a TE in rounds 4-8.
- QB rule: in 1QB leagues the QB cliff score is near zero until round 7+; do not let it drive a pick before then. In superflex, treat QB like RB.
- Recompute cliff scores live every pick; ADP drift in the room changes them.
Open questions
- This league's scoring (PPR? TE premium? 6-pt passing TD?) is unknown until mSettings is read. Confidence drops to 0.5 if superflex.
Claim
A player's draft value is his projected points minus the points of the best player at his position we could get for free (replacement level). Replacement level is a function of league size and roster slots, not a constant. The choice at any pick is an opportunity cost: player A now versus the best expected A-position player at our next pick, compared with the same for player B. A QB projected for 300 points over a 285-point replacement is worth less than an RB projected for 200 over a 100-point replacement.
Evidence for
- VBD has been the dominant analytical drafting framework since the 1990s and consistently beats raw-points drafting in simulation (analytics-projection).
- Multiple baseline definitions exist (VOLS: last starter; VORP: best waiver player; VONA: value over next available; man-games) and all outperform ignoring replacement level; they differ mainly in how they treat bench spots.
Evidence against
- VBD's output is only as good as the projections fed in; garbage projections produce confident garbage.
- Pure VORP ignores variance: two players with equal VORP can have very different ranges of outcomes (see outcomes/player-range-of-outcomes.md).
- Late in drafts, VORP compresses toward zero for every position and stops discriminating; upside and roster-fit take over.
How we use it
- Replacement level per position = projected points of the (starters + reasonable bench) rank at that position, using THIS league's settings:
- QB: rank (teams x QB slots) + 2 (e.g. 12-team 1QB: QB14)
- RB: rank (teams x (RB slots + FLEX x 0.5)) + 6 (12-team 2RB+1FLEX: about RB36)
- WR: rank (teams x (WR slots + FLEX x 0.5)) + 6 (12-team 2WR+1FLEX: about WR36)
- TE: rank (teams x TE slots) + 2 (12-team: TE14)
- DST, K: rank teams + 1 (streamable; VORP effectively 0 until the last rounds) Superflex: QB uses (teams x 2) + 2.
- VORP(player) = projection - replacement(position). Rank the board by VORP.
- Opportunity cost at a pick: for each candidate, compute VONA = projection - (best projection at that position expected to survive to our next pick, using ADP). Prefer the candidate with the largest VONA when VORPs are within one tier.
- Use the aggregate projection (multiple sources, multiple schools) as input; never a single source (PG-3).
- From round 10 on, when VORPs compress below ~15 points, switch to upside ranking (outcomes/player-range-of-outcomes.md rule 3).
Open questions
- Which projection set should be the primary input on draft day? Interim: FantasyPros aggregate, cross-checked against one usage-based source.
Resolved 2026-09-11 (Q-0001): this league is 10 teams, standard 1QB/1TE (no superflex), starters {QB 1, RB 2, WR 2, TE 1, FLEX 1, D/ST 1, K 1}. Plugging into rule 1's formulas: QB rank 12, RB rank 31, WR rank 31, TE rank 12 (DST/K stay streamable, VORP ~0 until the last rounds). Not superflex, so the QB line uses the single-QB formula throughout. 2026 was an ESPN autodraft (Q-0002); these baselines were not exercised this season and stand ready for 2027.
Claim
A draft builds a portfolio, not a list of best players. The roster that wins a managed H2H league has (a) starters chosen by value over replacement, (b) bench spots allocated to the positions with the highest weekly variance and injury rate (RB, then WR), (c) one starter at the onesie positions (QB, TE) unless an elite tier is available at a discount, and (d) zero draft capital on DST and K before the last two rounds.
Evidence for
- Value-based drafting (VBD) beats drafting by raw projected points because replacement level differs by position (analytics-projection).
- Best-ball data: 7-8 WRs and 5-6 RBs were the modal winning builds; positions with higher weekly variance deserve more roster slots (market-efficiency). Managed leagues have smaller benches, so the ratio, not the counts, transfers.
- Streaming DST/K matches or beats drafting the top-ranked unit (rules-of-thumb; see waivers/streaming.md).
Evidence against
- Best-ball builds optimize for weekly ceiling with no waivers; managed leagues can repair a roster via waivers, so bench allocation matters less than in best ball.
- Small-league (8-10 team) replacement levels are high enough that positional balance matters less than raw talent.
How we use it
Draft-day rules for the Builder, in order: 1. Compute replacement level per position from THIS league's roster settings (see replacement-theory.md). Rank the board by VORP, not projected points. 2. Fill starting slots by VORP tiers. Break ties toward the position with the steeper VORP drop between this pick and our next pick (see positional-scarcity.md). 3. Bench allocation targets for a 15-16 man roster: RB 4-6, WR 4-6, QB 1 (2 only if a top-3 rushing QB falls), TE 1 (2 only if elite TE plus late upside), DST 1, K 1. Deviate only when VORP gap exceeds one full tier. 4. DST and K in the final two rounds. Never earlier unless every remaining skill player is below replacement. 5. Do not draft for bye weeks. Use bye only as a last tiebreaker (see in-season/bye-and-playoff-planning.md). 6. Log the allocation actually drafted against these targets in the decision record.
Open questions
- (none open)
Resolved 2026-09-11 (Q-0001): 10 teams, standard 1QB (no superflex), starters {QB 1, RB 2, WR 2,
TE 1, FLEX 1, D/ST 1, K 1} = 7 starting slots on a 16-man roster (9 bench), confirmed from
mSettings. QB allocation rule (one QB unless a top-3 rushing QB falls) stands. 2026 was an
ESPN autodraft (Q-0002), so bench allocation targets (rule 3) were not exercised this season;
they remain the plan for 2027.
Claim
Draft talent, not bye weeks. Byes are solved on waivers in-season, and concentrating several starters' byes in one week costs one bad week rather than several weakened weeks. The fantasy playoff weeks (usually 15-17) are the money weeks: from mid-season on, matchups and weather in those weeks are a legitimate tiebreaker for trades and adds, but never the primary criterion.
Evidence for
- Consensus across schools: best player available, byes as a last tiebreaker; consolidating byes ("bye stacking") is defensible.
- Playoff strength-of-schedule and dome/warm-weather scheduling for weeks 15-17 are widely published and matter because single-elimination amplifies one week's variance (market-efficiency, rules-of-thumb).
Evidence against
- Preseason strength-of-schedule predictions of week 15-17 defenses are weak; defenses change a lot by December.
- Bye stacking is a small effect; in a 14-week regular season one forfeited week can cost a playoff spot in a tight race.
How we use it
- Draft: byes are ignored except as the final tiebreaker inside a tier. Do not avoid two studs sharing a bye.
- In-season, two weeks before a bye pileup (3+ starters out), the waiver rule shifts: prioritize a one-week streamer at the affected slot over a speculative stash (waivers/waiver-wire-strategy.md rule 4).
- If the pileup week is against a much stronger opponent (win probability < 30%), punt it: do not spend FAAB above streamer tier to patch it.
- From week 9 on, add a playoff-weeks projection column (weeks 15-17 matchups, dome/weather) to every trade and add evaluation; use it as tiebreaker within 5% of ROS projection.
- By week 13, prefer DST/K with favorable multi-week playoff matchups (waivers/streaming.md rule 6).
- Regular-season length, playoff weeks, and bye schedule are read from mSettings and the NFL schedule; never hard-coded.
Open questions
- Does this league's playoff run weeks 15-17 or 14-16, and is the final one week or two? Determines which weeks rule 4 targets.
Claim
A starter whose per-touch efficiency falls in weeks 1-6 (RB yards per carry down 0.5+, WR or TE yards per target down 1.0+ versus last season) while his snap share holds is not a rest-of-season points trigger beyond what his weeks 1-6 points already say. The raw signal looks real: P(weeks 7+ PPG under 0.8 x prior) is 0.42 for the fallen group against 0.24 for the held group (n 160 / 278, 2019-2025, half-PPR, our data). But the fallen group's weeks 1-6 PPG ratio is already 0.88 against 1.20, and once that ratio is held fixed the efficiency flag adds +0.02 to the weeks 7+ ratio (CI -0.05 to +0.09, shuffle floor 0.07, n 438): a measured null, at every position (RB -0.02, WR +0.07, TE -0.01). Inside early-ratio bins the fallen-vs-held gap is 0.02-0.07, inside noise. The early efficiency fall is about half noise and half real: the fallen group's weeks 7+ efficiency sits 0.98 yards below last season (early change -2.09), the held group's sits at last season's level, but that persistent half is already in the points ratio, so it is not a second signal. Dropout (fewer than 3 late games) is 0.05 either way. Separately, the weeks 1-6 PPG ratio itself carries into weeks 7+ at slope 0.53 (late = 0.44 + 0.53 x early), which is the in-season points version of regression to the mean.
Evidence for
- Our check
in-season-efficiency-points-fork.py(2026-09-11): coef +0.016, CI -0.052 to +0.088, floor 0.070, n 438; raw P(bad) 0.419 vs 0.237; early ratio 0.881 vs 1.204; late ratio 0.920 vs 1.078; bins early <0.8 0.60 vs 0.54 (n 73/28), 0.8-1.0 0.34 vs 0.28 (41/67), >=1.0 0.20 vs 0.18 (46/183); late efficiency vs prior -0.98 vs +0.04; gone rate 0.048 vs 0.041; RB coef -0.024, WR +0.068, TE -0.012. - The A-46 check (
outcomes-veteran-decline-inseason-efficiency.py, in outcomes/veteran-decline) found the same raw PPG-bad gap (0.27 -> 0.45) with role unmoved; this check explains it. - Harstad (Footballguys, rules-of-thumb): within one season a back at 5.00 YPC over 8 games projects to 4.37 in the other 8 and one at 3.50 projects to 3.93; YPC needs about 1,978 carries to be half skill (Tuccitto), so an in-season YPC swing is mostly sample noise. Same column on yards per target: three weeks of YPT is dominated by yards per reception, and the higher-target, lower-YPT group was the one to bet on.
- PFF (analytics-projection): year-to-year YPC correlation about 0.41 over five seasons; the projection is 0.59 x league average + 0.41 x last season.
- Sharp Football and 4for4 (search summaries, not fetched this pass): yards per target and other per-opportunity efficiency stats are the least stable receiver stats; targets and points per game are what carry.
- The season-level version agrees: theory/usage-stickiness and outcomes/veteran-decline measure YPC and YPT year over year at r 0.27-0.39 and find no next-season points effect from an efficiency fall.
Evidence against
- The controlled coefficient is a linear adjustment on a single covariate; a nonlinear interaction (efficiency fall only mattering for players whose early ratio is under 0.8) shows as 0.60 vs 0.54 in that bin, n 73 vs 28, which is too thin to exclude a +0.05 to +0.10 effect there.
- Half of the early efficiency fall persists into weeks 7+ (-0.98 of -2.09), so efficiency is not pure noise inside a season; it is redundant with the points ratio, not absent.
- The population is snap-share-held starters with 3+ late games; a fall that coincides with a snap drift is handled by veteran-decline rule 2, and role loss (not points) is where A-46 found a small RB/WR lead.
- Sample is 2019-2025, weeks 1-6 vs 7+ only; a trailing-4 window at other cut points is untested.
How we use it
- No efficiency trigger in-season. Do not lower (or raise) a starter's rest-of-season PPG because his YPC or yards per target moved in weeks 1-6 with snap share held; the trailing points average already contains it. Efficiency stats stay out of the lineup optimizer and the waiver ranker as inputs (theory/usage-stickiness rule 1).
- Shrink the in-season points ratio. When projecting weeks 7+ PPG from a weeks 1-6 PPG ratio for
a snap-held starter, use 0.44 + 0.53 x (early ratio) times prior PPG, not the raw early ratio: a
0.70 early ratio projects 0.81, a 1.30 projects 1.13. Superseded 2026-09-12 by
in-season/ppg-shrinkrule 1, the same line measured at every week and by position (pooled 0.46 + 0.49 x, RB 0.58, WR 0.49, QB 0.41, TE 0.30); this one-cut version sits within 0.02 points of MAE of it. Share projections keep the usage-window blend. - Role, not efficiency. The in-season decline signal remains snap-share drift (veteran-decline rule 2). A weeks 1-6 efficiency fall with snap share held is at most a weak role-loss lead for RB and WR (A-46) and gets no points weight.
- Re-run the check each August after the data refresh.
Open questions
- Answered 2026-09-12 (A-73,
in-season/ppg-shrink): the trailing-4 slope is 0.49 pooled at every week (0.41 weeks 5-8 to 0.53 weeks 13-18) and the shrink beats the share blend for projecting points by 4.8% of MAE. Window length and TE/QB weakness stay open there (A-75, A-76). - The early-ratio-under-0.8 bin (0.60 vs 0.54) needs 2026-2027 rows before a nonlinear fork can be excluded.
Claim
The injury designation is the headline; the practice report is the evidence, and the player's role matters as much as either. Measured on nflverse injury reports joined to nflverse weekly usage, 2019-2025 regular seasons, QB/RB/WR/TE (check above, run 2026-09-11):
| Final report | Last practice day | Starters active | n | All rostered active |
|---|---|---|---|---|
| Questionable | Full | 0.87 (RB 0.92, WR 0.92, QB 0.84, TE 0.77) | 174 | 0.66 |
| Questionable | Limited | 0.75 (RB 0.84, WR 0.78, TE 0.75, QB 0.52) | 897 | 0.60 |
| Questionable | DNP | 0.50 (WR 0.54, TE 0.49, RB 0.48, QB 0.29) | 237 | 0.38 |
| Doubtful | any | 0.01-0.04 | 137 | |
| none (on report, no game status) | DNP | 0.85 | 341 | |
| none (on report, no game status) | Limited | 0.95 | 591 |
Starter = snap share >= 0.40 in his last game. The pooled Questionable-starter rate is 0.72, so
practice status moves the needle in both directions. The headline "Questionable plays ~70%"
(Footballguys, 2017-2023, 2,000+ cases) is right for the whole report but wrong for any
particular player: a starter after a full practice is a 0.87-0.92 play, a depth player after a DNP
is under 0.40. Doubtful is effectively Out (under 4%). After a 2+ game absence, players score
well below their pre-absence baseline, and most of the deficit persists past the first game
(measured in outcomes/player-range-of-outcomes.md, check outcomes-injury-base-rates).
What the post-absence deficit is (checks in-season-post-absence-role-fork and
in-season-post-absence-injury-type, run 2026-09-11; RB/WR/TE regulars, 2+ game in-season
absence, same team on return, half-PPR ratio to the pre-absence 4-game baseline):
| Slice | Game 1 pts | Game 1 snap share | Game 1 pts/snap | Games 3-6 pts | Games 3-6 snap | Games 3-6 pts/snap | n |
|---|---|---|---|---|---|---|---|
| RB | 0.73 | 0.77 | 0.94 | 0.87 | 0.86 | 1.01 | 36 |
| WR | 0.76 | 0.81 | 0.95 | 0.79 | 0.90 | 0.94 | 75 |
| TE | 0.59 | 0.76 | 0.78 | 0.78 | 0.93 | 0.85 | 28 |
| QB | 0.78 | 0.69 | 1.03 | 1.04 | 0.92 | 1.12 | 65 |
| Games 3-6 snap share >= 0.90 x baseline (54% of skill cases) | 0.96 | 1.07 | 0.93 | 58 | |||
| Games 3-6 snap share < 0.90 | 0.62 | 0.69 | 0.96 | 49 | |||
| no-absence floor (regular games, week 6+) | 1.02 | 0.93 | 698 |
Three findings. (1) Game 1 is a snap ramp, not rust: RB/WR points per snap are 0.94-0.95 of baseline while snap share is 0.77-0.81; only TE loses per-snap efficiency (0.78). (2) The games 3-6 deficit is a role fork, not a decay: the half whose snap share is back score at the floor (0.96 vs 0.93), the other half score 0.62 with per-snap efficiency intact (0.96). The gap (0.34, CI 0.16-0.52) clears the shuffle floor (0.16). Game-1 snap share predicts nothing about the fork (games 3-6 pts 0.79 vs 0.82), but game-2 snap share does: >= 0.90 gives games 3-6 pts 0.88 and P(role restored) 0.75; < 0.90 gives 0.72 and 0.30. (3) Injury body part (the only type field public reports carry) is a measured null: soft-tissue 0.85 vs joint/bone 0.87 at games 3-6 (n 30 / 39, shuffle floor 0.22). Absence length is not monotonic: 2-game absences are the worst at games 3-6 (0.73, n 51), 3-4 games the best (0.92, n 40), 5+ 0.80 (n 16), so the length itself carries nothing usable. Absences never listed on the injury report (benching, personal, suspension, IR without a report row; n 14) score 0.43 in game 1 and 0.61 in games 3-6 (CI 0.42-0.79) against 0.84 for listed absences; upper-body injuries sit at 0.68 (n 15). QBs recover fully by game 3 whatever the cause.
Evidence for
- Measured (film-and-usage, nflverse): the table above. Full > Limited > DNP at every position; the full-practice rate beats the pooled rate at every position; Doubtful plays 1-4%.
- Footballguys injury index (rules-of-thumb): Questionable 71% played, Doubtful 5.9% played, 2017-2023; notes team-level tag habits (Tampa Bay and Chicago liberal, Philadelphia conservative).
- Betting-markets outlets treat the Questionable tag as spanning 25-100% and rely on practice participation and beat reporting rather than the tag, consistent with the spread above.
- Multiple outlets independently describe the same trajectory patterns and the ~90-minute inactive-list resolution (rules-of-thumb).
- Week 1 2026 returns from season-ending injuries (Malik Nabers, Patrick Mahomes, Alec Pierce) were cleared with full or near-full practice and "expected to play" framing, yet each carried an explicit workload caveat, so a full-participation designation for an offseason surgical return is availability evidence, not workload evidence (rules-of-thumb, analytics-projection, film-and-usage).
- Post-absence decomposition (film-and-usage, nflverse 2019-2025): the table above. The role fork, the game-2 snap-share signal and the body-part null each clear their shuffle floors or sit inside them as claimed.
- 4for4 Injury Index (analytics-projection, 2017-2021, 1,048 players): production -18.6% while carrying a designation (QB -8%, RB -16%, WR -22%, TE -18%); TE hamstring -38% in the first week but +8% by the three-game mark, RB high-ankle sprain -21% for the rest of the season, WR groin -36% when games were missed. Direction matches ours (first game worst, recovery uneven), but its per-body-part cells are 5-30 cases and it does not separate snaps from efficiency.
- Peer-reviewed fantasy-points return-to-play study (film-and-usage, 2017-2022, 2,523 time-loss injuries): -0.50 PPG next season overall, QB -1.95, RB -0.70, WR -0.33; no correlation between games missed and the deficit (R-squared 0.005), which agrees with our non-monotonic length split; injuries after the bye cost more (-1.22 PPG) than before it.
- NFL hamstring recurrence study (film-and-usage, 2009-2020, 2,075 injuries): a third recur, most within two weeks of return, and return within 2 weeks is the largest recurrence risk factor; WR is the highest-risk position. This is the mechanism behind our 2-game absences scoring worst at games 3-6 (a quick return is the rushed one), though we cannot yet separate recurrence from role loss (A-53).
- Week 1 2026, three camp-injury cases with no regular-season games missed yet (our roster/FA pool, 2026-09-11): Brock Bowers (meniscus trim, diagnosed late in camp, DNP all week, team language "a game or two," graded OUT) never reached a practice trajectory at all -- the procedure plus explicit team-stated absence length was the actionable signal days before the official tag existed. Jeremiyah Love (preseason high-ankle sprain, Aug 13) ran Limited-Limited-Limited into Sunday with the depth chart listing him behind the presumptive starter, while the coach repeated "feels good about his prospects" every media day; the player's own "50/70/100%, I'll play what they ask" quote is a stronger signal for availability than for full workload. Alvin Kamara (camp Grade 2 MCL sprain) went from a hard multi-week absence estimate to Limited-Limited by Thursday, with the coach explicitly declining to guarantee him either way. In all three, official-report participation (film-and-usage) tracked the medical severity better than the coach's public framing (rules-of-thumb), which stayed optimistic through the DNP/Limited days in every case.
Evidence against
- nflverse records only the final listed practice day, not the Wednesday-Thursday-Friday trajectory, so DNP-DNP-DNP versus DNP-DNP-LP cannot be separated here (WO-0013). The 0.50 for "DNP" mixes both.
- Play-rate percentages for Questionable vary by team and coach; a league-wide rate misleads for specific teams. Not yet split by team (small cells).
- The all-rostered full-practice rate drifted down over the window (0.74 in 2020 to 0.53 in 2025) while the DNP rate drifted up; the starter split is what to trust, but re-verify each season.
- QB Questionable-Limited is only 0.52 (n=100): a limited quarterback is a real coin flip, unlike RB/WR.
- Reports lag: beat-writer information beats the official report by hours.
- The role fork conditions games 3-6 points on games 3-6 snap share, so part of the gap is arithmetic (points follow snaps); the decision-relevant version is the game-2 signal (0.88 vs 0.72, n 57 / 50), which is weaker. "Snap share not restored" mixes a lost role with a player limited because he is still hurt; the per-snap efficiency (0.96) says the second group plays normally when on the field, but the two need a practice-trajectory or beat-report split (WO-0013) to separate.
- Body-part cells are 30-39 for the null and 6-15 for the rest; the unlisted-absence figure (n 14) is a flag, not a multiplier to trust to two decimals. Per-position body-part cells are 4-20 and are not used.
- The 4for4 index says calf strains (RB -41% at three games) and high-ankle sprains are the lasting ones; our soft/joint grouping is too coarse to see a single body part, and public reports never carry the grade.
How we use it
- Pull the official practice report Wednesday, Thursday, Friday (and Saturday for Monday games). Store the trajectory per player in
state/private/(WO-0013 automates this; until then the final-day status from nflverse stands in). - Availability probability for a starter (snap share >= 0.40 in his last game): Questionable + Full: 0.90 (TE 0.80). Questionable + Limited: RB 0.85, WR 0.80, TE 0.75, QB 0.55. Questionable + DNP on the final day: 0.50 (QB 0.30); if the trajectory is known to be DNP-DNP-DNP, use 0.35 until WO-0013 measures it. Doubtful: 0.03. Out: 0. On the report without a game status: 0.90 after a DNP (routine rest), 0.95 after Limited. For a non-starter cut each Questionable number by 0.20. Adjust by team tendency once we have 6+ weeks of team-level data.
- Multiply the player's weekly projection by availability probability for lineup decisions; if the product falls below the best bench alternative, bench him unless he plays in an early window and the alternative plays later (then hold and re-check at inactives).
- Post-absence discount, measured against the player's own pre-absence 4-game baseline after a 2+ game in-season absence (
outcomes/player-range-of-outcomes.mdrule 7 has the table): first game back RB x0.75, WR x0.75, TE x0.65, QB x0.75; second game RB x0.80, WR x0.75, TE x0.75, QB x0.85. If the ESPN projection is already below the pre-absence baseline, apply only the remaining gap, not the full multiplier (A-01 will measure whether ESPN prices this in). 4b. From game 3 on, the multiplier is a fork, not a blend (measured 2026-09-11). Read the game-2 box score: snap share >= 0.90 of the pre-absence baseline -> x0.90 for games 3-6 (restored group 0.96, floor 0.93, but only 0.75 of this group stays restored); snap share < 0.90 -> x0.70 and re-baseline on the observed post-return usage, since only 0.30 of that group gets the role back. Ignore the game-1 snap share (it is a ramp and predicts nothing about games 3-6). If game 2 has not been played yet, use the pooled RB x0.85, WR x0.80, TE x0.80. QB: x1.0 from game 3 regardless. Points per snap are back to normal by game 3 at RB/WR, so do not stack a "rust" cut on top of the snap-share fork; TE per-snap efficiency stays at 0.85, so TE gets x0.85 even when restored. 4c. Do not adjust for injury body part or absence length (measured nulls: soft-tissue 0.85 vs joint/bone 0.87 at games 3-6; 2-game absences 0.73, 3-4 games 0.92, 5+ 0.80, not monotonic). Two flags only: an absence with no injury-report row (benching, personal, suspension, unexplained) gets x0.65 for games 1-6 until the game-2 snap share says otherwise, because it is a role signal, not a health signal (n 14, so weight this at half strength when it conflicts with beat reporting); an upper-body injury gets x0.85 through game 6 (n 15, same caveat). A 2-game soft-tissue absence at WR is the hamstring-recurrence case: hold the game-1 multiplier through game 2 and do not restore before the game-2 snap share is in. - New Wednesday DNP for a previously healthy starter: flag for
…(truncated)
Claim
Start the higher aggregate projection. Matchup, weather, and home/away are already inside good projections; using them again double-counts. Volume beats talent for weekly floor: a player with 8+ targets has a higher floor than a more talented player with 4-5. Deviate from projection only to manage variance relative to our win probability (favorites want floor, underdogs want ceiling) and for late-breaking injury news.
Evidence for
- Every outlet's start/sit tool reduces to comparing projections; matchup is a tiebreaker for bubble players, not a reason to bench a stud (rules-of-thumb, analytics-projection).
- Volume drives floor; targets are the most stable weekly input (film-and-usage; theory/football-probability.md).
- Vegas-derived win probability for our matchup is a good input for choosing floor vs ceiling (team-range-of-outcomes.md).
Evidence against
- Aggregate projections lag late news (Saturday/Sunday inactives); the harness must re-run after the final injury report.
- In shallow leagues, the difference between bench options is small enough that these rules rarely matter.
How we use it
- Lineup is set Thursday morning from the aggregate projection, re-set after Friday's injury report, and re-checked 90 minutes before each kickoff for inactives (in-season/injury-processing.md).
- Projection first: start the highest p50 at each slot. Override only when (a) our win probability > 65% and a lower-variance alternative is within 1.5 projected points, or (b) win probability < 35% and a higher-p90 alternative is within 1.5 points.
- Volume tiebreaker: when p50s are within 1 point, start the player with the higher projected touches/targets.
- Never bench a top-12 positional player for matchup alone.
- Thursday-night players: start them only if their p50 exceeds the best Sunday alternative by 1+ point (losing the option to react to Sunday news has a cost).
- Playoff weeks: switch to the p90 rule from team-range-of-outcomes.md rule 4.
- Each lineup is a decision record with the alternative considered; REFLECT scores realized vs projected.
- Kickoff-window contingency. A Questionable starter locks at his own kickoff, and his inactive
status arrives about 90 minutes before it; every bench player whose game kicks off earlier is
already locked by then. So the contingency for a Questionable starter must name a bench player
in the same or a later window, and the expected value of holding him is
availability x p50(if active) + (1 - availability) x p50(same-or-later-window alternative), compared against the best earlier-window alternative's p50 started outright. Hold when the contingency EV is at least equal (the FLEX-timing tilt favors the later kickoff among near-equal options); start the earlier alternative outright when it is not. If the starter is downgraded before the earlier window locks, the earlier alternative is back in play (injury-processing.md rule 1 re-check). Week 1 2026 case: Love (ARI, 4:25 ET) with Williams and Gainwell at 1:00 ET and Golden at 4:25 ET.
Open questions
- Does ESPN's own projection outperform the aggregate for ESPN-scored leagues (it is scoring-aware)? Test in-season and record in
sources/.
Claim
A player with no qualifying prior season (rookie, or anyone without 8+ games at 6+ half-PPR PPG, 12 QB, last year)
who has 4+ games this season and a season-to-date PPG at that floor projects his next game best by shrinking the
season-to-date PPG toward the mean PPG of his position x draft-round bucket, with the prior worth n0 = 6 pseudo-games:
z = (6 T + n x) / (6 + n), PPG_hat = 0.88 + 1.04 z. Fit leave-one-season-out on 5,337 player-games (2019-2025, 1,385
of them rookies), this cuts MAE 2.9% against the raw season-to-date mean (5.48 vs 5.65 points; every held-out season
gains, 1.2% to 3.8%), 1.0% against a per-position line on the raw mean (5.54) and 0.7% against the same blend with a
position-only target (5.52). n0 = 6 is identified (cluster bootstrap 4-8, never at the grid edges), and the line on z is
mean-unbiased (prediction / outcome 0.99) where the raw mean runs 6% high, because a player who has cleared the 6-point
floor was selected partly on luck. The targets T (mean PPG of no-prior player-seasons with 4+ games) are: RB r1 12.7,
r2 9.4, r3 6.4, r4-7 4.6, udfa 4.0; WR r1 7.2, r2 5.7, r3 4.6, r4-7 3.8, udfa 3.0; TE r1 5.4, r2 4.6, r3 4.1, r4-7 3.1,
udfa 2.3; QB r1 12.5, r2 12.5, r3 8.6, r4-7 7.9, udfa 6.2. The draft-capital gain is an early-season gain: 5.2% over raw
at 4-7 games, 2.1% at 8-11, and at 12+ games the plain per-position line ties the blend (5.60 vs 5.61). Trailing-4 PPG
is worse than the season-to-date mean here too (5.88), rookie-specific targets are thinner and worse (5.49 vs 5.48),
and for QB the per-position line (5.6 + 0.62 x) edges the blend (6.62 vs 6.62). The textbook through-origin blend posts
the lowest MAE (5.46) only by predicting 2.4% under the mean, the same bias in-season/ppg-shrink-pseudo-games found.
Evidence for
- Our check
in-season-ppg-shrink-no-prior.py(2026-09-12, leave-one-season-out, MAE in half-PPR points): raw 5.648, trailing-4 5.881, per-position line 5.538, position-target blend 5.523, draft-round blend 5.485, draft-round x rookie blend 5.493, through-origin draft-round blend 5.460; the training folds pick the draft-round blend in all seven seasons (n0 5-6); mean prediction / outcome raw 1.057, line 0.994, blend 0.992, through-origin 0.976. By position (raw / line / blend): RB 5.911 / 5.873 / 5.741, WR 5.415 / 5.253 / 5.245, TE 4.674 / 4.552 / 4.470, QB 6.716 / 6.616 / 6.624. By games played: 4-7 5.720 / 5.552 / 5.420, 8-11 5.578 / 5.478 / 5.463, 12+ 5.631 / 5.599 / 5.613. Rookies 5.666 / 5.575 / 5.562, non-rookie no-prior 5.642 / 5.526 / 5.457. Week-1 rookie prior (PPG per active game, rookies with 1+ game, mean/median/n): RB r1 13.6/13.3/7, r2 10.0/10.8/15, r3 5.7/4.9/21, r4-7 4.0/3.0/89; WR r1 9.1/9.2/32, r2 6.4/6.1/37, r3 4.1/3.4/31, r4-7 3.0/2.3/94; TE r1 7.6/7.8/8, r2 4.2/4.1/14, r4-7 3.1/2.8/47; QB r1 14.4/13.6/22, r4-7 6.7/6.2/23. - Our check
in-season-ppg-shrink-no-prior-n0.py(2026-09-12): full-sample n0 6, y = 0.879 + 1.042 z; the MAE curve is flat from 4 to 8 (5.486-5.488) and climbs to 5.538 at 20 and 5.542 at 0.5; 60 cluster bootstraps over player-seasons put n0 at 4-8 (p10-p90), median 6, none at the floor or cap; position n0 RB 6, WR 4, TE 6, QB 2. - Harstad (Footballguys, rules-of-thumb): a weak prior should move more on one game than a strong one ("one game affects
our opinion of an unknown much more than of Gronkowski"); averaging ADP with early-season results beat either alone
(0.697 vs 0.548 and 0.659 across positions). Six pseudo-games against a prior season worth 4.5 for established starters
(
in-season/ppg-shrink-pseudo-games) is the same order; the prior here is a population mean, not the player's own line, so it is worth slightly more games, not fewer, because it is unbiased for the group. - FFA rookie projection analysis 2025 (analytics-projection): rookie RB projections had the lowest error, WRs were over-projected as a group (164.7 projected vs 123.4 scored), TEs under-projected on outliers (Fannin +7.7 PPG), QBs the least accurate; the accuracy "depended less on talent evaluation and more on anticipating opportunity and role." That is why a low group target with a fast-moving in-season weight (n0 6, so 4 games are already 40% of the blend) beats a fixed preseason number.
- Draft Value Analytics (consensus-adp): draft capital is the rookie prior; top-50 WRs reach WR2 production within two
seasons at 3:1 over rounds 3-7; first- and second-round RBs average 180+ rookie carries. Our targets carry the same
cliffs (RB r1 12.7 vs r3 6.4, WR r1 7.2 vs r3 4.6) and
outcomes/breakout-draft-capitalthe same hit rates. - Braun (analytics-projection): an ADP-fit Gamma prior updated weekly is the same Bayesian form; his model failed on role changes, which is the argument for the season-to-date term dominating by mid-season.
Evidence against
- The gain is 2.9% of MAE and the draft-capital part of it 1.0% over a plain line; the rest is regression toward the mean that any shrink supplies. At 12+ games and for QB the plain line is as good; the rule's value is weeks 5-12.
- Rows are selected on a season-to-date PPG at the 6-point floor, so the raw mean is biased high by construction (1.057); the same filter is what a lineup question implies, but the 2.9% would be smaller on an unfiltered pool.
- n0 trades off against the slope b (1.04) and the intercept (0.88); a through-origin form picks the same n0 but a different slope. MAE in points rewards a predictor below the mean (A-79); the through-origin blend "wins" by 0.5% on that bias alone and is not used.
- The "no prior" pool mixes rookies (26% of rows) with veterans coming off a lost or low-usage season, whose older seasons carry information this rule ignores (A-81). The rookie-only fit gains less over the line (0.2%) and the rookie-specific targets are worse, so for rookies specifically the draft-round target is a tie-break more than an edge.
- Week-1 prior cells are thin at the top (RB r1 n 7, TE r1 n 8, QB r2 n 4) and are means of active games only, so a rookie who does not play is not in them.
- No school publishes an out-of-sample test of a rookie shrink; Harstad and Braun offer the form, FFA one season of preseason-projection error, Draft Value Analytics hit rates.
How we use it
- No-prior player with 4+ games (no 8+ game, 6+ PPG / 12 QB season last year) and a season-to-date PPG at the floor:
PPG_hat = 0.88 + 1.04 x (6 T + n x) / (6 + n), with T from the position x round table in the claim and n the games
played.
lineup.ppg_no_prior_n0carries the 6. A round-1 WR at 12 PPG after 4 games projects 10.4; a UDFA WR at the same 12 PPG projects 7.7; after 10 games the two are 11.5 and 9.9. - Games 1-3 (unmeasured, A-80): extrapolate the same blend; at 1 game the prior is 86% of z, so a rookie's opening
20-point game moves a round-3 WR from 4.6 to about 8.0, not to 20. At 0 games use the week-1 rookie prior table's median
column, and prefer the ESPN projection when one exists (
in-season/lineup-optimization). - QB: use the per-position line 5.6 + 0.62 x instead of the blend; the draft-round target adds nothing at QB.
- 12+ games: the per-position line (RB 2.7 + 0.73 x, WR 2.0 + 0.70 x, TE 3.3 + 0.52 x) ties the blend; either is fine, and the draft-round target may be dropped.
- Never the trailing-4 mean for a no-prior player: it loses to the season-to-date mean by 4% here as it does for
established starters (
in-season/ppg-shrink-window). - Supersedes
in-season/ppg-shrinkrule 6 (raw trailing-4 mean for rookies and no-prior players). - Do not use the through-origin blend even though its MAE is lowest; it predicts 2.4% under the mean (A-79 decides which loss the optimizer wants; until then every ppg-shrink line stays mean-unbiased).
- Re-run both checks each August after the data refresh; the 2026 season adds about 100 no-prior player-seasons.
Open questions
- Games 1-3 and week 1 (A-80): does the blend extrapolate below 4 games, and does the week-1 prior table beat ESPN's rookie projection?
- Lost-season returners (A-81): a discounted stale season as the target for the non-rookie no-prior group.
- Whether a rookie's preseason rank or ESPN's preseason projection beats the draft-round mean as T (needs the archive, A-01 / A-28).
- Loss function (A-79): the ranking of forms within 0.5% of MAE is provisional.
Claim
For a starter with a prior season (8+ games at 6+ half-PPR PPG, 12+ QB) and four or more games this season, blending
last season's PPG and this season's PPG by games played, PPG_hat = a + b x (n0 x prior + n x season-to-date) / (n0 + n),
does not beat the fixed season-to-date line of in-season/ppg-shrink-window rule 1. Fit leave-one-season-out on 9,255
RB/WR/TE/QB player-games (2019-2025), the pooled blend posts MAE 5.337 half-PPR points against 5.333 for the fixed line,
the training-selected pseudo form gains 0.15% pooled and loses one held-out season (-0.3%), and the games-bucket lines
of the old rule 2 are worse still (5.343). This is a measured null: the single line is the rule at every games count
from four on. The Bayesian reading itself survives: with a population regression term the prior season is worth n0 =
4.5 pseudo-games (cluster bootstrap 2-14), the 4-5 the bucket slopes implied, and the slopes it predicts by games
played (0.57 at 4-7, 0.70 at 8-11, 0.78 at 12+) track the observed (0.54, 0.66, 0.83). Over the 4-17 games a season
offers, though, the blend is nearly a straight line in the season-to-date ratio, so the fixed line already carries it.
Per-position n0 is not identified: RB sits at the 0.5 floor, TE and QB at the 20 cap with degenerate lines (a + b z with
b near 2 as z collapses toward 1). The textbook through-origin blend c x z(n0) picks n0 = 1, c = 0.953 and the lowest MAE
of all (5.319, -0.27%) only because it predicts 2.3% under the mean; MAE rewards a predictor that sits near the median
(0.85 x expectation per outcomes/scoring-distribution-shape), and its implied slopes (0.81-0.89) are far above what
the data show. It is a biased predictor, not a better model.
Evidence for
- Our check
in-season-ppg-shrink-pseudo-games.py(2026-09-12, leave-one-season-out, MAE in points): h6 line 5.3330, season-to-date line 5.3459, games-bucket lines 5.3425, pooled pseudo blend 5.3374, per-position blend 5.3278, through-origin blend 5.3188; the training folds pick the through-origin form five times (n0 1-1.5) and the per-position form once (2025, n0 0.5/5/12/20); per-season gain of the selected form over h6 -0.0030 to +0.0054. Mean prediction over mean outcome: 1.002-1.003 for every form except the through-origin one at 0.983. By position, h6 vs pooled blend: RB 5.417 vs 5.459, WR 5.208 vs 5.201, TE 4.521 vs 4.492, QB 6.204 vs 6.197 (TE and QB gain 0.1-0.7%, RB loses 0.8%). By games played, h6 vs bucket lines vs pooled blend: 4-7 games 5.233 / 5.243 / 5.248, 8-11 5.289 / 5.284 / 5.287, 12+ 5.551 / 5.579 / 5.547. - Our check
in-season-ppg-shrink-pseudo-n0.py(2026-09-12): intercept form n0 4.5, y = -0.069 + 1.030 z, mean prediction / outcome 1.000; 80 cluster bootstraps over player-seasons put n0 at 2-14 (median 4.5, never at the 0.5 floor or the 20 cap). Through-origin form n0 1 (bootstrap 0.5-2), c 0.953, prediction / outcome 0.977; its MAE curve is flat from 0.5 to 2 (5.319-5.316) and climbs to 5.488 at n0 20. Position n0 intervals: RB 0.5-3, WR 1.5-17.5, TE 3.5-20, QB 3-20. - Harstad (Footballguys, rules-of-thumb): averaging preseason ADP with the first four weeks beat either alone (0.697 vs 0.548 and 0.659) across 2013-2015, and a fitted mix gave early-season performance 21% of the weight against 79% for the two preseason views combined. Four games against a prior worth 4-5 games is a 45-50% in-season weight; his 21% sits below that because his prior is two sources, not one.
- Birnbaum on the Tango method (analytics-projection): the right regression adds a fixed number of average pseudo-games regardless of sample size (74 in MLB). The same constant-n0 form is what fits here, with n0 = 4.5 for a fantasy season against a prior season, which is why the season-to-date slope rises 0.54 to 0.80 across the year.
Evidence against
- Per-position n0 hits the grid edges for three of four positions; with a free intercept and slope, n0 trades off against b and a 9,000-row sample cannot pin it. The pooled 4.5 is the only usable value and its interval spans 2-14.
- The per-position blend wins the 2025 fold and gains 0.1% pooled, so a longer sample might turn the null into a small positive for WR/TE/QB; it would still be a fraction of the 1.1% the season-to-date window bought over trailing-4.
- Every form is fit by least squares on ratios and scored by MAE in points; a median-target fit or a squared-error score could reorder the forms by a few thousandths (A-79). The bias diagnosis of the through-origin form does not depend on it.
- Harstad's numbers are three seasons with ADP as the prior; the Tango method assumes binomial per-game luck; Braun and Fantasy Projection Lab offer forms, not results. No school publishes an out-of-sample test of a pseudo-game blend.
- Rows need a prior season of 8+ games at 6+ PPG (12 QB); rookies and returners, where a pseudo-game prior would matter most, are not covered (A-78).
How we use it
- Projection with 4+ in-season games:
in-season/ppg-shrink-windowrule 1's single line at every games count. No games-bucket lines, no pseudo-game blend, no games term. This claim withdrew that file's old rule 2. - What the prior season is worth:
lineup.ppg_shrink_pseudo_games_n0(4.5 games) is the number to quote when a reader asks how much a hot or cold start should move a view: after four games the season to date and last season count about equally; after nine games the season to date carries two thirds. It explains rule 1; it does not replace it. - Never use the through-origin blend (c x blend, no intercept). Its MAE edge is a 2% under-prediction, and the lineup optimizer sums means across a roster, where a low-biased input is a bias, not a hedge.
- Rookies and players without a qualifying prior season:
in-season/ppg-shrinkhandles them as before; the pseudo-game form with a position or draft-capital target is the open item (A-78), not this claim. - Re-run both checks each August after the data refresh; move to
researchingif the per-position blend's gain clears 0.005 in every held-out season.
Open questions
- No-prior shrink target (A-78): for rookies and returners, shrink the season-to-date mean toward what (position replacement PPG, the draft-capital bucket mean, ADP rank) and with how many pseudo-games?
- Scoring rule (A-79): MAE in points rewards a median predictor; the whole ppg-shrink family should be re-scored under squared error or re-fit by least absolute deviations before any 0.1-0.3% ranking of forms is trusted.
- Does a longer sample (2026, 2027) identify per-position n0, and does the WR/TE/QB gain become real?
Claim
For a starter with a prior season and four or more games this season, the in-season half of the PPG shrink
(in-season/ppg-shrink) should use every game of the season to date, not the last four. Fit leave-one-season-out
on 9,255 RB/WR/TE/QB player-games (2019-2025), a shrink on the half-life-6 exponentially weighted season-to-date
PPG projects the next game with 1.1% lower MAE than the trailing-4 shrink (5.333 vs 5.395 half-PPR points), wins
in every held-out season (0.5% to 1.7%), and is the form the training seasons pick every time. The plain
season-to-date mean is within 0.2% of it (5.346), so recency weighting is nearly irrelevant: what matters is not
throwing games away. The line is this game's ratio to prior PPG = 0.33 + 0.63 x (weighted season-to-date ratio),
slope CI 0.60-0.68, shuffle floor 0.04; the in-season side carries more than the trailing-4 line's 0.49 because a
longer average is less noisy. Two things do not help: a games-played term on the trailing-4 line (MAE 5.398 vs
5.395) and fitting in points instead of ratios (5.401). Both are measured nulls. The slope on the season-to-date
mean rises with games played (0.54 at 4-7 games, 0.65 at 8-11, 0.80 at 12+), which is what a Bayesian blend
predicts when the prior season is worth a fixed number of pseudo-games (about 4 to 5 here).
Evidence for
- Our check
in-season-ppg-shrink-window.py(2026-09-12, leave-one-season-out, MAE in points): k4 5.395, k6 5.363, k8 5.347, season-to-date 5.346, half-life 2 5.361, 3 5.342, 4 5.336, 6 5.333, k4 + games term 5.398, points-scale 5.401; training-fold pick h6 in all six folds; per-season gain over k4 0.0051 to 0.0174. By position h6 vs k4: RB 5.416 vs 5.476, WR 5.208 vs 5.269, TE 4.521 vs 4.578, QB 6.204 vs 6.277; the gain grows late in the season (games 12+: 5.551 vs 5.644) because the k4 window discards the most there. - Our check
in-season-ppg-shrink-ewma.py(2026-09-12): pooled y = 0.328 + 0.634 x (mean x 0.936, mean y 0.921); RB 0.260 + 0.697 x, WR 0.318 + 0.632 x, TE 0.527 + 0.448 x, QB 0.400 + 0.582 x; games 4-7 0.413 + 0.536 x, 8-11 0.303 + 0.651 x, 12+ 0.182 + 0.802 x; half-life 2 slope 0.554, 3 0.599, 4 0.619, 8 0.639, season-to-date 0.641. Slopes converge past half-life 4: the weighting is a detail. - Harstad (Footballguys, rules-of-thumb): preseason ADP and the first four weeks correlate with the stretch run about equally (three seasons 2013-2015: 0.649 vs 0.655, 0.466 vs 0.560, 0.548 vs 0.659), and "it's not until week four that the new information approaches equilibrium" with the preseason view. Four games is where the in-season side becomes worth as much as the prior, which is why the shrink line crosses 0.5 near games 4-7 here and keeps rising after.
- Stats and Snake Oil (analytics-projection, soccer xG): a fixed rolling window estimates the metric at the window's midpoint and jumps when a large game drops out; an exponential decay keeps every game at a fading weight. The same failure is what a trailing-4 average does when a 30-point game leaves it.
- Fantasy Football Analytics (analytics-projection): weighting sources by past accuracy vs an equal average was a wash over 12 seasons (MAE 47.8 vs 47.4). Not the same question, but the same lesson: with noisy inputs, a plain average is hard to beat by clever weighting, as the 0.2% gap between h6 and season-to-date shows.
Evidence against
- The gain is 0.06 points per game, 1.1% of MAE. It will not change many lineup calls; it mostly stops the harness from over-reacting to a four-game streak.
- Season-to-date and half-life 6 are indistinguishable in-sample (0.013 points); the h6 pick in every fold could be a coincidence of tiny margins. Treat the two as the same rule.
- A games-played term was tested on the trailing-4 line, where it did nothing; on the season-to-date line the
slope moves 0.54 to 0.80 across the season, but neither games-bucket lines nor a pseudo-game blend beat the fixed
line out of sample (A-77,
in-season/ppg-shrink-pseudo-games: gain 0.15% of MAE, one held-out season negative). - Rows need a prior season of 8+ games at 6+ PPG (12 QB), so rookies and returners are not covered, and the window question for them (no prior to shrink toward) is open.
- Harstad's table is one number per season with no position split past 2016; Stats and Snake Oil is soccer.
How we use it
- Points projection with 4+ in-season games (prior season of 8+ games, 6+ half-PPR PPG, 12+ QB): replace
in-season/ppg-shrinkrule 1's trailing-4 with the season-to-date PPG, weighted by half-life 6 games if convenient (weight 0.5^(age / 6), newest game 1) and unweighted otherwise: PPG_hat = prior PPG x (0.33 + 0.63 x season-to-date PPG / prior PPG). Position lines: RB 0.26 + 0.70 x, WR 0.32 + 0.63 x, TE 0.53 + 0.45 x, QB 0.40 + 0.58 x.lineup.ppg_shrink_ewma_slopecarries the pooled slope. - No week term on the season-to-date line either (A-77,
in-season/ppg-shrink-pseudo-games, 2026-09-12): the games-bucket lines (8-11 games 0.30 + 0.65 x, 12+ 0.18 + 0.80 x; leave-one-season-out MAE 5.343) and a pseudo-game blend (5.337) both lose to rule 1's single line (5.333) out of sample. Use rule 1 at every games count from 4 on. The rising slope by games bucket is real in-sample and is what a 4-5 pseudo-game prior predicts, but exploiting it buys nothing. - No week term on a trailing-4 line, no points-scale fit. Both are measured nulls; do not add them back.
- Fewer than 4 games:
in-season/ppg-shrinkrule 2 stands (k=1 0.73 + 0.21 x, k=2 0.60 + 0.34 x, k=3 0.52 + 0.43 x); those windows already use every game played. - Streaks: a player whose last four games run hot but whose season-to-date mean is ordinary projects to the season-to-date line, not the streak. A four-game hot streak inside a twelve-game season moves the projection by a third of what the trailing-4 rule implied.
- TE and QB keep the weakest in-season carry (TE 0.45, QB 0.58 even on the full season);
ppg-shrinkrule 5 stands. - Re-run both checks each August after the data refresh.
Open questions
- Pseudo-game blend (A-77): measured null, see
in-season/ppg-shrink-pseudo-games(n0 4.5 with a population term, but the fixed line already carries it). - Rookies and no-prior players (A-78): the season-to-date mean shrunk toward what (position replacement level, the draft-capital prior), and with how many pseudo-games?
- Scoring rule (A-79): every form here is fit by least squares on ratios and scored by MAE in points, which rewards a predictor below the mean (a through-origin blend "won" 0.27% on bias alone in the A-77 check); a median fit or a squared-error score may reorder the forms.
- Does the ESPN weekly projection already use the whole season (A-57, blocked on the projection archive)?
Claim
Inside a season, a starter's trailing-4-game half-PPR points per game carries into his next game at about
half strength: this game's ratio to prior-season PPG = 0.46 + 0.49 x (trailing-4 ratio), pooled over RB, WR,
TE and QB with 4+ prior in-season games and a prior season of 8+ games (n 9,255 player-games, 2019-2025,
slope CI 0.46-0.53, shuffle floor 0.03). Written as points that is 0.46 x prior PPG + 0.49 x trailing-4 PPG:
a near-even blend of last season and the last four games, with a 5% haircut because starters' PPG drifts
down year over year. The blend, fit leave-one-season-out, projects the next game with 4.0% lower MAE than the
raw trailing-4 average (5.40 vs 5.62 points), beats prior-season PPG alone (5.85), beats every raw window
from 1 to 6 games, and beats the usage-window share blend converted to points (5.23 vs 5.50 on the 7,799
RB/WR/TE rows). The slope differs by position: RB 0.58, WR 0.49, QB 0.41, TE 0.30. For TE and QB the prior
season alone already beats the trailing-4 average, so in-season TE and QB points are mostly noise. Longer
windows carry more (slope 0.21 at k=1, 0.34 at k=2, 0.43 at k=3, 0.49 at k=4, 0.60 at k=6, 0.66 at k=8),
and the same-k slope grows late in the season (0.41 weeks 5-8, 0.49 weeks 9-12, 0.53 weeks 13-18) as the
prior season goes stale. The weeks 1-6 / 7+ rule in in-season/efficiency-trigger (0.44 + 0.53 x) is the
same line measured at one cut point; it lands within 0.02 points of MAE of the fitted version.
Evidence for
- Our check
in-season-ppg-shrink.py(2026-09-12): pooled y = 0.462 + 0.491 x, mean x 0.933, mean y 0.921, n 9,255; RB 0.374 + 0.575 x (n 2,599), WR 0.458 + 0.486 x (n 4,015), TE 0.657 + 0.297 x (n 1,197), QB 0.572 + 0.409 x (n 1,444); weeks 5-8 0.559 + 0.413 x, weeks 9-12 0.430 + 0.492 x, weeks 13-18 0.436 + 0.529 x; k=1 0.727 + 0.206 x, k=2 0.603 + 0.340 x, k=3 0.522 + 0.428 x, k=6 0.364 + 0.597 x, k=8 0.302 + 0.663 x. - Our check
in-season-ppg-shrink-oos.py(2026-09-12, leave-one-season-out): MAE trail4 5.618, prior 5.851, shrink 5.395, fixed 0.44 + 0.53 rule 5.414, trail1 6.906, trail2 6.055, trail3 5.771, trail6 5.490; share blend 5.498 vs shrink 5.233 on the same RB/WR/TE rows; improvement over trail4 positive in every held-out season (2.9% to 5.3%); held-out fits stay inside a 0.45-0.47 / 0.48-0.51 band. By position: RB shrink 5.476 vs trail4 5.599 vs share blend 5.778; WR 5.269 vs 5.507 vs 5.538; TE 4.578 vs 4.848 (prior alone 4.806, share blend 4.746); QB 6.277 vs 6.604 (prior alone 6.578). - Harstad (Footballguys, rules-of-thumb, 2016 data): preseason ADP correlates 0.599 with late-season points, games 1-4 correlate 0.585, and a simple average of the two 0.682, beating either alone. That is the same near-even blend as our 0.46 / 0.49 weights. His position split (RB early-season 0.77 vs ADP 0.60; WR 0.45 vs 0.55) has the same RB-over-WR order as our slopes, on 18-42 players per position.
- Fantasy Projection Lab (analytics-projection, no numbers): in-season data carries substantially more weight from week 4, accuracy gains are largest weeks 1-4 and plateau after 6-8 observed games. Matches the slope rising with k and flattening past 6.
- The season-level version agrees:
outcomes/scoring-distribution-shapeandtheory/usage-stickinessfind the same partial carry of points and usage year over year.
Evidence against
- MAE is in points, so high-PPG players dominate; the ratio form uses a single linear coefficient and a ratio over an uncapped prior, so a starter whose prior PPG sits near the 6-point floor can post a ratio above 3. A rank- or log-scale fit was not run.
- The gain is 4% of MAE. The remaining error is weekly variance (
outcomes/scoring-distribution-shape), not something a better window will remove. - A longer window does beat this one (A-75,
in-season/ppg-shrink-window, 2026-09-12): the half-life-6 weighted season-to-date shrink projects with 1.1% lower MAE (5.333 vs 5.395) in every held-out season, and the plain season-to-date mean is within 0.2% of it. The trailing-4 line below is the fallback, not the rule. - TE (slope 0.30) and QB (0.41): prior-season PPG alone beats the trailing-4 average, and the shrink gain is small in points. Whether a TD-stripped trailing PPG carries more at TE is A-76.
- Rows need a prior season of 8+ games at 6+ PPG (12 QB), so rookies, returners from a lost season and low-usage players are not covered. Harstad's numbers are one season with thin position cells.
How we use it
- Points projection for a starter with a prior season (8+ games, 6+ half-PPR PPG, 12+ QB) and 4+
in-season games: PPG_hat = prior PPG x (a + b x trailing-4 PPG / prior PPG), with the position line
RB 0.37 + 0.58 x, WR 0.46 + 0.49 x, TE 0.66 + 0.30 x, QB 0.57 + 0.41 x; pooled 0.46 + 0.49 x when a
position line is missing.
lineup.ppg_shrink_slopecarries the pooled slope. A 0.70 trailing ratio projects 0.80, a 1.30 projects 1.10. Superseded when the season-to-date PPG is available: usein-season/ppg-shrink-windowrule 1 (0.33 + 0.63 x season-to-date ratio, half-life 6 weighting optional); this trailing-4 line is the fallback when only the last four games are on hand. - Fewer than 4 in-season games: use the matching window line, k=1 0.73 + 0.21 x, k=2 0.60 + 0.34 x, k=3 0.52 + 0.43 x. In week 1 the prior-season PPG stands alone (times 0.93 for the year-over-year drift).
- Points, not share. For projecting a player's points, this shrink replaces the usage-window share
blend (
in-season/usage-windowrule 1 stays for role and share questions: who starts, target share). - Supersedes
in-season/efficiency-triggerrule 2 (0.44 + 0.53 x at the weeks 1-6 / 7+ cut); same line, now measured at every week and by position. - TE and QB hot streaks are mostly noise: a TE at a 1.5 trailing ratio projects 1.10, a QB 1.19. Do not
bench a prior-season TE1 or QB1 for a four-game streak unless the role changed (
theory/ball-share-shifts). - Rookies and players without a qualifying prior season: superseded by
in-season/ppg-shrink-no-prior(2026-09-12), which shrinks the season-to-date PPG toward the position x draft-round mean with 6 pseudo-games. Not the trailing-4 mean. - Re-run both checks each August after the data refresh.
Open questions
- Window length and weighting (A-75): answered in
in-season/ppg-shrink-window(season to date beats k=4; a week term on the k=4 line and a points-scale fit are nulls). The pseudo-game blend is A-77. - TE and QB (A-76): is the weak TE carry the TD share, so that a TD-stripped trailing PPG carries more?
- Rookies and no-prior-season players (A-78): answered in
in-season/ppg-shrink-no-prior(the position x draft-round mean, n0 6; the position-only target is 0.7% worse). Games 1-3 are A-80. - Does the ESPN weekly projection already apply this shrink (A-57 projected-gap shrink, blocked on the projection archive)?
Claim
A usage shift is real after about three games. For WR and TE target share, averaging the last 3-4 games predicts the next game better than the last game alone (r 0.68 vs 0.61) and better than last season's share (about 0.60); adding a fifth game or more adds nothing. RB carry share and snap share for every position saturate after one to two games because they are already sticky week to week. Last season's share is the best single predictor in weeks 1-2, is matched by the in-season window by week 3, and is beaten by it from week 4 on. Measured on nflverse 2019-2025, regular season.
Evidence for
r between trailing-k-game mean and this game's value (WR/TE/RB active games, same season, our data):
| metric | k=1 | k=2 | k=3 | k=4 | k=6 | k=8 |
|---|---|---|---|---|---|---|
| WR target share | 0.61 | 0.66 | 0.67 | 0.68 | 0.68 | 0.68 |
| TE target share | 0.59 | 0.65 | 0.66 | 0.67 | 0.67 | 0.67 |
| RB carry share | 0.75 | 0.77 | 0.77 | 0.77 | 0.76 | 0.75 |
| RB snap share | 0.74 | 0.75 | 0.75 | 0.74 | 0.73 | 0.71 |
| WR snap share | 0.77 | 0.78 | 0.78 | 0.77 | 0.76 | 0.75 |
| WR PPR points | 0.42 | 0.48 | 0.51 | 0.53 | 0.54 | 0.54 |
| RB PPR points | 0.46 | 0.52 | 0.54 | 0.55 | 0.55 | 0.54 |
Prior-season per-game target share vs this week's share for WRs: r 0.62 in week 1, 0.60 in week 3, 0.59 in week 4, 0.57 by week 12 (n about 480-560 per week). So the 4-game window (0.68) overtakes the prior season around week 4 and the gap widens slowly after that.
Head-to-head on identical rows (WR games with 4+ prior games and a prior season of 8+ games, n 6360, check run 2026-09-10): trailing-4 r 0.66 (CI 0.65-0.67), prior season 0.58, last game 0.56. RB carry share on the same basis: trailing-4 0.75, last game 0.72, prior season 0.56.
Rules-of-thumb writers (FantasyData, Yahoo usage columns) independently use a 3-to-4 game window before calling a usage change structural; Sharp Football finds per-game rates, not season totals, carry the signal. Both agree with the plateau above.
Evidence against
- The window is measured on active games only; a player returning from a multi-week injury has stale trailing games that may not reflect the current role. Treat the return game as k=0.
- Fantasy points keep improving to k=6; the usage plateau is at k=3-4 but scoring noise is larger.
- Rows require a prior season on file for the head-to-head with last season, which drops rookies; for rookies the in-season window is the only usage signal and should be trusted a game earlier.
How we use it
- Blend rule for share projections (WR/TE target share, RB carry share): weeks 1-2, 0.6 prior season + 0.4 in-season mean; week 3, 0.5/0.5; week 4 on, 0.25 prior season + 0.75 trailing-4 mean. Rookies and players with under 8 prior-season games: in-season mean only from week 2.
- Snap share and RB carry share: use the last 2 games, weighted equally; one game already tells most of it.
- A single-game target spike that is not accompanied by a snap-share rise gets no extra weight.
- After a return from injury, restart the window (use the prior blend until 2 post-return games exist).
- Re-run the check each August after the data refresh.
Open questions
- PPG level, not share: answered 2026-09-12 (A-73,
in-season/ppg-shrink). For projecting points the trailing-4 PPG shrink (0.46 + 0.49 x ratio, by position) beats this share blend converted to points by 4.8% of MAE, so rule 1 here is for role and share questions, not for the points projection itself. - Does the window shorten when the trigger is a known event (teammate injury, QB change)? (A-06.)
- Should the blend use team-total context (pass attempts) rather than share alone in low-volume offenses?
Claim
Draft round is the strongest single preseason breakout prior we have, and it is a step function, not a slope. On our 2019-2025 data (half-PPR, total points ranked within season and position, career year = season minus entry year plus one), a round-1 WR posts a top-24 season in 38% of his first three player-seasons (n=89, CI 0.29-0.49); rounds 2 and 3 pooled post 12% (n=195); rounds 4-7 post 3%; undrafted WRs 0%. Per player, 63% of round-1 WRs from complete classes have at least one top-24 season in years 1-3, against 25% for rounds 2-3 and 2% for everyone else. RB is steeper and earlier: round-1 RBs post a top-24 season in 70% of early player-seasons, round 2 in 53%, round 3 in 21%, rounds 4-7 in 5%. The WR round-1 rate is roughly flat in years 1 and 2 (0.31, 0.32) and peaks in year 3 (0.54); the round-2 rate does not rise until year 4 (0.26, thin). "Late-round sleeper" WR breakouts are rare enough to treat as noise: 5 top-24 seasons in 495 player-seasons from rounds 4+ and UDFA.
TE and QB follow the same cliff with a different clock. TE top-12 season rate in years 1-3: round 1 0.38 (n=24), round 2 0.23 (n=40), round 3 0.10 (n=48), rounds 4-7 0.01 (n=144), UDFA 0.01 (n=136). Round-1 TEs hit as rookies (0.62, n=8) and then settle near 0.25; round-2 TEs are near zero as rookies (0.07) and rise in year 2 (0.38) and year 4 (0.60, n=10); rounds 4-7 TEs are a 1% bet in years 1-3 and an 8% bet in years 4-5, which is where the "TE breaks out late" folklore lives. Only 31% of TE top-12 seasons fall in career years 1-3, against 38% WR and 43% RB, so the position is older but not by much. QB top-12 season rate in years 1-3: round 1 0.33 (n=70; 0.23 as a rookie, 0.37-0.38 in years 2-3), round 2 0.18 (n=11), rounds 3-7 and UDFA 0.01 (1 of 98). Per player, 62% of round-1 QBs from complete classes post a top-12 season by year 4 (n=13); no QB drafted in round 3 or later did in years 1-3 (n=36 players).
Evidence for
- Our check
outcomes-breakout-draft-capital.py(rewritten 2026-09-12 to apply all three prongs of the falsifier; the earlier version scored WR only and was judged unsound): WR years 1-3 top-24 rate by round r1 0.38 (n 89) vs rounds 2-3 pooled 0.12 (n 195); TE rounds 1-2 top-12 rate 0.28 (n 64) vs rounds 3+/UDFA 0.02 (n 328); QB round 1 top-12 rate 0.33 (n 70). All three pass. Per round: WR r2 0.12 (115), r3 0.11 (80), r4-7 0.03 (252), UDFA 0.00 (243). RB r1 0.70 (27), r2 0.53 (45), r3 0.21 (62), r4-7 0.05 (219), UDFA 0.02 (174). Any top season in years 1-3, classes through 2023: WR r1 0.63 (27), r2-3 0.25 (67), r4+/UDFA 0.02 (237); RB r1 0.82 (11), r2 0.82 (17), r3 0.38 (21), r4+/UDFA 0.07 (167); TE r1 0.56 (9), r2 0.43 (14); QB r1 0.42 (24). - Inside round 1 the prior is flat (measured null,
outcomes/breakout-within-tier-signals.md): WR picks 1-16 0.385 (n 39) vs picks 17-32 0.380 (n 50); QB picks 1-10 0.33 vs 11-32 0.32. The cliff is at pick 32, not inside the round. - Our check
outcomes-breakout-te-qb-draft-capital.py: TE rounds 1-2 years 1-3 top-12 rate 0.28 (n 64, CI 0.18-0.40) vs rounds 3+ 0.02 (n 328), pooled floor 0.06; round-1 QB years 1-3 rate 0.33 (n 70). Passes. Notes carry the round x year grids. - Fantasy Footballers (2000-2018 classes, 1,349 players, PPR PPG over first three years): WR round 1 10.09 PPG, round 2 8.16 (-19%), round 3+ 4.10; WR2-level (13.54+ PPG) hit rate 18.4% / 10.7% / under 4%. RB round 1 13.06 PPG, round 2 8.59 (-34%), round 3+ 4.33; RB2-level hit rate 55% / 28% / 12% / under 5%. Only one WR outside rounds 1-2 (Tyreek Hill, round 5) hit WR1-level early-career production in 19 classes. TE has the widest round-1 to round-2 gap.
- Fantasy Footballers TE trends (2000-2021, 67 TE breakouts, breakout = first top-12 PPR/g season with 8+ games): career hit rate round 1 88.9%, round 2 44.4%, rounds 3-4 about 20%, rounds 5-7 6.8%; 73.1% of breakouts within three years, 89.6% by year 5; 86.6% broke out with the drafting team. Our years 1-3 ordering (0.38 / 0.23 / 0.10 / 0.01) matches; their career-long rates are higher because they count any later season.
- Footballguys rookie TEs (2016-2025, 141 drafted): rookie-year top-12 rate round 1 54.5% (6 of 11), rounds 2-3 4.7% (2 of 43), rounds 4-7 0% (0 of 87); top-24 81.8% / 14.0% / 4.6%. Our rookie-year cells (r1 0.62, r2 0.07, r3 0.06, r4-7 0.00) agree.
- Fantasy Classroom QB study (2000+): 55.6% of round-1 QBs post a QB1 season in their career, 33% of round-2 QBs (6 of 18), picks 1-5 about 70% vs picks 6-32 about 45%; "the rest of the draft is a crapshoot." A secondary summary of Dynasty Nerds (page not retrievable) puts round-1 QBs above the QB12 average in their first three seasons at 24.5% for 2000-2018, below our 0.33 for 2019-2025, consistent with rookies starting sooner now.
- Apex Fantasy (rookie WR breakouts 2011-2024, 200+ PPR points, n 32): 59.4% were round-1 picks, 84% rounds 1-2, none after round 5. The composite profile (age 22.1, pick 39, college dominator 33.9%) is the film-and-usage school's version of the same cliff.
- Campus2Canton (since 2001): round-1 WRs hit a top-24 season in their first three years 25% of the time on a per-season basis; round-2 WRs hit a top-12 season 22.7% vs round 3 11.7%. Our per-season rate (0.38) is higher because it covers 2019-2025, when early-round WRs were used sooner.
- FantasyLife (2011-present, 45 WR1 and 46 WR2 post-rookie breakouts): 68% of WR1 breakouts happen by year 3, 76% of WR2 breakouts within the first three seasons, so the years 1-3 window is where draft capital gets its test.
Evidence against
- Sample: 6 season pairs; the round-1 RB cell is 27 player-seasons and the per-player cells are 11-27. The round-1 TE cell is 24 player-seasons (8 per year), round-2 QB is 11. Rates carry +/-0.10 of error, TE and QB more.
- The RB round-2 rate (0.53, and 0.82 per player) is as high as round 1 on our window; the 2000-2018 study shows a 34% drop. Either the 2019-2025 RB market moved down a round (teams stopped drafting RBs in round 1) or the cell is noise. Treat rounds 1-2 as one tier for RB until more seasons accrue.
- The round-2 TE year-4 spike (0.60, n=10) and the rounds 4-7 TE years 4-5 rise (0.08, n=72) are the "late TE breakout" signal and both are thin; the folklore's 73%-within-three-years figure is about first breakouts, ours is about all top-12 seasons, so the two do not conflict but neither pins the clock.
- QB top-12 is dominated by veterans (34% of QB top-12 seasons come in career year 8+), so a round-1 QB's 0.33 early rate is a bet on a starting job, not on being a top-12 QB at peak; it says nothing about years 5+.
- Draft capital is a market price, not a mechanism. It works because teams give early picks snaps and targets; where a late pick is handed the role (injury, trade), the usage claims in
theory/usage-stickiness.mdandtheory/ball-share-shifts.mdoverride this prior. - Total-points ranking penalizes injured rookies; a per-game cut (min 8 games) was explored in the earlier pass and moves rates by a few points, not the ordering.
How we use it
- Preseason prior for any player in career years 1-3. P(top-24 season this year): WR round 1 0.35, rounds 2-3 0.12, rounds 4+ and UDFA 0.03. RB rounds 1-2 0.60, round 3 0.20, rounds 4+ and UDFA 0.04. P(top-12 season this year): TE round 1 0.35, round 2 0.20, round 3 0.10, rounds 4+ and UDFA 0.02; QB round 1 0.30, round 2 0.15, rounds 3+ and UDFA 0.02. Combine with the rookie-year carryover claim (
outcomes/breakout-rookie-carryover.md) once a rookie season exists; that claim supersedes this one from year 2 on. - Position clocks. Round-1 TE and rounds 1-2 RB can hit as rookies; round-2 TE and round-1 QB are year-2 bets (rookie rate 0.07 and 0.23); round-1 WR peaks in year 3. Do not roster a rookie round-2 TE or a rookie round-3+ QB for production.
- Years 4-5. For WR and RB the round tiers hold at roughly the same rates. For TE, rounds 3-7 rise to 0.08-0.14, so a year-4 or year-5 TE with a new role is a live 0.10 bet where a year-2 one is a 0.02 bet. For QB, nothing outside round 1 hits (rounds 2+ 0.04, n=56).
- Waiver and bench triage. A rounds 4+/UDFA WR with no usage signal is a 3% bet; do not hold a bench spot for him on draft capital alone. A round-1 WR in year 2 or 3 who has not broken out is still a 0.30-0.50 bet and is worth a bench spot ahead of any veteran with the same projection. A round-1 TE or QB in year 2 without a hit is a 0.25-0.35 bet on the same logic.
- Do not slope inside the tiers. Round 1 versus round 2 is a cliff for WR (0.38 vs 0.12) and QB (0.33 vs 0.18); pick number within round 1 is measured flat for WR and QB (
outcomes/breakout-within-tier-signals.md), and entry age adds nothing within a round. RB and TE pick slot inside rounds 1-2 is unmeasured (cells of 8-16); hold the round rate.
Open questions
- Pick number inside round 1: measured flat for WR (picks 1-16 vs 17-32 gap +0.005, CI -0.20..0.21) and QB, so Fantasy Classroom's 70% vs 45% QB split does not show in years 1-3 on our window; the picks 1-10 WR cell (0.48, n 25) is the one to re-test as classes accrue. RB and TE slot cells are too thin (agenda A-74).
- Is the 2019-2025 RB rounds 1-2 tie real? Re-test when 2026 lands.
- Late-round TE years 4-5 rise: is it a role change (a starter leaving) rather than development, and does target share in year 3 predict it (agenda A-50)?
Claim
The "year-2 WR leap" is real only for receivers who already produced as rookies. On our 2019-2025 data (half-PPR, total points ranked within season and position), a WR who ranked top-36 as a rookie finishes top-24 in year 2 45% of the time (n=29, CI 0.28-0.62) against an all-rookie base rate of 7.5%. The gradient is steep: rookie rank 1-24 gives 0.50, 25-36 gives 0.40, 37-48 gives 0.25, 49-72 gives 0.06, and 73rd or worse gives 0.01 (1 of 173). A rookie WR who was invisible does not leap in year 2. Year 3 is the late-bloomer window, and only for round-1 picks: WRs who missed the top 24 in both years 1 and 2 hit it in year 3 42% of the time if drafted in round 1 (n=12), 7% for rounds 2-3 (n=42), 1% for rounds 4+ and UDFA (n=113). RB carryover is stronger at the top and noisier below: rookie rank 1-24 gives 0.73, then 0.20 / 0.42 / 0.05 / 0.03 for the lower buckets (cells of 10-21).
Rookie target share is not a better year-2 predictor than rookie points, but it is a hard floor. Among WR rookies with 8+ active games (n=138), share and points rank year-2 top-24 hits about equally (AUC 0.90 share, 0.91 total points, 0.89 PPG; r with year-2 PPG 0.68 share vs 0.78 points). The share gradient: mean per-game target share 0.22+ gives 0.53 (n=15), 0.18-0.22 gives 0.31, 0.14-0.18 gives 0.16, and under 0.14 gives 0.00 (0 of 82). Share adds at the margin: inside rookie rank 37-72, the above-median-share half hits 0.25 and the below-median half 0.00 (n=12 each). Per-game ranking with an 8-game floor does not sharpen the signal (top-36 per-game rookies hit 0.42), and rookies with under 8 games never hit in year 2 (0 of 88), so an injury-shortened productive rookie season is too rare in this window to price. For RB, points beat share (AUC 0.84 vs 0.73). The late-season slice is not the signal: mean target share over a rookie's last 6 active games ranks year-2 hits slightly worse than the season mean (AUC 0.87 vs 0.90, gap inside the shuffle floor), the under-0.14 floor holds on the late slice (0 of 73), and a late rise does not rescue a low season share (season under 0.14 but late 0.14+: 0 of 12). Risers (late share 0.04+ above the earlier games) hit 0.14, the same as flat rookies (0.13). For TE (top-12), rookie rank 1-24 gives 0.43 in year 2 (n=7) and 25+ gives 0.04 (n=101), the same shape at a lower level. For QB, rookie rank 1-12 gives 0.60 (n=5), 13-24 gives 0.43 (n=7), 25+ gives 0.09 (n=43); round-1 QBs who missed the top 12 in years 1 and 2 did not hit in year 3 (0 of 10), the opposite of the WR round-1 late-bloomer pattern.
The QB rookie efficiency filter (sack rate under 8.7% and adjusted yards per attempt over 6.5, the Fantasy Classroom rule) is a weak tiebreak, not a prior. On 2019-2024 rookie QBs with 150+ attempts (n=27), rookies who pass both legs finish top-12 in year 2 0.56 of the time (n=9) against 0.22 for those who fail either (n=18); the gap sits exactly at the shuffle floor (p 0.10). Over a year-2-or-3 horizon it is 0.50 vs 0.14 (n 8 vs 14). Fantasy Classroom's 94% is a career-long any-QB1 rate on first-rounders since 1999 and does not reproduce at a fixed horizon. Sack rate alone carries nothing (0.35 vs 0.30) and flips sign inside rookie rank 13-30, where the high-sack rookies who hit are the runners (Fields, Maye, Williams); adjusted yards per attempt is the leg that works (0.50 vs 0.20). Rookie fantasy rank (AUC 0.75) and rookie rushing yards per game (0.74) both out-rank every efficiency measure (sack rate 0.57, AY/A 0.64, EPA per dropback 0.69). Both efficiency metrics are sticky rookie-to-year-2 (r 0.63 sack rate, 0.54 AY/A), so they are real passer traits; they just measure passing efficiency, and a young QB's fantasy jump runs through rushing.
Evidence for
- Our check
outcomes-breakout-rookie-finish.py: value 0.448 (n 29), base rate 0.075, 73+ rate 0.006 (n 173); buckets and late-bloomer rates as above. Both falsifier legs pass. - Our check
outcomes-breakout-rookie-target-share.py: under-0.14-share rate 0.000 (n 82, CI 0-0.04), 0.18+ rate 0.42 (n 31), AUC share 0.899 vs points 0.911. All three legs pass. Notes carry the RB, TE and per-game versions. - Our check
outcomes-breakout-rookie-late-share.py(A-51): AUC last-6 share 0.873 vs season 0.899 (gap -0.027, bootstrap CI -0.07 to +0.01, shuffle floor 0.047); early-games share 0.889; r with year-2 share 0.76 season vs 0.70 late; late < 0.14 rate 0.000 (n 73). Late share alone adds only at the margin: inside the 0.14-0.18 season-share bin, late 0.18+ hits 0.23 vs 0.08 below (n 13 vs 12); inside 0.22+ it adds nothing (0.50 vs 0.57). RB (AUC 0.73 vs 0.72) and TE (0.75 vs 0.79) show the same no-difference. Both falsifier legs pass. - Fantasy Index (2000-2021 rookie classes, n 51): rookie WRs with 750+ receiving yards repeat 750+ in year 2 76% of the time, reach 1,000 yards 60%, finish top-25 PPR 67% and top-12 37%; most misses were injuries. A full-season production bar, not a late-season one, and consistent with our rookie top-36 -> 0.45 top-24 (a stricter bar).
- FantasyLife (2011-present): 76% of post-rookie WR1 breakouts (34 of 45) came from a receiver who already had a WR2 or WR3 season; for years 2-3 breakouts, 71% (20 of 28) had a WR3 or better finish already. Year 2 accounts for 31% of WR1 breakouts and 50% of WR2 breakouts. Prior-season PPR PPG is the best single predictor of next-season points (r 0.67), then yards per route run and target share (0.60 each). Our r 0.78 points vs 0.68 share reproduces that ordering.
- QB List sophomore-surge review: the second-year jump concentrates among rookies who already earned targets; "simply drafting a talented second-year receiver isn't enough."
- Fantasy Footballers TE trends (2000-2021, 67 breakouts, breakout = first top-12 PPR/g season, 8+ games): 73.1% of TE breakouts come within the first three years, 89.6% by year 5, modal breakout age 25. Consistent with our TE year-2 carryover being real but thin.
- Fantasy Classroom QB draft-capital study (first-round QBs since 1999): a rookie QB with a sack rate under 8.7% and adjusted air yards per attempt over 6.5 later posts a QB1 season 94.4% of the time (17 of 18, Mac Jones the miss); failing the sack leg gives 35% (65% bust), passing sacks but failing air yards gives 30-47%. Career-long horizon, first-rounders only, and the "adjusted air yards" definition is not published; we read it as PFR adjusted yards per attempt (the 6.5 bar sits at league average on that scale).
- Our check
outcomes-breakout-qb-rookie-efficiency.py(A-49): pass-both rate 0.556 (n 9, CI 0.27-0.81) vs fail-either 0.22 (n 18); gap 0.33 at the shuffle 95th percentile (p 0.10); filter AUC 0.667 vs rookie points AUC 0.747, so both falsifier legs pass. Year-2-or-3 horizon 0.50 vs 0.14 (n 8 vs 14). Round 1 only: pass 0.67 (n 6) vs fail 0.36 (n 11), against rookie rank 1-12 0.60 vs 13+ 0.42. AY/A alone 0.50 vs 0.20 (n 12 vs 15); sack rate alone 0.35 vs 0.30. Rushing yards per game AUC 0.735. Rookie-to-year-2 r: sack rate 0.63, AY/A 0.54. - NBC Sports / Football Insights (film-and-usage): pressure-to-sack rate is a quarterback trait, not a line trait (Mahomes faces top-5 pressures with the second-lowest P2S), and is "pretty sticky" college-to-NFL (coefficient 0.34); college P2S of 20%+ is a red flag (2 of 9 first/second-round picks reached a second contract vs 12 of 22 under 20%). Consistent with our rookie-to-year-2 sack-rate r of 0.63: the trait persists, which is why it cannot be the breakout signal by itself.
- PFF rookie QB prospect model (123 prospects since 2017): 15.4% of drafted QBs post a top-12 season, 8.9% a top-4, and 82% of top-4 finishers did it within three seasons; sack rate is one input alongside rushing production, scramble rate and draft capital. Rushing enters the model as a positive; we find it out-ranks passing efficiency.
- Sports Info Solutions (via SI, 22 QBs since 2016, first 2-3 seasons, 300+ attempts): intermediate (11-20 yard) on-target rate predicts overall QB performance best (r 0.74 vs 0.55 short, 0.24 deep). Charted, so not on disk; a passing-quality signal that would improve on raw AY/A if it ever is.
- Consistent with
theory/usage-stickiness.md: target share carries y/y at about 0.7 (our rookie-to-year-2 share r 0.74), so a rookie who earned share keeps it; one who did not has nothing to carry.
Evidence against
- Cells are small. The WR top-36 rookie set is 29 players across six classes; the round-1 late-bloomer set is 12; the TE and QB rookie-rank cells are 4-7. A +/-0.15 swing on the headline rates would not surprise.
- Total points, not per game, so a rookie who missed half the year lands in a low bucket. The per-game version was tried (8-game floor) and moved nothing, but with 0 of 88 sub-8-game rookies hitting, the window cannot say whether an injured-but-productive rookie carries over; the prior should be "treat as unknown, use draft capital" rather than 0.
- The 0.14 share floor is a 2019-2025 fact on 82 players; one hit would put the rate at 0.012, still inside the falsifier, but the rule should read "near zero", not "impossible".
- FantasyLife finds 43% of WR2-level breakouts come from players whose previous best was WR5 or worse. That is a lower bar (13.5 PPR PPG) than our top-24 and includes years 4+, so it does not contradict the year-2 finding, but it warns that mid-tier improvements are less predictable than top-24 ones.
- Fantasy Points' sophomore manifesto was paywalled at retrieval; its weight is reduced and its claims are carried only through secondary summaries. Its "late slice" argument does not reproduce on 2019-2025 with raw target share; a secondary summary of its 2025 second-year rankings says catchable targets (ball placement charted) beat raw targets as the year-2 predictor, which we cannot test without charting data (agenda A-66).
- FF Dataroma (breakouts 2011+, non-rookie WRs under 28 beating positional ADP by 60%): the median breakout came with a +4.8 point target-share jump, 4 more routes per game and 2 more TDs in the breakout year, and 75% had better QB play by EPA per dropback. The growth that makes the breakout happens in the breakout season, so a rookie's own late-season share cannot foreshadow most of it; the environment change (QB, vacated targets per
theory/ball-share-shifts.md) is the other half. - The late-slice cells are thin: 42 risers, 17 fallers, 12 "rescued" low-share rookies. A single hit in the rescued group would put it at 0.08, still inside the floor rule but not zero.
- The QB efficiency cells are the thinnest in this file: 9 passers, 18 failers, six draft classes. The pass-fail gap is at the shuffle floor, so a 2026 class that goes 1 for 4 on passers would erase it. The Fantasy Classroom rule was fit on the same kind of data it reports, so its 94% carries selection.
- Sack rate is confounded with mobility. The rookies who took sacks and still hit (Fields, Maye, Williams) are runners; a sack-rate screen would have dropped three of nine year-2 hits. BrainyBallers (top-30 fantasy QBs since 2003) finds college sack rate has no correlation with fantasy finish (r 0.005); only a mid band (2.4-6.5%) shows up more often among top-10s. A sack screen is a passing-quality screen, and passing quality is not what makes a year-2 QB1.
- "Adjusted air yards per attempt" in the source may be a different statistic from PFR adjusted yards per attempt; on raw nflverse air yards per attempt the filter's AUC is 0.46 (worse than chance), so the reading matters, and we chose the one that reproduces the source's direction.
How we use it
- Year-2 prior (supersedes the draft-capital prior once a rookie season exists). P(top-24 this year) for a second-year WR by rookie rank: 1-24 0.50, 25-36 0.40, 37-48 0.25, 49-72 0.06, 73+ 0.02. RB: 1-24 0.70, 25-48 0.30, 49+ 0.04. TE (top-12): rookie rank 1-24 0.40, 25+ 0.04. QB (top-12): rookie rank 1-12 0.55, 13-24 0.40, 25+ 0.08.
- Target-share floor and tiebreak (WR). If the rookie's mean target share over 8+ g
…(truncated)
Claim
Four breakout signals circulate that we cannot yet measure: (1) training-camp reports and coach praise, (2) preseason usage with the first unit, (3) college breakout age and dominator rating for rookies, and (4) the age at which NFL breakouts happen. The measured signals (draft round, rookie-year rank) already explain most of what these are used for, so until they are quantified they get small weight. The best-supported of the four is college breakout age, which one school reports as a 4x multiplier on top-24 odds but without controlling for draft capital; the weakest is camp hype, for which no school we found has published a hit rate at all.
Evidence for
- College breakout age (film-and-usage, PlayerProfiler lineage, PFF): WRs reaching a 20% share of team receiving production before age 20 average 10+ PPR PPG in their best NFL season versus 4.9 for those breaking out at 21+; "early breakouts are 4.2x more likely to register a top-24 PPR season." PFF used dominator rating and breakout age to rank rookie classes 2021-2022 with mixed results.
- FantasyLife (2011-present): year-4+ WR1 breakouts coincide with an offensive change in the breakout year 53% of the time (9 of 17), so a new coordinator or QB is a legitimate late-breakout trigger; that overlaps
theory/coaching-change-share-reset.md. - Age at breakout: the earlier scratch pass on our 2019-2025 file found most first top-24 WR seasons come in career years 1-3, matching FantasyLife's 68-76% within three years; age adds little beyond career year because entry age varies by only about two years.
- Camp and preseason: outlets (RotoBaller, Athlon, SI) publish risers and fallers each August built on first-team reps, route order and coach quotes; the same outlets warn against "double-counting camp hype" already in ADP. No hit rate was published by any of them.
Evidence against
- The breakout-age multiplier is confounded: early college producers are drafted earlier, and draft round alone gives a 3x gap (0.38 vs 0.12) on our data. Nobody we found reported breakout age conditional on round.
- Camp reports are narrative, come from beat writers with access incentives, and are already inside ADP by draft day (consensus-adp). Their marginal information over ADP is unknown and plausibly zero.
- Preseason snap share with starters is a real observable, but the harness does not capture preseason games (nflverse weekly is REG only in our file) and the 2026 preseason has passed.
- Age at breakout duplicates career year for most players; the exceptions (older rookies, late entrants) are too few to measure in 6 seasons.
- Age at NFL entry, the measurable cousin of breakout age, is a measured null once round is held (
outcomes/breakout-within-tier-signals.md, 2026-09-12): WR round 1 young (22.5 or under) 0.38 vs old 0.38, rounds 2-3 0.10 vs 0.14, pooled gap -0.02 (n 284). If college breakout age carries anything beyond round, it is not through the player simply being younger.
How we use it
- Weight while unquantified: camp reports and preseason usage move a player's P(top-24) by at most +0.03 and never move him across a tier from
outcomes/breakout-draft-capital.mdoroutcomes/breakout-rookie-carryover.md. Treat a camp report as a reason to watch weeks 1-2 usage, not as a reason to roster. - College breakout age: for rookies only, and only as a tiebreaker inside the same draft-round tier (a round-2 WR with breakout age 19 over a round-2 WR with breakout age 21). Cap the adjustment at +/-0.05. Plain entry age is not a tiebreaker at all (measured null,
outcomes/breakout-within-tier-signals.md). - Offensive change as a late-breakout trigger (year 4+): a round-1 or round-2 WR entering a new offense gets +0.05 on top-24 odds; otherwise year-4+ players with no hit stay at the position base rate.
- Never diagnose from a single preseason game or one beat-writer quote. Two independent reports of first-team usage is the minimum before rule 1 applies.
Open questions
- Does college breakout age add anything once draft round is held fixed? Needs the draft_picks parquet (cfb_player_id) joined to a college production table (not in our data; WO-0016).
- A logged preseason signal set would quantify camp hype in one season; build the log from August 2027 (WO-0016) and score it against
outcomes-breakout-draft-capital.pytiers. - Preseason snap share with the first unit: is it available from nflverse snap counts for preseason games? The cache holds REG only.
Claim
Once draft round is held fixed, two signals that analysts slope on inside the round add nothing we can measure: the pick number inside round 1, and the player's age when he enters the NFL. On our 2019-2025 data (half-PPR, total points ranked within season and position, career years 1-3), round-1 WRs taken at picks 1-16 post a top-24 season in 38.5% of player-seasons (n=39) and picks 17-32 in 38.0% (n=50); the gap is +0.005 against a shuffle floor of 0.19, so the cliff sits at pick 32, not inside round 1. The WR cliff below round 1 is sharp and immediate: overall picks 33-48 post 16%, 49-64 10%, 65-100 12%. Entry age (entry season kickoff minus birth date, median 22.9, SD 1.0) does not separate WRs within a round either: round 1 young (22.5 or under) 0.38 vs old 0.38 (n 68 vs 21), rounds 2-3 young 0.10 vs old 0.14 (n 108 vs 87), pooled gap -0.02. Both are measured nulls: the round tier is the whole prior, and the tiebreakers the schools sell inside it (a top-10 pick over a pick-28, a 21-year-old over a 23-year-old) do not show up in early-career hit rate at our sample size.
Evidence for
- Our check
outcomes-breakout-within-tier.py: WR round-1 picks 1-16 vs 17-32 gap +0.005 (CI -0.20..0.21, n 89, shuffle floor 0.19); entry-age gap pooled over round 1 and rounds 2-3 -0.024 (n 284). Passes both prongs. Overall-pick cells for QB agree (picks 1-10 0.33, n 51; 11-32 0.32, n 19), and the round-1 cell for QB is nearly all top-10 picks anyway. - Footballguys (323 drafted WRs, 2016-2025): rookie-year top-24 rate round 1 27.9%, round 2 10.0%, round 3 0.0%, rounds 4-7 under 3%; the same step function as ours with no pick-slot split published, which is consistent with none being visible.
- Fantasy Footballers (117 WR breakouts, 2000-2021, first top-24 PPR/g season): breakout age peaks at 23 and 91.5% happen by 26, but 77.9% happen by career year 3 and entry age varies by two years, so age and career year carry the same information for nearly every player; nobody in that study reports age conditional on round.
- 4for4 production curves (players with 2+ top-12 half-PPR seasons, 25 years): rookies produce about 80% of their prime baseline, ages 23-24 about 92%, so a one-year difference in entry age is worth a few percent of production, below what a 0.10 hit-rate gap would need.
Evidence against
- Power. The WR round-1 cell is 89 player-seasons (about 30 players); the CI on the pick-slot gap spans -0.20 to +0.21, so a real +0.10 slope inside round 1 would not be detected. This is a null at the 0.15 threshold, not proof of a flat line. The picks 1-10 cell alone reads 0.48 (n 25) against 0.33-0.35 for 11-32; keep watching it as classes accrue.
- RB and TE are not tested. Their round-1 pick-slot cells are 8-16 player-seasons and read the other way (RB picks 1-16 0.91, n 11, vs 17-32 0.56, n 16; TE 1-16 0.60, n 10, vs 17-32 0.21, n 14). Those are thin and unmeasured, not null; agenda A-74.
- College breakout age (film-and-usage) is a different variable from NFL entry age: it is production age, and one school claims day-1-2 picks with a late college breakout underperform. That remains unquantified in
outcomes/breakout-soft-signals.md(needs WO-0016); this claim does not settle it. - Rounds 4-7 WRs show a +0.05 young-over-old gap (0.06 vs 0.01, n 84 vs 168). It is inside the noise for a 3% base rate and excluded from the pooled prong by design (a day-3 WR is a 3% bet either way), but it is the one cell that leans the schools' way.
How we use it
- Do not slope inside round 1 for WR or QB. A pick-28 WR and a pick-6 WR in years 1-3 carry the same top-24 prior from
outcomes/breakout-draft-capital.md(0.35). The only cliff is round 1 vs round 2 (0.38 vs 0.12). - Entry age is not a breakout tiebreaker. Do not prefer the younger of two same-round WRs on age alone; use rookie-year rank (
outcomes/breakout-rookie-carryover.md) and in-season usage instead. Age matters at the other end of the curve (outcomes/veteran-decline.md), not at entry. - RB and TE pick slot: unmeasured, hold the tier rate. Do not upgrade a top-16 RB or TE on slot until A-74 has 30 player-seasons per cell.
- Sample caveat. A trade or draft-day argument that rests on "he was a top-10 pick" over a pick-25 is, on our data, worth at most the width of this CI; treat it as +0.00 to +0.05, never a tier move.
Open questions
- Picks 1-10 vs the rest of round 1 for WR (0.48 vs 0.34, n 25 vs 64): re-test with the 2026 and 2027 classes.
- RB and TE pick slot (A-74) once cells reach 30.
- Does entry age matter for the length of the window (a 21-year-old round-1 WR has more year-4-5 chances) rather than the years 1-3 rate? Needs the per-player any-hit cut by age, which the sample cannot support yet.
Claim
Every player projection is a distribution. Two players with the same median differ in value by their spread and skew, and the right one to draft depends on roster slot: starters want a higher floor, late bench spots want a higher ceiling. Availability base rates, measured on nflverse 2019-2025 regular seasons for regulars (snap share >= 0.40 in each of the team's last two games; check above, run 2026-09-11):
| Position | P(miss next game) | n player-games | Absence: mean / median games | P(2+ games) | No return that season |
|---|---|---|---|---|---|
| RB | 0.052 | 3,015 | 2.2 / 1 | 0.49 | 0.36 |
| WR | 0.074 | 8,112 | 1.9 / 1 | 0.40 | 0.26 |
| QB | 0.081 | 2,985 | 2.4 / 2 | 0.56 | 0.42 |
| TE | 0.099 | 3,742 | 1.7 / 1 | 0.34 | 0.27 |
| pooled | 0.077 (CI 0.073-0.080) | 17,854 |
Miss counts any cause (injury, rest, benching); 56% of regular misses appear on that week's
injury report, the rest are mostly IR placements that never hit the report. Age adds almost
nothing to the per-game miss rate: RB 0.050 at 25 and under versus 0.053 at 29-31; WR 0.068
versus 0.084; QB and TE flat. Mean absence lengths are among absences that ended within the
season; a quarter to two-fifths do not, so the unconditional expected loss is well above the
2-game mean. After a 2+ game absence, first-game scoring is 0.72 (RB), 0.74 (WR), 0.59 (TE),
0.78 (QB) of the pre-absence 4-game baseline, against a no-absence floor of about 1.0-1.1, and
games 3-6 after return stay at 0.79-0.86 for RB/WR/TE while QB recovers to 1.0. That games 3-6
figure is a fork, not a level: the half whose snap share is back score 0.96, the rest 0.62, and
the game-2 snap share is the tell (in-season/injury-processing.md rules 4b-4c). RBs decline
sharply around age 28-29; WRs later; QBs latest (outcomes/veteran-decline.md). A player's
own history does carry: regulars with two or more multi-game absences over the prior two
seasons miss 1.3 more games the next season than regulars with none (CI 0.8-1.9, shuffle
95th percentile 0.56; analysis/checks/outcomes-injury-history-carryover.py, run 2026-09-11):
| Position | Next-season games missed: 2+ absences / 1 / none | n (2+ / none) | P(miss 4+): 2+ vs none |
|---|---|---|---|
| QB | 5.1 / 3.5 / 3.1 | 49 / 57 | 0.53 vs 0.35 |
| RB | 4.4 / 3.7 / 3.4 | 52 / 74 | 0.54 vs 0.34 |
| WR | 4.7 / 4.2 / 3.4 | 106 / 230 | 0.47 vs 0.35 |
| TE | 4.3 / 4.1 / 3.2 | 70 / 84 | 0.50 vs 0.36 |
Prior season alone works as well: 3+ games missed versus 0-1 gives +1.4 games. The season-to-season Pearson r of miss counts is only 0.16, so history sorts the tails, not the middle. Three refinements were measured on the same pool (both checks run 2026-09-11):
- Injury type is a null, in the wrong direction. Among regulars with at least one
injury-report episode over the prior two seasons, those whose episodes include a soft-tissue
injury (hamstring, groin, calf, quad, thigh, oblique, pectoral, achilles) miss 0.64 fewer
games the next season than those with only bone, joint or head episodes, matched on position
and absence count (CI -1.26 to -0.18, shuffle 95th percentile 0.52, n 947;
analysis/checks/outcomes-injury-history-type.py). At 0 / 1 / 2+ absences: soft 3.1 / 3.5 / 4.2, other-only 3.7 / 4.2 / 4.7. Same-type recurrence (two or more episodes with one body part repeating) is +0.4 over two episodes of different types (4.15 vs 3.76, n 208, inside the shuffle floor). Recency is flat: a last report row in weeks 14+ gives 4.0, weeks 9-13 3.9, weeks 1-8 3.7, none 3.8. - The absence with no report row is the strong signal. Regulars whose multi-game absences never appeared on the injury report (IR placed without a report, mostly season-ending) miss 4.7 (one absence) and 5.4 (two or more) games the next season, above every on-report group at the same count. Severity, not body part, is what the count is carrying.
- History is not exposure. Matching regulars on within-position tercile of snaps per game
in season S as well as position, the 2+-absence gap is +1.05 games (CI 0.56-1.66, shuffle
95th percentile 0.63, n 722;
analysis/checks/outcomes-injury-history-workload.py) against +1.32 position-only. Workload itself runs the other way: the top snaps-per-game tercile misses 2.4-2.9 games the next season and the bottom tercile 4.7-5.5 at every position (touches per game the same), which is selection on all-cause misses (fringe regulars lose jobs, get scratched, rotate), not durability. The SIS "playing-time paradox" does not reproduce on all-cause misses.
Evidence for
- Measured (film-and-usage, nflverse): the table above. Pooled per-game rate 0.077, roughly twice the 4% figure the earlier literature-based version of this file carried, because "miss" here includes rest and benching and the regular filter is snap-based, not ADP-based.
- ADP-tier games missed (analytics-projection, Rotobanter, 2021-2025, top-100 ADP): RB 2.4 games per season in rounds 1-2 vs WR 2.2; rounds 6-8 RB 3.8 vs WR 3.3; RBs miss about 10% more games than WRs. Consistent in direction with the per-game rates being similar but RB absences longer and more often season-ending.
- Injury studies: ~4.1% per-game injury probability, ~3.1 games mean absence, ~14.2 healthy games per 16 (analytics-projection), which is the injury-only rate and matches our 56% on-report share times 0.077.
- Post-injury PPG: -0.50 PPG next season across skill positions, QB -1.95 (film-and-usage / sports medicine, PMC); PFF's 10 serious-injury RBs came back at 4.34 YPC from 4.71 with 6 of 10 holding yards after contact.
- Injury history (analytics-projection): the Rotoworld decision-tree study (1,332 soft-tissue, 350 bone, 249 concussion cases) found the count of prior same-type soft-tissue injuries and days since the last one the top predictors of soft-tissue injury, with prior injuries of a different type near useless; Sports Info Solutions' XGBoost model (RMSE 3.6 games one year out for skill players) ranks games missed in the past 1 and 3 years 9th-10th of its features behind age, BMI and playing time; Draft Sharks reports R^2 0.40 / MAE 1.6 games for predicted games missed, with time since last soft-tissue injury beating cumulative counts.
- Hamstring recurrence (film-and-usage, sports medicine, NFL 2009-2020 and a 12-year review): about a third of NFL hamstring injuries recur (33-38%), a quarter of those in the same season, and the largest recurrence risk is return within two weeks of the injury (13.4% in that window); risk factors for any-season recurrence are a more recent injury year, longer experience and playing WR/RB/TE/DB. That is a within-injury, short-horizon effect (weeks); our season-ahead check finds no soft-tissue premium and no recency effect at the season horizon, so the two are compatible: the recurrence risk lives inside
in-season/injury-processing.md, not in the season availability multiplier. - Fantasy Points ("injury prone is a lie", analytics-projection): prior-year games played predicts next-year games played with slope 0.069 (next = 13.11 + 0.069 x prior), near zero; matches our Pearson r 0.16 and the "tails, not middle" reading. Footballguys (rules-of-thumb, 2017-2024, 43 RBs with 300+ touches): the average RB plays 13.7 of 16-17 games and the 300-touch group is not visibly worse; consistent with our null-to-negative workload slope.
- Injury-risk tools (film-and-usage): PlayerProfiler's top risk bands miss 2.0 (WR), 2.9 (QB), 3.6 (TE) games versus 0.7 in the low bands, and RBs above the 60th percentile miss about 3x the games of those below; these bands are built mostly from history plus projected workload.
- Weekly variance by position: TE highest, QB lowest, RB and WR similar in half-PPR; measured in
outcomes/scoring-distribution-shape.md, where the TE gap turns out to be an expectation-size effect (at equal expectation TE spread equals RB/WR) and SD in points is about 3.8 + 0.28 x expectation.
Evidence against
- The per-game rate does not show "RBs injured most often": among regulars, RB is the lowest per-game miss rate. RB fragility shows up in absence length and season-ending share, not frequency. Snap-share filters also select the most durable RBs (workhorses), so the RB row is survivorship-shaded.
- TE and QB rates are inflated by non-injury misses (TE rotation and healthy scratches, relief QBs who played 40%+ then sat when the starter returned). The injury-only rate is lower.
- Age gradient is measured on players who were still regulars, so it is conditional on surviving to be a regular; it does not say a 31-year-old is as likely to be a regular next year.
- Published percentile projections are rare and methodologically opaque; most "ceiling/floor" talk is narrative.
- Return-game ratios are means of ratios with n of 28-78 per position; the TE first-game figure especially is soft.
- The history check counts all-cause misses and requires the player to be a regular in season S, so the no-history group includes healthy scratches and the history group is survivorship-shaded (players who never came back are not in it). The history group is 0.3-1.1 years older, though age alone was a null. Two sources say injury type and recency carry more than the count; measured 2026-09-11, neither does at the season horizon (type reversed, recency flat), but the nflverse report field is body part only (a "hamstring" row is a strain of any grade) and about 44% of misses have no report row at all, so the type test is on the on-report share and the recency test uses report week, not injury date.
- The workload control (snaps per game in S) is itself shaded by the outcome definition: a low-snap regular is nearer the roster bubble, and all-cause misses include losing the job. An injury-only outcome (misses with a report row within two weeks, exploration only) showed the history gap near zero and the workload slope flat, but that proxy catches only a third of misses (IR placements without a report are lost), so it is not the arbiter. The transaction feed (WO-0013 scope) would be.
How we use it
- Represent each player as (p10, p50, p90) season points. If a source offers only a point projection, widen it with the measured tier widths in
outcomes/scoring-distribution-shape.mdrule 4 (RB/WR 14+ PPG about -32% / +32%, 10-14 PPG -35 to -45% / +32 to +44%, TE -35% / +54%, QB -24% / +26%, sub-10 PPG -50% / +70-90%); the old flat +/-20/30/35% defaults were right only for top-tier starters. Then apply availability (rule 2) on top, which skews RBs downward. - Availability multiplier for season-long expectation: expected games = 17 x (1 - per-game miss rate x expected absence). Use per-game miss rates RB 0.05, WR 0.07, QB 0.08, TE 0.10 and expected absence (including no-return) RB 3.0, WR 2.4, QB 3.2, TE 2.2 games, giving multipliers RB 0.85, WR 0.83, QB 0.75, TE 0.78. QB and TE include non-injury misses; for a locked-in starter at those positions use RB's 0.85 instead.
- Slot rule: for starting slots (rounds 1-6) rank by p50 adjusted for availability; for bench (round 10+) rank by p90; rounds 7-9 use 0.5 x p50 + 0.5 x p90.
- Age flags: superseded by
outcomes/veteran-decline.mdrule 1 (RB 28+ cut 12%, RB 29 cut 15%, WR 29+ and TE 30+ cut 8%). Do not add an age-based availability cut: the per-game miss rate is flat with age (measured null, 2026-09-11). - Injury-history flag (measured 2026-09-11): two or more multi-game absences (runs of 2+ missed games) over the last two seasons, or 3+ games missed last season: cut the availability multiplier by a further 0.07 (1.3 of 17 games; the measured range 0.8-1.9 games is 0.05-0.11). Exactly one multi-game absence: 0.03. Take the larger, never both. Applies to regulars at every position; QB shows the largest gap, RB the smallest.
- 5b (measured null, 2026-09-11): do not add for soft-tissue history, hamstring included, and do not add for a late-season injury; body part and rep
…(truncated)
Claim
A player's weekly score scatters around a calibrated expectation with a spread that is neither purely additive (constant SD in points) nor purely multiplicative (constant SD in ratio). It is right-skewed with a hard wall at zero, the median sits below the mean, and the spread grows with the expectation at roughly a quarter of a point per point. Season to season, a starter's PPG ratio is centred below 1 (regression) with a p10 near 0.55 that is far wider than weekly sampling noise alone would produce.
Measured on nflverse 2019-2025 regular seasons, half-PPR, active player-weeks (snaps or touches
0). Expectation is the leave-one-out season mean (8+ games), the closest thing to a perfectly calibrated projection. Checks:
analysis/checks/outcomes-scoring-distribution-weekly.py(weekly, run 2026-09-11) andanalysis/checks/outcomes-scoring-distribution-season.py(season, run 2026-09-11), both passing.
Weekly, RB/WR/TE pooled: SD(points/expectation) falls 0.44 (CI 0.41-0.47, n 29,778) from the 4-8 bin to the 16+ bin; a purely additive model predicts 0.75, a multiplicative one 0. SD in points rises 3.6 over the same bins. Linear fit on bin means: SD_points ~= 3.8 + 0.28 x E.
Weekly points / expectation, p10 / p50 / p90 (SD of ratio; P(score <= 0)), by leave-one-out bin:
| Pos | E 4-8 | E 8-12 | E 12-16 | E 16+ |
|---|---|---|---|---|
| RB | 0.11 / 0.76 / 2.39 (0.97; 4%) | 0.21 / 0.83 / 2.00 (0.70; 1%) | 0.34 / 0.89 / 1.71 (0.56; 0%) | 0.35 / 0.82 / 1.57 (0.49; 0%) |
| WR | 0.05 / 0.76 / 2.32 (0.93; 10%) | 0.22 / 0.88 / 1.97 (0.71; 3%) | 0.26 / 0.83 / 1.67 (0.56; 1%) | 0.32 / 0.81 / 1.52 (0.48; 1%) |
| TE | 0.14 / 0.78 / 2.21 (0.88; 8%) | 0.20 / 0.82 / 1.84 (0.66; 3%) | 0.27 / 0.73 / 1.57 (0.54; 1%) | n < 50 |
| QB | n 59 | 0.22 / 1.12 / 2.10 (0.71; 4%) | 0.36 / 1.01 / 1.72 (0.53; 2%) | 0.41 / 0.95 / 1.48 (0.43; 1%) |
SD in points by bin: RB 5.6 / 6.9 / 7.6 / 9.1, WR 5.4 / 7.0 / 7.6 / 8.7, TE 5.0 / 6.2 / 7.1, QB
flat at 7.3-8.2 across bins. Skew of the ratio is about 1.0 for RB/WR/TE at 8+ and 0.2-0.4 for
QB. With the operational trailing same-season mean (4+ prior games) instead of leave-one-out, the
table is the same to +/-0.05 except the top bin's median drops (WR 16+ p50 0.75, RB 0.80): a hot
trailing mean overstates the expectation and must be shrunk (in-season/usage-window.md).
Season to season (8+ active games in both seasons, so the injury tail is excluded and
handled by player-range-of-outcomes.md rule 2), PPG(s+1) / PPG(s), p10 / p25 / p50 / p75 / p90:
| Pos, PPG in s | n | ratio quantiles |
|---|---|---|
| RB 14+ | 63 | 0.61 / 0.73 / 0.89 / 1.06 / 1.18 |
| RB 10-14 | 88 | 0.48 / 0.62 / 0.88 / 1.03 / 1.27 |
| RB 6-10 | 104 | 0.35 / 0.56 / 0.82 / 1.16 / 1.56 |
| WR 14+ | 50 | 0.60 / 0.76 / 0.87 / 0.99 / 1.13 |
| WR 10-14 | 138 | 0.60 / 0.74 / 0.91 / 1.06 / 1.20 |
| WR 6-10 | 197 | 0.44 / 0.63 / 0.87 / 1.18 / 1.46 |
| TE 10+ | 34 | 0.53 / 0.69 / 0.82 / 1.04 / 1.26 |
| TE 6-10 | 86 | 0.56 / 0.68 / 0.89 / 1.08 / 1.25 |
| QB 14+ | 116 | 0.73 / 0.83 / 0.96 / 1.08 / 1.21 |
| QB 10-14 | 34 | 0.80 / 0.91 / 1.08 / 1.32 / 1.56 |
| pooled RB/WR/TE 10+ | 373 | 0.54 / 0.71 / 0.88 / 1.05 / 1.21 |
Resampling each player's own season-s weeks into a fake s+1 gives a p10 of 0.82, so most of the downside is real change (role, health, offense), not sampling noise. Of 8+ PPG regulars, the share who play fewer than 8 games the next season is RB 0.14, WR 0.12, QB 0.26 (includes benching), TE 0.04 (n 70); zero games RB 0.03, WR 0.02, QB 0.05.
Evidence for
- Measured (film-and-usage, nflverse): the two tables above; both checks pass with the falsifier margins stated.
- Weekly projections (analytics-projection, FFA, 2015-2025, 6,000+ source-weeks): best-source MAE QB 6.1-6.2, RB 5.1-5.2, WR 4.8-4.9, TE 3.7-3.9; R^2 3-23% with RB highest (19-23%) and QB/WR/TE in single digits; mean error -0.10 overall (QB -0.76, WR +0.29). Consistent with our SD in points of 5-9 and with position ordering of MAE following expectation size, not position per se.
- Error shape (analytics-projection, Subvertadown): the distribution of projection errors has the same shape as the distribution of points; projections narrow QB RMSD only from 8.0 to 7.5, "the tails continue to extend just as wide". Our QB SD in points of 7.3-8.2 matches.
- Coefficient of variation (market-efficiency, Underdog, 2015-2021): top-tier TE 0.63, WR 0.58, RB 0.54, QB 0.36; deeper pools TE 0.70, WR 0.67, RB 0.63, QB 0.39; "the more volume the position gets, the less variance". Our ratio SDs at the 12-16 bin (TE 0.54, WR 0.56, RB 0.56, QB 0.53) say the position gap is mostly an expectation-size gap.
- Skew (rules-of-thumb and analytics, Footballguys and FFA): gamma-like, hard wall at zero, easier to score 20 over than 20 under the average. Our median/mean ratio of 0.8-0.9 for RB/WR/TE and skew ~1.0 agree.
- Value of variance (contrarian-upside, Fantasy Footballers, best ball 2017-2022): at equal points, +100% SD raises top-6 odds RB 9%, WR 8%, QB 6%, TE 4%; points coefficients are 2-3x larger. In managed H2H sims (analytics-projection, snippet only), consistent rosters won 51.4% (2024) and 51.2% (2025) of matchups versus 48.5% and 48.9% for boom/bust rosters at equal means.
Evidence against
- Leave-one-out expectation is more calibrated than any live projection; the operational spread around an ESPN or trailing-mean projection is wider (trailing 4-8 bin ratio SD is 0.03-0.06 higher) and its median is further below 1 at the top. The ESPN-specific version waits on A-01 (A-30).
- Season table cells have n 34-197; the 14+ tiers and TE 10+ are thin, so p10/p90 there carry +/-0.05-0.08 bootstrap error.
- The season ratio conditions on 8+ games both years, so it understates total downside; the checks' games-played rows and rule 2 of
player-range-of-outcomes.mdcarry the rest. - Half-PPR only. Full-PPR compresses WR/TE ratio spread slightly (more receptions, fewer zero weeks); standard widens it.
- QB 4-8 and TE 16+ bins are too thin to report.
How we use it
- Weekly simulation (spec 10): for RB/WR/TE draw from a right-skewed distribution (gamma or lognormal) with mean E and SD = 3.8 + 0.28 x E, floored at zero; for QB use SD 7.5 with mild skew. Calibrate against the p10/p50/p90 table by bin rather than the SD alone. Median is 0.82-0.89 x E for RB/WR/TE and 0.95-1.0 x E for QB.
- Weekly floor/ceiling for lineup rules (
in-season/lineup-optimization.mdrule 2): p10 and p90 come from the table by position and expectation bin. Do not add a TE-specific variance penalty; at equal expectation TE spread equals RB/WR. - Trailing-mean inputs: shrink a 4+ game trailing mean toward the position prior before applying the table (usage-window rules); un-shrunk hot streaks have a median outcome of 0.75-0.80 x the trailing mean at 16+.
- Draft-time season (p10, p50, p90) from last-season PPG, for players expected to play: use the season table by position and tier and then apply availability separately. This supersedes the +/-20/30/35% defaults in
player-range-of-outcomes.mdrule 1. As widths around a regressed p50 (a projection that already regresses), p10/p50 and p90/p50 are: RB 14+ 0.68 / 1.33, WR 14+ 0.69 / 1.30, QB 14+ 0.76 / 1.26, TE 10+ 0.65 / 1.54, RB 10-14 0.55 / 1.44, WR 10-14 0.66 / 1.32, RB 6-10 0.43 / 1.90, WR 6-10 0.51 / 1.68. The old symmetric +/-30% is right only for 14+ PPG starters. - Value of consistency: in this managed H2H league, at equal mean prefer the lower-SD starter by default (about +3 points of win rate per the H2H sims); flip to the higher-p90 option only under the lineup-optimization rule 2(b) and playoff rule 6. Best-ball variance premiums do not transfer.
- Record the (p10, p50, p90) used in every decision record so REFLECT can score the widths against outcomes (spec 03 rule 7).
Open questions
- Does the ESPN weekly projection's error follow the same SD ~= 3.8 + 0.28 x E line and the same median-below-mean skew, or is ESPN calibrated in the median (A-30, needs A-01 data)?
- Does lower weekly SD at equal mean raise H2H win rate in this league's format by the 2-3 points the sims claim (A-31, spec 10 simulator)?
- Full-PPR and standard versions of the weekly table (scoring columns exist in the weekly file; a one-line change to the check).
- Are boom weeks (>= 2 x E) predictable from TD share or air yards, or pure noise (would refine rule 1's skew)?
Claim
A team's season is a distribution produced by simulating every remaining matchup thousands of times from per-player weekly distributions, the real schedule, and the league's playoff format. The decision-relevant outputs are P(playoffs) and P(championship), not projected wins. Because H2H playoffs are single elimination over 2-3 weeks, ceiling matters more than floor once seeded, and correlated players (same-team stacks) raise the ceiling without changing the mean.
Evidence for
- Monte Carlo playoff simulators (10,000+ seasons) are standard practice and produce well-behaved playoff odds from team mean/SD and remaining schedule (analytics-projection).
- Stacked QB-WR pairs have the same season median as unstacked pairs but a higher weekly boom rate (~15% vs 12%) and higher bust rate: correlation is a week-to-week variance lever (market-efficiency).
- Schedule-luck simulation ("who would have won on every schedule") separates team strength from schedule noise.
Evidence against
- Simulations are only as good as the per-player distributions; naive normal draws understate tail outcomes and ignore injury clustering.
- Most public simulators use recent-5-week mean and SD, which overreacts to small samples.
- Our own player distributions are uncalibrated until at least one season of REFLECT scoring.
How we use it
harness/owns a season simulator (work order once the draft is done). Inputs: per-player (p10, p50, p90) weekly, availability, schedule, league playoff format. Output: P(playoffs), P(bye), P(championship), expected wins, for all teams by tid.- Every roster decision (trade, waiver, lineup) is evaluated as delta P(championship). Before week 8 also report delta P(playoffs).
- Risk posture (Q-0005, answered by the league owner 2026-09-09,
school: league-owner-preference, weight: 1.0): maximize P(championship) first, then P(playoffs). Pre-playoffs, accept a lineup or trade that lowers mean points if it raises P(championship). The owner expects this to be re-evaluated as the season progresses: when P(playoffs) falls low enough that missing the playoffs dominates the title odds (interim rule: P(playoffs) below 0.45 after week 8), switch to maximizing P(playoffs) and say so in the decision record. Once eliminated from playoff contention, the harness stops trading and plays for points-for tiebreakers only. - Playoff weeks: maximize weekly p90 subject to p50 not dropping more than 5% below the alternative; prefer positively correlated starters (theory/correlation-and-stacking.md).
- Publish P(playoffs) and P(championship) for every tid to
state/public/(no names; tid labels only).
Open questions
- Playoff format (teams, weeks, byes, reseeding) unknown until mSettings; changes rule 4 weighting.
- How much should we shrink recent-form estimates toward preseason projections? Interim: weight preseason 0.6 until week 6, then linearly to 0.3 by week 12.
Claim
Age is a real but weak and lumpy decline signal; observed role loss is a much stronger one; per-touch efficiency is a weak leading indicator of role loss and no indicator at all of next-season points. On our 2019-2025 data (half-PPR, per active game, qualifying prior season), the median next-season PPG ratio is 0.88 for RBs aged 25-27 and about 0.74 for RBs aged 28+ (gap 0.16, n=43 old RBs, just above the shuffled-age noise floor of 0.15). Dropout (no 6-game season the next year) rises from 0.17 to 0.23 for RBs 28+ and to 0.22-0.27 for WRs and TEs at 30+. WR and TE medians fall at 29 and again at 30+, but the WR 29 bucket (0.67, n=27) is noisier than the 30+ bucket (0.80, n=58), so the WR curve is better read as "risk rises past 28, timing unpredictable" (Harstad's mortality-table model) than as a smooth curve. The stronger in-season signal is snap-share drift: a weeks 1-4 snap share at least 0.10 below the prior-season mean raises P(rest-of-season PPG under 80% of prior) from 0.31 to 0.58 (n=687, CI on the gap 0.19-0.35, noise floor 0.08), and it works at every position (RB 0.61 vs 0.29, WR 0.59 vs 0.34, TE 0.55 vs 0.28). Age compounds it modestly: 27+ with drift 0.64, under 27 with drift 0.53.
Efficiency (A-19). A player who held his role in season s (snap share within 0.05 of s-1) but whose efficiency fell (RB yards per carry down 0.5+, WR/TE yards per target down 1.0+) loses the role in s+1 (snap share down 0.10+ or no 6-game season) 0.56 of the time against 0.43 when efficiency held (gap 0.125, CI 0.01-0.22, shuffle floor 0.10, n=459). It is an RB and WR effect (RB 0.67 vs 0.45, n=33 fallen; WR 0.56 vs 0.43, n=63) and a TE null (0.45 vs 0.42). It leads rather than coincides: the same-season correlation between the efficiency change and the snap-share change is 0.06 (RB), 0.00 (WR) and -0.14 (TE). But it does not move the next-season PPG ratio (P(ratio < 0.8) 0.56 vs 0.55) or dropout (0.24 vs 0.26): the bad efficiency already lowered season-s points, and in s+1 efficiency regresses back (y/y r: RB YPC 0.30, WR YPT 0.27, TE YPT 0.39; points per touch 0.28/0.42/0.34) while the role shrinks, and the two cancel. Efficiency by age is a measured null on this sample: the year-over-year YPC change for RBs 27+ minus 23-26 is -0.08 yd (floor 0.25, n=35 old pairs) and the yards-per-target change for WRs 25+ minus 23-24 is -0.40 yd (CI -0.82 to +0.01, floor 0.44, n=346). The direction matches RotoWire (WR yards per target peaks at 25; RB YPC falls 1/8 yd a year) at every position from 25 on, but the effect is inside the noise, so it gets no weight on top of the age prior.
Prior-season decline (A-20). A 20%+ PPG fall from one season to the next is neither noise that reverses nor the start of a slide; it is a fork. Matched on the new (lower) PPG level, decliners and holders have the same composite bad rate next season (gap -0.04, CI -0.12 to +0.04, floor 0.07, n=738), but the composite hides two opposite moves: decliners drop out of the league or the role three to four times as often (0.34 vs 0.09; matched dropout gap +0.16), while the ones who keep a 6-game role score at the new level or slightly above it (median PPG(s+1)/PPG(s) 0.97 vs 0.85 for holders, i.e. no further regression because the regression already happened). Only 0.13 of decliners get back to 0.90 x the pre-decline PPG (RB 0.16, WR 0.10, TE 0.17). The fork is steeper for old players (RB 28+, WR 29+, TE 30+: dropout 0.44, survivors' median 0.92, full recovery 0.04) and for decliners whose snap share also fell 0.10+ (dropout 0.45, full recovery 0.08); a young decliner who kept his snap share drops out 0.20 of the time and fully recovers 0.20 (survivors' median 1.01). That is Harstad's mortality model in our data: 17 of 100 top careers ended with two consecutive declines against 25 expected by chance, and here survivors of a decline decline again (ratio < 0.8) 0.33 of the time against 0.43 for holders.
Decliner fork drivers (A-48). The dropout side of the fork is health and age, not draft capital and not the shape of the decline. An injury-shortened decline season (12 or fewer active games) drops out the next year 0.46 of the time against 0.24 for a healthy 14+ game decline (gap -0.22, CI -0.32 to -0.12, shuffle floor 0.12, n=266); the short-season decliner who never appeared as "Out" on a weekly report (an IR placement or an unlisted absence) is the worst case (dropout 0.49, full recovery 0.05, n=85). Age inside the old bucket compounds it: 2+ years past the threshold (RB 30+, WR 31+, TE 32+) dropout is 0.56 and a third are not on any roster the next year; the old short-season decliner drops out 0.51. Draft capital is a measured null (round 1-2 dropout 0.31 vs 0.36). Survivors hold the new level whatever the cause (median r2 0.96 healthy, 0.98 short): the injury does not lower next-season points among those who play, it raises the chance they do not play, which is what the return-to-play literature finds (games missed explain 0.5% of the next-season PPG change among returners). The team's own decision is the best public rebound signal we have: a decliner kept by his team posts a 6-game season at or above the new level 0.44 of the time against 0.28 for one who moves (matched on the new PPG, gap +0.14, CI 0.02 to 0.24, floor 0.11, n=266), with median survivor r2 1.05 vs 0.92 and full recovery 0.22 vs 0.09. The kept, healthy decliner is the Yahoo rebound case (up 0.48, dropout 0.13, full recovery 0.25, n=63); the young healthy decliner who also held his snap share is the best of all (dropout 0.12, n=43; young and healthy overall: dropout 0.18, r2 1.06, recovery 0.24, n=91). The old kept decliner gets none of it (up 0.32 vs 0.26 moved, recovery 0.06).
Evidence for
- Our check
outcomes-veteran-decline-age.py: RB median ratio by age at season start: <=24 0.91, 25-27 0.88, 28 0.74, 29 0.71, 30+ 0.80 (survivors only; the 30+ survivors are selected). WR: 0.98, 0.83, 0.86, 0.67, 0.80. TE: 1.07, 0.86, 0.79, 0.72, 0.82. Dropout RB 0.09/0.17/0.22/0.40/0.13; WR 0.08/0.11/0.17/0.11/0.22; TE 0.06/0.06/0.04/0.05/0.27. - Our check
outcomes-veteran-decline-snap-drift.py: drift flag P(bad) 0.58 vs 0.31 base; the gap 0.27 is three times the shuffle floor. - Our check
outcomes-veteran-decline-efficiency-lead.py: role-loss gap 0.125 (floor 0.10); RB 0.667 vs 0.447, WR 0.556 vs 0.434, TE 0.452 vs 0.419; PPG-ratio gap +0.014 and dropout gap -0.02 (both null); coincident r 0.06 / 0.00 / -0.14. - Our check
outcomes-veteran-decline-efficiency-age.py(the null): y/y efficiency change by age at s. RB YPC: <=22 -0.36 (n=27), 23-24 +0.01 (64), 25-26 -0.11 (53), 27-28 -0.20 (23), 29-30 +0.09 (9). WR yards per target: <=22 -0.20 (68), 23-24 +0.08 (96), 25-26 -0.31 (80), 27-28 -0.43 (54), 29-30 -0.35 (27), 31+ -0.11 (21); WR points per touch -0.15 at 29-30 is the largest drop. TE yards per target: 25-26 -0.19 (38), 27-28 -0.13 (30), 29-30 -0.61 (16), 31+ -0.24 (12). Every 25+ bucket is negative, none clears its floor. - Our check
outcomes-decline-persistence.py(A-20): pooled matched composite gap -0.037 (CI -0.117 to +0.042, shuffle floor 0.071, n=738: 298 decliners, 440 holders); matched dropout gap +0.158; decliner dropout 0.34 / survivor median r2 0.97 / full recovery 0.13 vs holder 0.09 / 0.85 / 0.48. RB: dropout 0.38 vs 0.12, r2 1.02 vs 0.87, recovery 0.16 vs 0.45. WR: 0.37 vs 0.07, 0.93 vs 0.85, 0.10 vs 0.48 (matched dropout gap +0.20, the largest). TE: 0.24 vs 0.10, 1.06 vs 0.88, 0.17 vs 0.53. Old: dropout 0.44 vs 0.14, r2 0.92 vs 0.77, recovery 0.04 vs 0.31. Young: 0.27 vs 0.07, 1.05 vs 0.88, 0.20 vs 0.54. Snap fell 0.10+: dropout 0.45 vs 0.14, recovery 0.08. Snap held: 0.20 vs 0.08, r2 1.01 vs 0.86, recovery 0.20 vs 0.51. - Our check
outcomes-decline-fork-splits.py(A-48): healthy 14+ games dropout 0.24 / gone 0.05 / r2 0.96 / recovery 0.18 (n=141); short <=12 games 0.46 / 0.18 / 0.98 / 0.07 (n=125); short with no "Out" rows 0.49 / 0.19 / 1.05 / 0.05 (n=85); 2+ Out weeks 0.33 / 0.11 / 1.01 / 0.20 (n=64, the reported-and-returned injury is not the problem); healthy+snap held dropout 0.12 (n=43); round 1-2 0.31 vs 3+/UDFA 0.36; young 0.27 / r2 1.05 / recovery 0.20 (n=176); old +0-1 yrs 0.35 / 0.89 / 0.04 (n=68); old +2 yrs 0.56 / gone 0.33 / 0.93 / 0.04 (n=54); young+healthy 0.18 / 1.06 / 0.24 (n=91); old+short 0.51 (n=61). By position the healthy-short dropout gap is RB -0.06, WR -0.27, TE -0.36. - Our check
outcomes-decline-fork-team-retention.py(A-48): matched up gap +0.14 (floor 0.11); kept n=125 up 0.44 dropout 0.21 r2 1.05 recovery 0.22; moved n=141 up 0.28 dropout 0.30 r2 0.92 recovery 0.09. RB +0.10, WR +0.12, TE +0.13; young +0.16 (kept r2 1.10, recovery 0.27) vs old +0.02; healthy +0.20 (kept dropout 0.13, up 0.48) vs short +0.09. - Return-to-play study (PMC12804431, 2,523 time-loss injuries at QB/RB/WR/TE/FB, 2017-2022): next-season PPG change after an injury is -0.50 overall (CI -0.67 to -0.33), RB -0.70, WR -0.33, QB -1.95; games missed have no linear relation to the change (R^2 0.005). Consistent with our survivors holding the level whatever the games count; the study only sees players who returned, so it cannot see our dropout gap.
- Footballguys Injury Index (2017-2023, 2,000+ injuries): first-game-back production versus the healthy average and same-season re-injury rate by injury type; QB back injuries -39%, hamstring -18% with a high re-injury rate. Position tables are paywalled; the free part supports "the injury persists" as the dropout mechanism only in direction.
- Harstad, "We're Probably Thinking about Age the Wrong Way" (top 50 retired RBs and top 50 WRs of the last 30 years, EVoB): 11 of 50 RBs and 6 of 50 WRs declined in each of their last two relevant seasons (17%) against 25% if declines were random; players stay steady until a sudden drop. Our survivor-only consecutive-decline rate (0.33 vs 0.43) has the same sign.
- FantasyLife WR age study (2011-2023, 200+ routes, production over the best three-year stretch): after a sub-WR3 season at 30+, one elite WR (of those with 3+ WR1 finishes) returned to WR1 and one "good" WR rebounded at all; 71% of elite cliffs had gradual, turbulent or injury warnings first; PPR points per game beat every efficiency stat as the predictor for 30+ WRs. Consistent with our old-decliner full recovery of 0.04.
- Yahoo's 2026 rebound list (rules-of-thumb): rebounds are argued case by case from injury, QB or role change; it admits "some misses" and states no hit rate. It is the narrative the 0.13 full-recovery rate should temper.
- 4for4 production curves (47 RBs with 2+ top-12 seasons, 25 years): RB peak 26, decline from 29; WR peak 26-28, decline from 29, sharp at 32-33; TE peak 26, decline from 31, still 89% of baseline at 34 (n=10, survivorship acknowledged).
- PFF (2010-2016, 150+ attempt backs): decline begins at 29 for heavy-workload backs, 31 for low-volume backs; the "age 30" rule is "off by a year on either end".
- PFF WR age of decline (top-30 finishers since 1970, standard PPG): WR peak 26-27, then "a fairly steady downhill slope" 28-34 and a drop after 34; no split of efficiency from volume.
- Northwestern Sports Analytics (2007-2016, half-PPR, backs with 3+ top-20 finishes): 22-23 top-20 seasons at ages 27-28, 11 at 29, 8 at 30; over half the age-28 top-10 backs fall out of the top 10 the next year. Rank-based, no YPC by age.
- RotoWire delta-method curves: RB yards per carry falls about 1/8 yard per year from 22 to 30; RB points per touch fall 0.02 from 24 to 31 (about 5 points a season on 250 touches); WR yards per target peak at 25 and fall faster after (-0.054 points per target per year). Our per-bucket signs agree; our magnitudes are inside noise.
- Harstad (Footballguys): WR "death rate" 2-5% per year through 25, 17% at 30, 22-55% at 31-35; players stay productive until they suddenly are not.
- Harstad, "Be very wary of yards per carry": YPC needs about 1,978 carries to be half skill, half luck; a 5.00 YPC half-season projects to 4.37 in the
…(truncated)
Claim
Public season-long projections are all weak and all similar. Across 2014-2025 (11 sources, Fantasy Football Analytics, FFA) they explain 14-26% of within-position variance in season points (RB most, QB least), the top sources at WR sit within 2.1 MAE of each other, and an equal-weight average of sources beats any single source in 69% of head-to-head comparisons. Which single source is best does not persist: CBS went from 4th over 12 seasons to 9th over the last three, and FFA concluded "source accuracy simply does not persist reliably enough" to justify weighting. ESPN's season projections were the worst QB source over the last three seasons (MAE 86.7 vs 61-71 for the leaders). All sources exaggerate the spread between players (calibration slope of actual on projected: QB 0.67, TE 0.72, RB 0.79, WR 0.85) and are optimistic on average (mean error +21.6 season points; QB +46.5 recently), which is mostly games missed, not per-game error. The numbers are FFA's, not ours: no projection or ranking archive is on disk, so this claim is unquantified until WO-0019 lands.
Evidence for
- FFA, 2014-2025, 11 sources, season-long MAE: equal-weight average beat individual sources in 69% of pairwise comparisons, mean rank 4.2 of 11; equal weight (47.4 MAE) matched accuracy-weighted (47.8).
- FFA: R2 within position 0.14-0.26; calibration slopes 0.67-0.85; ME +21.6 pts.
- FFA: CBS fell from 4th (all seasons) to 9th (last three); ESPN worst QB source recently.
- FantasyPros: 212 experts scored on 2025 draft rankings, leaderboard turns over year to year; the only repeat winner across three seasons is the IDP ranker (a thin field). In-season, one Yahoo ranker has nine top-10 finishes, so in-season ranking skill persists more than draft-ranking skill.
- Consistent with theory/usage-stickiness (usage r ~0.70 y/y) and outcomes/scoring-distribution-shape (next-season PPG p10/p90 0.54/1.21): a rank order that explains 20% of variance is what those imply.
Evidence against
- FFA's own site sells the aggregate it finds best; the 69% is the study author's product.
- FantasyPros reports no ECR-vs-individual number in the 2025 write-up, so "ECR beats most experts" is inferred from the FFA aggregation result, not observed on the ranking side.
- FFA measures season points, so optimism and low R2 fold injuries in; per-game accuracy is likely higher and is what the weekly optimizer needs (A-30).
- Weekly projections were not in the 12-season study; the weekly calibration slope is unknown.
How we use it
- Do not weight a source by last season's accuracy rank. Source accuracy does not persist; a source inherits its school's accuracy (schools.md rule 1) and nothing more. This holds for individual draft rankers; give in-season rankers with a multi-year record at most +0.05.
- Average when two sources exist, equal weight. ESPN's projection is one voice; where the harness holds a second projection or ECR for the same player-week, use the mean.
- Shrink projected gaps within a position when a decision rides on the gap between two players: multiply the projected season-point gap by the calibration slope (QB 0.67, TE 0.72, RB 0.79, WR 0.85). Weekly use at half that shrink (slope 0.85-0.92) until A-30 measures ESPN's weekly calibration; a gap under 1.0 weekly point is inside the noise (scoring-distribution-shape).
- Season projections are optimistic by about 20 points, mostly games missed. Do not subtract it per game; price games missed from outcomes/player-range-of-outcomes rule 5 and use the projection as a per-game rate.
- Consensus rank is a weak prior, not a verdict: ~20% of variance explained means usage evidence (in-season/usage-window) should dominate the rank by the window's game count, and a top-24 preseason rank is not by itself a reason to hold a player whose usage has fallen (veteran-decline rule 2).
- Total weight of this claim in any decision stays under 0.6 (unquantified cap) until the check passes.
Open questions
- Does ADP beat ECR as a preseason prior? No public source scores both on the same pool; needs an ADP archive (A-56).
- Weekly calibration slope of ESPN projections (A-30) and whether the season-long optimism appears weekly at all.
- Is the QB optimism the rushing component (projected rushing TDs regress, theory/td-regression rule 5)?
Claim
Fantasy football analysis is produced by identifiable schools of thought that make different kinds of errors. Weighting a source by its school's demonstrated accuracy (PG-3) beats trusting any one school. Every school starts at accuracy 0.50, a prior, not a measurement. REFLECT updates it in 0.02 steps as decision outcomes are scored (spec 03 rule 7).
The schools
consensus-adp (accuracy 0.50, prior)
Crowd wisdom: Expert Consensus Rankings (ECR), ADP across platforms, aggregated rankings. Outlets: FantasyPros ECR, platform ADP (ESPN, Sleeper, Yahoo, Underdog). Gets right: averaging many rankers reduces individual error; ECR is hard to beat on median outcome. Gets wrong: by construction it cannot find sleepers; slow to move on news; herds late in draft season.
analytics-projection (accuracy 0.50, prior)
Model-driven point projections and value-based drafting math (VORP, VOLS, VONA). Outlets: Fantasy Football Analytics, 4for4, Establish The Run, FantasyPros projection aggregate, PFF. Gets right: replacement-level reasoning; positional scarcity; season-long expectation. Gets wrong: point estimates hide distributions; models inherit their inputs' biases; overfit to last season.
film-and-usage (accuracy 0.50, prior)
Snap share, route participation, target share, air yards, red-zone touches, offensive line and scheme. Outlets: PlayerProfiler, 4for4 usage reports, Fantasy Points data, Sumer Sports, nflverse-derived work. Gets right: opportunity metrics are the stickiest predictors (target share ~0.70 y/y correlation). Gets wrong: usage does not price in coaching changes or efficiency collapses; small-sample overreaction in-season.
contrarian-upside (accuracy 0.50, prior)
Zero RB, late-round QB, ceiling chasing, anti-consensus structural bets. Outlets: Zero RB writers (Shawn Siegele lineage), late-round QB writers (JJ Zachariason lineage), various. Gets right: leverage in large fields; exploits RB fragility; avoids paying for scarce positions at inflated prices. Gets wrong: strategies that work in best-ball tournaments transfer poorly to 10-12 team H2H; often stops being contrarian once popular.
market-efficiency (accuracy 0.50, prior)
ADP arbitrage and structural findings from millions of best-ball drafts (advance rates, roster construction). Outlets: Underdog Network research, Establish The Run best-ball, 4for4 "How Winners Draft" series. Gets right: large-N empirical structure results (e.g., Round 1-2 RB advance rates); prices in the market's blind spots. Gets wrong: best-ball scoring (no lineup setting, no waivers, 18-man rosters) differs from managed H2H; results shift year to year with the player pool.
rules-of-thumb (accuracy 0.50, prior)
Heuristics: positional scarcity, handcuffing, streaming DST/K/TE, "play your studs", tiers over ranks. Outlets: ESPN Fantasy staff (Clay, Karabell, Cockcroft), CBS (Richard, White), Footballguys, Fantasy Footballers. Gets right: robust, low-variance procedures; excellent for time-limited decisions like the draft clock. Gets wrong: heuristics lag rule/format changes; rarely quantified; overweight recency and narrative.
betting-markets (accuracy 0.50, prior)
Vegas win totals, implied team totals, player props as projection inputs. Outlets: sportsbook lines, Sharp Football implied totals, FantasyLabs Vegas dashboard, Draft Sharks. Gets right: sharpest available estimate of team scoring environment and game script; updates fast on news. Gets wrong: props are shaded for public action; thin markets for depth players; little to say about touch distribution within a team.
Evidence for
- Aggregating sources beats any single source (FFA 2014-2025: an equal-weight average of 11 projection sources won 69% of head-to-head comparisons), which is the premise behind weighting by school rather than by one outlet. Detail in sources/projection-accuracy.md.
- Best-ball structural results (market-efficiency) and Zero RB (contrarian-upside) currently disagree, so schools are distinguishable.
Evidence against
- Schools overlap heavily; many analysts straddle two or three. Attribution of a source to one school is a judgment.
- One season of one league produces few scored outcomes; accuracy updates will be noisy.
- Individual source accuracy does not persist year to year (FFA: CBS 4th over 12 seasons, 9th over the last three; FantasyPros' 212-expert draft leaderboard turns over yearly). If sources inside a school are that unstable, a school-level score measured over one season may be noise too; hence the 0.02 steps and the [0.25, 0.85] bounds. The earlier bullet "reliable rankers score higher" was an assumption, not a measurement, and is withdrawn.
How we use it
- Every source recorded in a claim file gets a
schooltag and inherits that school's accuracy as its defaultweight. - When schools disagree on a decision-relevant question, ANALYZE records both positions, weights by accuracy, and files the disagreement in
_contested.mdif the weighted gap is under 0.10. - Never let one school exceed 0.40 of total weight on a single decision (PG-3).
- REFLECT adjusts school accuracy by +/-0.02 per scored decision, bounded [0.25, 0.85].
- A source never carries a weight above its school's accuracy on the strength of its own past accuracy rank (sources/projection-accuracy.md rule 1); the one exception is an in-season ranker with a multi-year top-10 record, at most +0.05.
- Where two projection sources cover the same player-week, average them at equal weight before applying any school weighting; the average is the analytics-projection input, not either source.
Open questions
- Should accuracy be tracked per school per position (a school may be good on WR, bad on RB)?
- Should betting-markets get a higher prior given sharp-money incentives? (Held at 0.50 until scored.)
Claim
When a regular pass catcher (trailing target share 0.15 or more) misses a game, his roughly 21% of team targets does not land on the next man up. The top remaining pass catcher gains about +0.03 share over what he would have had anyway (about +0.7 targets and +1.5 PPR points per game); the next two gain +0.02 to +0.03 each. The rest is spread across the receiver room and bodies who had no prior role. The vacated share stays inside the position: a missing WR's targets go to WRs (+0.12 to the WR group, +0.02 to TEs, +0.01 to RBs), even when the team's top remaining option is a TE. Team pass volume barely moves (-2%). The shift appears in the first game and holds the same size in the second. The backfield is the opposite case: when a lead RB (trailing carry share 0.40 or more, on average 0.55) misses a game, the top remaining RB takes about half of what was vacated, +0.27 carry share and +0.25 snap share over control, worth +1.5 targets and about +6 PPR per game; the second remaining back takes +0.13 share (+2.3 PPR), and backs with no prior role about +0.08. Almost nothing leaves the position (the RB fraction of team carries dips 0.02) and the team runs about 5% less. The RB shift also holds while the absence lasts, easing only slightly by the third game (+0.24 then +0.21). An in-season change of starting QB, by contrast, moves the established pass catchers' target shares and points no more than an ordinary week does: a measured null, pooled over upgrades and downgrades. Coaching or play-caller changes are not measured yet (see theory/coaching-change-share-reset). The promoted back also inherits the goal-line role in full: his share of team goal-line carries rises +0.33 over control (versus +0.27 in overall carry share), to 0.58 of team goal-line carries, the same level a healthy lead back holds (0.56); the team does not move short-yardage work to a bigger body. Committee partners and pure backups inherit it alike, and the team's run rate does not change the size. Measured on nflverse 2019-2025 regular season. Companion checks: analysis/checks/theory-ball-share-shifts-qb-change.py, analysis/checks/theory-ball-share-shifts-rb1.py and analysis/checks/theory-ball-share-shifts-goal-line.py.
Evidence for
Teammate absence (single absence of a >=0.15-share WR/TE/RB who was active the prior three games and did not resurface elsewhere; n 348 events, 2462 control team-games; changes are versus each player's trailing-4-game share, control = the same statistic in weeks with no such absence):
| who | event | control | excess |
|---|---|---|---|
| top remaining catcher, share | +0.006 (CI -0.004..+0.016) | -0.021 | +0.027 (+0.73 targets/g, +1.55 PPR/g) |
| second remaining | +0.011 | -0.009 | +0.020 |
| third remaining | +0.020 | -0.008 | +0.028 |
| team targets, ratio to trailing 4 | 0.989 | 1.011 | -2% |
Where the vacated share goes (position-group share change net of control): absent WR (n 263, vacated 0.213): WR +0.124, TE +0.017, RB +0.014. Absent TE (n 67, vacated 0.207): WR +0.082, TE +0.062, RB +0.023. Absent RB with a receiving role (n 18): WR +0.068, RB +0.062. When the top remaining option is a WR his excess is +0.032 (n 255); a TE +0.018 (n 67); an RB +0.001 (n 26). The remaining 0.05-0.06 of vacated share goes to players with no trailing baseline (a WR4 or practice-squad call-up).
Persistence: in the second consecutive absence game the top remaining catcher's change is +0.014 versus +0.013 in the first (n 160), so the shift is immediate and stable in the mean; the exploration pass found the individual first-game reaction only weakly predicts the second (r about 0.30).
Lead-RB absence (single absence of a >=0.40-carry-share RB active the prior three games, not traded; n 172 events, 2669 control team-games; changes versus each player's trailing-4-game baseline, net of the same statistic in control weeks):
| who | carry share | snap share | targets/g | PPR/g |
|---|---|---|---|---|
| top remaining RB | +0.265 (raw +0.259, CI +0.226..+0.292; control -0.006) | +0.250 | +1.52 | +5.92 |
| second remaining RB | +0.127 | +0.152 | +0.76 | +2.25 |
| RBs with a baseline, as a group | +0.366 | |||
| RBs with no baseline (call-ups) | +0.080 |
Mean vacated carry share 0.547. RB fraction of team carries 0.799 in absence games vs 0.815 in control; team carries ratio to trailing four 0.977 vs 1.027. Persistence: the top remaining RB's carry-share change is +0.243 in the second absence game (n 88) and +0.209 in the third (n 45), so most of the first-game promotion survives, with a slow drift back as the staff spreads the work.
Goal-line inheritance (same lead-RB events, restricted to games where the team had at least one carry from inside the 5 and the promoted back had a trailing goal-line baseline; n 81 events, 2669 control team-games; analysis/checks/theory-ball-share-shifts-goal-line.py, 2026-09-11):
| statistic | promoted back | control (healthy lead back) |
|---|---|---|
| goal-line share change vs own trailing 4 | +0.335 (CI +0.235..+0.434) | +0.001 (excess +0.334) |
| red-zone carry share change | excess +0.277 | |
| overall carry share change | excess +0.265 (gl/carry ratio 1.26) | |
| goal-line share level in the game | 0.578 | 0.560 |
| rushing TD per game | 0.40 | 0.48 |
| PPR per game | 12.5 | 13.9 |
The absent back's own trailing goal-line share was 0.560, so the promoted back lands exactly where the lead back was. Split by prior role (excess over the pooled control): committee partner +0.327 (n 54) vs pure backup +0.348 (n 27); by team run rate (trailing-4 rush fraction 0.45+): +0.321 (n 33) vs +0.343 (n 48) below it. Neither split moves the answer. Team goal-line carries per game fall from 1.39 to 1.22 (-12%) in absence games, which is why the promoted back's TD rate sits a little under a healthy lead back's despite the same share. No school publishes a goal-line inheritance rate: the handcuff pieces (Bonini, Yahoo/Athlon 2026; FantasyPros committee review, Aug 2026) say a backup "should" inherit goal-line work or that committees fragment it, and ESPN's insurance-RB list (Moody, Aug 2025) gives production examples only.
In-season QB change (new starter after the old one started the prior three games, and the new one kept the job for the next two; n 76 switches, 195 top-3 pass-catcher rows; control 1487 team-games): share change -0.009 vs control -0.012 (gap +0.003); PPR-per-game ratio after/before 0.937 vs 0.946; team attempts +0.06 vs +0.01. By rank: the top catcher +0.007 share and a better PPR ratio than control (0.970 vs 0.913), third catcher -0.004 and 0.862 vs 0.982. TE +0.010, WR +0.001, RB +0.003.
Public sources agree on direction: Quant Lab finds a WR2 promoted to WR1 gains only +1.3 targets, +0.9 receptions, +1.5 PPR per game, far less than an RB2 promoted to RB1, and in its backfield-split series (Di Bon, Oct 2025) puts the RB snap share that "provides sufficient opportunity to thrive" at about 60% and the elite level above 80%; our promoted back reaches roughly 0.25 above a backup's share, which lands a true backup near that 60% line and a committee partner above it. Yahoo (Smyth) treats vacated targets as a map of where opportunity may surface and warns they are split among several players; PFF's 2025 review (Conway) found backups sometimes helped the WR1 (Olave, Jefferson) and left an RB1 flat (McCaffrey), consistent with no systematic penalty. The handcuff pieces (Yahoo/Athlon, Bonini, Apr 2026; PFF, Bodiford, Aug 2026) are rules-of-thumb with no numbers: they say a backup only pays when he would inherit most of the carries, routes and goal-line work rather than split them, which is the committee caveat our second-RB row quantifies.
How long the RB bump lasts and what happens on the return (A-33, checks under waivers/): in absences of 4+ games the top remaining RB's carry share runs 0.52 / 0.46 / 0.41 / 0.45 over games 1-4 (n 35); the decline is all in pure backups (0.45 to 0.33), committee partners hold (0.58 to 0.56). In the lead back's first game back the promoted back is at 0.22 share against 0.25 before the absence and 0.51 during it (n 105), and the returning lead back is at 0.87 x his own baseline share; the vacated share therefore reverts in one game, not over several.
Evidence against
- The QB null pools upgrades and downgrades and keeps only switches where the new QB lasted three games, so a bad backup who was benched after two starts is excluded; direction-specific effects may exist.
- Events require a trailing baseline of 2+ active games among the last 4, so a player promoted from inactive (the true next man up on a thin depth chart) is not in the top-3 and his gain is unmeasured.
- Multi-absence weeks (36) are excluded; the response when two regulars are out may not be additive.
- Absence includes healthy scratches, suspensions and benchings, not only injuries.
- Public sources on vacated targets are preseason roster-math pieces with no historical hit rate.
- The RB event needs a lead back at 0.40+ carry share, so a true 50/50 committee losing one half is not covered; the response there is probably smaller. Carry share is of all team carries including QB scrambles, so a mobile QB deflates every RB's share equally. The third-game persistence rests on 45 events and only counts absences that lasted three games (longer injuries), so it is not a clean decay.
- The goal-line check keeps only 81 of the 172 absence games (the team must have had a goal-line carry that week and the promoted back a trailing baseline with one), so it is the mean over teams that reached the 5; a team that stalls outside it contributes nothing. Goal-line carries per game are about 1.3, so the share is a coarse statistic and the CI (+0.24..+0.43) is wide. The run-rate split uses a 0.45 rush fraction cut on 33 vs 48 events; a finer cut would be noise.
- The 4+ game hold path rests on 35 absences (17 pure backups); the per-game decay for backups is a direction, not a slope to be trusted to the second decimal.
How we use it
- When a >=0.15-share teammate is ruled out, raise the top remaining WR's projected target share by +0.03 (about +0.7 targets, +1.5 PPR), the next two pass catchers by +0.02 each. Never assign the whole vacated share to one player; a projection that adds more than +0.08 to anyone needs specific evidence.
- Keep the bump inside the position. A missing WR does not raise the TE or the RB (+0.02 at most for a TE who is the top remaining option, nothing for an RB). A missing TE raises the WRs as a group and the TE2, who is usually a low-baseline player and a weak start.
- Do not lower the team's pass-volume projection for the absence (-2% is inside noise).
- Apply the bump from the first game and keep it flat for as long as the absence lasts; do not scale it up or down from the first game's box score (in-season/usage-window still governs the 4-game window).
- An in-season QB change is not a reason to move an established pass catcher's share or projection, and never a reason to bench a WR1 or TE1. Any direction-specific adjustment (rookie behind a bad line, proven veteran replacing a struggling starter) is a judgment capped at +/-10% of PPR, stated in the decision record as such.
- Coaching and play-caller changes: use theory/coaching-change-share-reset (unquantified) until WO-0014.
- When a lead RB (0.40+ carry share) is ruled out, project the top remaining RB at his trailing baseline plus 0.27 carry share, plus 0.25 snap share, plus 1.5 targets (about +6 PPR); the second remaining back at plus 0.13 share (about +2 PPR). Do not give either back the whole vacated share. Keep the bump for every game the lead back misses. Decay depends on who was promoted: a committee partner (baseline share 0.20+) holds his game-1 level through game 4 (carry share 0.58 / 0.49 / 0.48 / 0.56 in 4+ game absences, n 18), so keep his bump flat; a pure backup slides about 0.0
…(truncated)
Claim
A new head coach or play-caller probably erodes some of a returning player's prior-season share, but less than a change of team does, and the size is unmeasured. The sources that say coaching changes are "one of the most underpriced edges" are preseason narrative pieces with no historical hit rate; the one data point we have is that a player who changes teams keeps r 0.72 of target-share persistence instead of 0.81 (theory/usage-stickiness), and a staff change is a weaker version of the same disruption. For 2026 the pieces above count 10 new head coaches and about 20 new offensive coordinators, so the question touches a third of the league every year.
Evidence for
- theory/usage-stickiness: same-team WR target share r 0.81 vs 0.72 after a team change, RB carry share 0.77 vs 0.58. A staff change keeps the QB and the depth chart, so it should sit between those.
- Draft Sharks and Fantasy Life (search snippets, not fetched in full) argue usage, run/pass balance, tempo and red-zone tendencies move with the play-caller even when the roster is intact.
Evidence against
- No source quantifies it, and the schools that push it (rules-of-thumb) have a preseason incentive to find narrative movers.
- Scheme carry-over is real: many new coordinators come from the same tree as the old one.
- Ball-share shifts after an in-season QB change are a measured null (theory/ball-share-shifts), which argues that share is a role property and may survive a new voice on the headset too.
How we use it
- Until WO-0014 lands, when the play-caller changed and the player did not, shrink prior-season share toward the positional mean by 10% (half of the 25% team-change discount in usage-stickiness rule 3). Do not stack it with the team-change discount; a player who changed teams gets that one only.
- Give a coaching-change narrative no more than 0.10 of the weight in any decision that cites it, and record the claim's unquantified status in the decision record.
- Re-open when the coaching table exists; the falsifier converts this to a measured number or a null.
Open questions
- Head coach vs offensive coordinator vs play-caller: which one matters? The table must name the caller.
- Does the effect concentrate in the first four games (a new offense installing) and fade, or hold all season?
Claim
Same-team QB and WR scores are strongly positively correlated (~+0.86 at the team-total level); QB-RB (~+0.40) and QB-TE (~+0.36) less so. Stacking does not change season-long expected points but raises weekly variance: more boom weeks and more bust weeks. In a managed H2H league that is a playoff-week tool, not a draft-day goal: a stack is worth pursuing only at neutral cost and is worth actively building only for weeks 15-17.
Evidence for
- Four years of data: stacked QB-WR pairs had the same season median (~443 PPR points) as unstacked pairs, but boom rate 15.3% vs 12.0% (market-efficiency).
- Pass-heavy offenses produce much stronger QB-WR linkage than run-heavy or rushing-QB offenses; the correlation is conditional on scheme.
- Best-ball tournaments reward stacking because a single spike week advances a team; the same logic applies to single-elimination fantasy playoffs.
Evidence against
- Managed leagues score cumulative wins over 14 weeks; extra variance in the regular season is neutral to slightly negative for a favorite and positive only for an underdog.
- Reaching for a stack partner costs VORP; the variance benefit rarely exceeds the value lost.
- Negative correlations (our WR vs opponent's DST) are small and mostly irrelevant with weekly streaming.
How we use it
- Draft: never reach more than one tier for a stack partner. If two candidates are within the same tier, prefer the one stacked with our QB only if that offense projects pass-heavy (implied pass rate above league average).
- Regular season, as a favorite (our win probability > 60%): prefer uncorrelated starters when medians are equal. As an underdog (< 40%): prefer correlated starters.
- Playoffs (weeks 15-17): build the lineup to maximize p90 (team-range-of-outcomes.md rule 4); same-team QB-WR is the first lever, QB-TE second.
- Avoid negative correlation inside our own lineup: do not start our DST against our own QB's team unless projection gap exceeds 3 points.
- Record whether a lineup was stacked in the decision record for REFLECT scoring.
Open questions
- Does this league use a bracket with reseeding or a two-week championship? Two-week finals reduce the value of variance and weaken rule 3.
Claim
Fantasy points are opportunity times efficiency times luck. Opportunity (target share, air-yards share, weighted opportunity, snap share) is sticky year to year: measured per-game r of 0.78-0.81 for WR/TE target share, WOPR and air-yards share, 0.73 for RB carry share, about 0.70 for snap share (2019-2025; detail and horizons in theory/usage-stickiness). Touchdowns are only half as sticky (per-game TD r 0.44-0.51 for RB/WR/TE, 0.69 for QB) and TD totals regress toward the rate implied by a player's opportunity (expected TDs). Weekly variance is highest at TE and lowest at QB. Team scoring environment, best estimated by Vegas implied team totals, scales every player on the team roughly proportionally.
Evidence for
- Measured (our data, 8+ active games in both seasons): WR target share y/y r 0.79 (n 632), TE 0.78 (n 333), RB carry share 0.73 (n 375), per-game TDs 0.44-0.51 for RB/WR/TE. Shuffle floor under 0.06. The check's narrower population (WRs with 50+ targets both seasons) gives 0.67 (n 347), passing its 0.60 floor; 2026-09-10 run passed. Check:
analysis/checks/theory-usage-stickiness-season.py. - Target share, air-yards share, and weighted opportunity rating all exceed 0.70 y/y correlation with themselves and with fantasy points; opportunity metrics out-predict efficiency metrics (film-and-usage).
- TD regression reports from several outlets independently compute expected TDs from location-weighted opportunity and find high-TD outliers regress (analytics-projection, betting-markets).
- Measured (our pbp, 2019-2025, consecutive-season pairs): TD-over-expected y/y r 0.04-0.13 at RB/WR/TE, within-season split-half r 0.05; top-decile over-performers regress in 88-96% of cases and keep about 10% of the excess (RB 0, WR 0.12, TE 0.18). QBs keep 0.35 of passing and 0.31 of rushing excess (y/y r 0.22-0.25, top quintile regresses 77%). The 2026-09-10 figures (r 0.28, keep 30%) came from a check that paired non-consecutive seasons; repaired 2026-09-11. Detail and rules in theory/td-regression. Checks:
analysis/checks/theory-td-regression.py,analysis/checks/theory-td-regression-qb-season.py(both passed 2026-09-11). - Implied team totals are used by rankers as the environment prior; a 10% higher implied total is treated as ~10% more fantasy points for the team's players (betting-markets). Our measurement says about 2% (elasticity 0.16 against a player's own recent form; theory/situational-factors), because the line explains only part of actual team scoring and recent form already carries the team's usual environment.
- Weekly SD by position: TE > RB ~ WR > QB, attributed to touch volume.
Evidence against
- Stickiness measured on survivors (players who kept their role); role changes (trade, new OC, rookie arrival) break it, and those are exactly the cases where projection matters most.
- Some players sustain TD rates above expectation for years (red-zone specialists, elite QBs); regression is a population statement.
- Home/away effects on fantasy scoring are small and inconsistently documented; treat as noise.
How we use it
- Projection inputs are ranked: opportunity metrics first, efficiency second, last season's TDs last. When two sources disagree, prefer the one that reasons from opportunity.
- TD regression adjustment: projected TDs = expected TDs from projected opportunity + 0.1 x (prior-season TDs - prior-season expected TDs) for RB/WR/TE and 0.35 x for QB passing (0.3 for QB rushing), at every magnitude and in both directions. The earlier "halve the excess" and "keep 30%" rules both under-regressed for RB/WR/TE (measured carry-over 0 to 0.18); see theory/td-regression rules 1-2. In season, a RB/WR/TE's TD rate over expected carries no information (split-half r 0.05): update only on field-position usage. A QB's carries a quarter weight (split-half r 0.12; td-regression rule 5).
- Environment: the full proportional multiplier (implied team total / team season average) over-corrects. Measured 2019-2025 (theory/situational-factors,
analysis/checks/theory-situational-factors-implied-total.py): a starter's points relative to his own trailing 3-game average move with elasticity about 0.16 (CI 0.08-0.24; TE 0.30, WR 0.10) to the implied total relative to his team's season norm. Favorites score more at every position (+0.4% per point favored with the game total held fixed, 2019-2025 starters,analysis/checks/theory-situational-factors-spread-direction.py); most of that is the implied total arriving through the spread, and the RB/WR game-script residual is about +0.3% per point (situational-factors rule 7). Garbage time is 10% of a starter's points and the only slice that leans to underdogs; it does not rescue underdog WRs. On a form-based baseline use multiplier = 1 + 0.2 x (implied total / team season-mean implied total - 1), capped [0.92, 1.08]. On top of an ESPN projection apply nothing until A-30 shows whether ESPN already prices the line. - Variance by slot: when picking between equal-median players for a starting slot, prefer the lower-variance position profile (QB, then WR/RB, then TE); for bench upside, invert.
- Ignore home/away as a standalone factor; it is already inside the Vegas line. Weather, dome, and rest adjustments live in theory/situational-factors (wind is the one that matters; Thursday is a measured null).
- Re-verify the stickiness numbers each August with nflverse data (work order; sets
last_verified).
Open questions
- Does this league's scoring (PPR vs half vs standard, TE premium, bonuses) change which opportunity metric is most predictive? Half-PPR shifts weight toward yards; check once settings are known.
Claim
A quarterback's season interception rate carries almost no information about his next one. Over 2019-2025 (QBs with 200+ non-sack attempts in consecutive seasons, n 155 pairs) the year-over-year r of INT per attempt is 0.09 and the next season keeps 0.09 of the prior excess over the league rate (2.26% of attempts). Every top-decile INT-rate season (3.5% of attempts) was followed by a lower rate (16 of 16, to 2.2%), and 88% of bottom-decile seasons (1.0%) were followed by a higher one (to 2.0%). Within a season the odd/even-week r of INT rate is 0.05, so a hot or cold first half says nothing about the second. Raw interception counts do not persist at all (r 0.00): next season's INTs are predicted by attempts (r 0.19), not by last season's INTs. What does persist is the depth mix: the expected INT rate implied by a QB's air-yards distribution (0.7% behind the line, 6.3% on throws of 20+ air yards) has y/y r 0.39 and split-half r 0.55, but it moves next season's INT rate by only r 0.11, because the mix's spread across QBs is a tenth of the noise spread. A two-season attempt-weighted rate does better than one season (r 0.20 vs 0.02 on the same 98 pairs), which is the shape of a small skill term buried under one-season sampling error (Burke: up to 81% of the variance is sampling error; PFF: the throw-quality rate is stable, its conversion to interceptions is not, r 0.12). Completion rate is not a better predictor than the INT rate itself (r -0.10 vs 0.09); both are near zero.
Evidence for
- Measured,
analysis/checks/theory-qb-interception-regression.py, run 2026-09-11 on 2019-2025 pbp: top decile 100% lower (n 16, CI floor 0.5 = fully persistent skill), bottom decile 88% higher (n 16); y/y r 0.09, slope 0.09; r(INTs, next INTs) 0.00, r(att, next INTs) 0.19. With the 250-attempt cut (n 142) r falls to 0.05 and the top decile to 93%. - The Footballguys bands replicate on 2019-2025 with their own definition (INT per incompletion, league 6.1%): 9.5%+ fell 9 of 9 (10.1% -> 6.2%); under 4% rose 18 of 20 (3.1% -> 6.3%). Their 2005-2016 sample (261 QBs) had 28 of 28 and 26 of 28.
- PFF (2016-2025, 209 pairs, charted throws): 49% of turnover-worthy throws become interceptions, 0.8% of clean throws do, and the conversion rate's y/y r is 0.12; roughly a third of all interceptions are on throws PFF does not grade as turnover-worthy.
- Burke (2002-2009): the observed SD of season INT rate (0.94%) is barely above what a coin with the league rate produces over the same attempts (0.85%); at most 19% of the variance is ability. Our 2019-2025 season SD is 0.70% at 200+ attempts, the same picture.
- Depth mix is real and sticky (y/y r 0.39, split-half 0.55) and is the one structural input: a deep-throwing QB carries a higher INT floor. It is small (r 0.11 with next season) because xINT rate ranges about 1.9%-2.7% across QBs while realized rates range 1.0%-3.8%.
Evidence against
- Skill is not zero: the two-season prior lifts r from 0.02 to 0.20, Burke's ability share is up to 19%, and PFF's turnover-worthy-throw rate (which we cannot compute from nflverse) is the stable part. Career-long outliers exist (Rodgers 1.4% over a career vs 2.3% league, Footballguys). The rule below prices skill through the two-season rate and the depth mix, not through last season's INT count.
- The fantasy stakes are modest. At -2 per INT (ESPN default), the top-decile-to-mean regression is about 6-7 fewer INTs over 550 attempts, 12-14 points a season, under one point a game; the season SD of INTs (3.6) is 7 points, 0.4 a game. It moves a QB across a tier boundary only when two QBs are already close.
- n is 155 pairs and 16 per decile on 2019-2025. The two-season-prior line rests on 98 pairs. The Footballguys and PFF samples (236-261 and 209 pairs, different eras) agree in direction and size, which is why confidence is 0.80 rather than 0.70.
How we use it
- Never project a QB's INTs from his last-season INT rate. Project INTs as attempts x the league rate (2.3% per non-sack attempt, refreshed yearly from the check notes), then adjust: half of the two-season attempt-weighted deviation from league when 400+ prior attempts exist (a QB at 1.5% over two seasons is priced at about 1.9%, not 1.5%), and the depth-mix xINT rate in full (it is a rate the QB's role sets, not luck).
- Trailing-window INT rate gets zero weight in-season. The usage-window trailing average (in-season/usage-window.md) should be built on yards, TDs and attempts; an INT-heavy or INT-free 3-6 game stretch is not form (split-half r 0.05). This mirrors the TD rule (theory/td-regression.md rule 4) with a smaller keep (0.09 vs 0.35 for passing TDoE).
- Prior-season INT luck is a small QB-tier tiebreak, not a ranking input. A QB whose prior INT count sat 5+ below attempts x league rate (PFF's 2026 list: Stafford, Ward, Dart, Prescott, Mayfield) loses about 0.5-0.7 points per game of expectation at -2 per INT; a QB 5+ above gains it. Apply only between QBs within 1.0 projected points per game of each other.
- When a projection source's QB INT line sits under 1.5% or over 3.2% of attempts, shrink it to the rule-1 number before scoring the projection; a source that carries last season's INT rate forward is carrying noise (projection-accuracy rule 3 shrink applies).
- Where the league's INT penalty is larger than -2 (check the scoring items before the season), scale rules 1-3 in proportion; where INTs are not penalized, ignore this claim.
Open questions
- PFF's turnover-worthy-throw rate is the stable component and would be the right prior, but it needs charted data (docs/DATA-OPTIONS.md lists PFF as paid). Is a pressure-adjusted or CPOE-adjusted rate from nflverse a usable public proxy? (qb_hit share had r 0.04 with next INT rate here, so hits alone are not it.)
- Does the depth-mix xINT rate sharpen when built from the richer buckets of the TD work (down, distance, score state)? The TD result (theory/td-regression.md) says no for TDs; untested here.
- Rookie and second-year QBs: PFF's 2026 list has two second-year QBs (Ward, Dart) below expected INTs. Does the regression run harder or softer for young QBs, and does a young QB's INT rate fall with experience independent of luck (an age curve, not a regression)? Needs the outcomes/breakout-rookie-carryover.md QB pairs (A-49 has the same data).
Claim
A passer's quality moves his receivers' points a little, and only the quality he actually delivers that season: about 0.25 PPR/g per 0.1 EPA per dropback for WRs (one standard deviation of starting-QB quality is 0.14 EPA/dropback, so a full-SD better passer is worth about 0.35 PPR/g, 3-4% of a starting receiver's line). The quality a new passer showed the season before predicts none of it: among same-team WRs whose primary passer changed, next-season PPR/g holding prior PPR/g fixed moves 0.03 PPR/g per 0.1 EPA/dropback of ex-ante difference (CI -0.30 to +0.38, shuffle floor 0.34, r 0.02, n 183), and the same-QB pairs show the same realized-quality slope (0.20) as the changed pairs (0.25). So the small points dent after a QB change is not a quality effect that can be priced from the new passer's history, and QB quality is an offense-level factor that the QB's own projection already carries. Measured on nflverse 2019-2025, regular season, 8+ active games both seasons, team = modal team, primary QB = most pass attempts on that team that season; a passer with under 150 prior-season dropbacks (rookie, backup) is set to the 20th-percentile replacement level (-0.09 EPA/dropback).
Evidence for
Same-team consecutive-season pairs, PPR/g in s+1 regressed on PPR/g in s plus a QB quality term, slopes in PPR/g per 0.1 EPA/dropback (bootstrap 95% CI; shuffle floor = 95th percentile of |slope| with the quality term permuted):
| position | n changed / same | ex-ante new-minus-old (changed) | realized new-minus-old (changed) | realized, same QB (his own y/y change) |
|---|---|---|---|---|
| WR | 183 / 294 | +0.03 (-0.30..+0.38), floor 0.34 | +0.25 (-0.04..+0.62) | +0.20 (-0.07..+0.48) |
| TE | 109 / 159 | +0.21 (-0.05..+0.46), floor 0.29 | +0.16 (-0.11..+0.41) | +0.32 (+0.01..+0.65) |
| RB | 114 / 168 | +0.03 (-0.50..+0.63), floor 0.45 | +0.34 (-0.18..+0.84) | +0.33 (-0.10..+0.75) |
- Restricting to changed pairs whose new passer had 150+ dropbacks the season before (a veteran whose quality is actually known, n 87 WR) does not rescue the ex-ante slope: +0.18 (CI -0.26..+0.59).
- The change-per-se dent at 8+ games is inside noise: WR residual against the same-QB line -0.22 PPR/g (CI -0.67..+0.25), slope gap -0.044 (CI -0.126..+0.038); TE -0.02. It appears only among 50+ target starters in both seasons (slope 0.575 vs 0.715, n 105, the theory/team-change-discount finding), and adding the ex-ante quality term there leaves the slope at 0.583: quality does not recover it.
- QB EPA/dropback is itself only half-sticky (y/y r 0.49, n 173 passer-seasons with 150+ dropbacks both years), and PFF's catchable-target rate, the accuracy component receivers feel most, has a y/y correlation of 0.29 since 2023 (Macri). A signal that unstable cannot carry a receiver projection.
- Harmon (NFL.com, 2011-2015, 120 top-24 WR seasons): 44% of top-24 WRs played with a passer ranked 15th or worse in adjusted yards per attempt; "if a wide receiver is a proven good player and in line for locked-in volume, we shouldn't fret over their quarterback situation." Same direction.
- Fantasy Footballers (DiSorbo, 2021-2023): WR points correlate moderately with team EPA/play in the same season; that is the realized effect above, and it says nothing about predicting it.
Evidence against
- Realized quality does matter, a little, and a good QB projection is partly predictable (y/y r 0.49): a receiver whose new passer projects a full SD above the old one should still gain about 0.3 PPR/g. The null here is about the new passer's raw prior season, not about a proper QB projection.
- Rookie and backup passers (52% of WR changes, 60% of RB changes) are all set to one replacement value, which throws away whatever rookie draft capital or preseason play would say; their receivers' residual is -0.54 PPR/g vs +0.13 when the new passer is a known veteran (WR, difference inside noise at n 96/87). That split is the direction question A-25 should test, not raw quality.
- RB PPR/g carries a real change-per-se dent (-0.71, CI -1.36..-0.10) that this claim does not explain; a QB change often comes with a new offense or a losing team, and rushing points move with both (A-68).
- Survivors only (8+ games both seasons); EPA/dropback mixes the passer with his line and receivers, so part of a passer's "quality" is the receivers themselves.
How we use it
- Do not adjust a same-team receiver's projection from the new passer's prior-season stats (EPA, PPR/g, catchable rate). The ex-ante slope is a measured null.
- Let the QB's own projection carry the offense: if the new passer projects more than one SD (0.14 EPA/dropback, roughly QB8 vs QB24) above or below the old one, move the receiver's PPR/g by at most 0.35 per SD, and only when the QB projection is not already inside the receiver's projection (ESPN's weekly projection is; nothing is added on top of it).
- Keep theory/team-change-discount rule 2 as written: no share discount for a QB change, and the extra 15% shrink on above-average points applies only to 50+ target starters. It is a change-per-se effect at the top of the distribution, not a quality effect, so it does not scale with who the new QB is.
- Same-team QB change is a reason to widen a receiver's projection interval, not to move its median.
- Re-run the check each August after the data refresh.
Open questions
- Direction split for rookie or backup passers vs veterans (A-25): the -0.54 vs +0.13 residual gap is the hypothesis; needs a QB projection or rookie draft capital as the quality prior for unknowns.
- Does a proper QB projection (not raw prior season) predict the receiver residual better than 0.03 per 0.1? Needs a QB projection archive (A-01/A-30).
- Why RB points drop after a same-team QB change when share does not (A-68).
Claim
Once a player's own recent form, home/away and opponent are held fixed, only a few situational factors move his weekly PPR points by more than a point, and none by more than about 15%:
- Wind is the one weather factor that matters, and it is a gradient, not a 20 mph cliff. QB/WR/TE lose about 0.65% of their expected points per mph of sustained wind (about 0.07 points per mph in level terms). At 15-20 mph a QB is down 13-17% against his own form, a WR down 14%, a TE down 15-20%. RBs are unaffected (flat to slightly up at 20+ mph).
- Cold in still air is small. Most of the cold-game deficit is wind. Netting out wind, sub-freezing air costs a QB about 5% and a WR/TE about 1-3%; RBs are unaffected.
- Domes and closed roofs lift passing-game players 6-9% (QB +8%, WR +9%, TE +6%, RB +2%) relative to a baseline built on mixed conditions. This is the absence of wind and cold, not a separate effect.
- The Vegas implied total is far less than proportional. A starter's points move with elasticity about 0.15 (CI 0.07-0.23) to his team's implied total relative to that team's season norm; TE 0.27, WR 0.09. A 15% higher implied total is worth about +3%, not +15%.
- Favorites score more, at every position, and garbage time does not rescue underdogs. With the game total, wind and dome held fixed, a starter's points against his own form rise 0.4% per point his team is favored by (CI 0.3-0.6%; QB 0.4, RB 0.4, WR 0.4, TE 0.7). Seven-point favorites run 0.97 x form, seven-point underdogs 0.89. Garbage time is about 12% of plays and 10% of a starter's points, and it is the only slice that leans to underdogs; the other 90% leans to favorites by 0.11 points per point of spread, ten times the garbage-time lean the other way. Between half and three-quarters of the favorite effect is the implied total arriving through the spread (rule 1 already prices that); the residual game-script part is about +0.3% per point for RB and WR, nothing for TE, and negative for QB: -0.6% per favorite point once the implied total is fixed (-0.10 PPR per point, CI -0.18 to -0.02; a 7-point underdog QB is +0.7 PPR, about +4%, against a 7-point favorite at the same implied total). The channel is volume, not efficiency: at the same implied total the underdog QB throws 1.7 more passes per 7 points of spread at unchanged yards per attempt. Raw, without the total held fixed, favorites' QBs still score far more (13.4 PPR at 7+ underdog vs 19.6 at 7+ favorite) and throw the same number of passes (30.6 vs 32.3); "underdog QBs throw 5-7 more passes" is not in the data. The earlier "two models disagree on sign" note was a label bug in the effects module, not a disagreement (WO-0020).
- Thursday (short rest) is a measured null. Starters on 4 or fewer days of rest score +3% (CI -1% to +7%) against their own form; QB +8%, TE +9%, WR -1%, RB +2%. The 2012-2014 "Thursday knocks QBs and TEs" finding does not hold in 2019-2025.
- Post-bye is not a boost, and the dip is too small and too unstable to price. With the week of the season held fixed, a starter's first game after his team's bye runs -4% against his own form (CI -8% to 0%, n 1,019), right at the shuffle floor (p95 3.8%); six of seven seasons are negative but 2025 was +14%. The team off a bye scores no more points than the week's norm and covers 46% of spreads. The 2008-2012 "post-bye boost" (+4% all players, QB +8%) is absent in 2019-2025. The post-Thursday mini-bye (9-11 days) is a clean null (-1%). The residual dip sits in TE (-16%, n 134) and in byes from week 11 on (-7.5% vs -1% for byes through week 10); neither is a rule yet (A-69). The rest-day slope the level regression reports is this small bye dip plus the Thursday bump.
- Precipitation, travel distance, time zones, altitude, and the kicker/DST versions of all of the above are unquantified (see Open questions); they get at most a small nudge meanwhile.
Evidence for
Measured, our data (2019-2025 REG, nflverse weekly joined to schedules on game_id; "ratio" below
is a player's PPR points divided by his own trailing 3-game average, starters with a trailing
average of 8+ PPR, n 14,309 player-games; regression subsections further down pool every player
with level-of-points controls):
- Wind, outdoor games. Level model: -0.068 pts per mph (CI -0.088 to -0.049, n 16,385, placebo
floor 0.013, held-out 2025 -0.102, same sign). Check
analysis/checks/theory-situational-factors-wind.pypassed 2026-09-11. Joint ratio model (wind + temp + week, QB/WR/TE, n 6,485): -0.0065 ratio per mph (se 0.0015), QB -0.009, WR -0.0063. Ratio by wind bin, QB: 0-5 mph 1.05, 5-10 0.99, 10-15 0.93, 15-20 0.87, 20+ 0.91 (n 37); WR: 0.94, 0.89, 0.85, 0.80, 0.87 (n 80); TE: 0.85, 0.86, 0.87, 0.78, 0.69 (n 23); RB: 0.92, 0.95, 0.89, 0.94, 1.10 (n 54). The decline is visible from the 5-10 mph bin onward, against Fantasy Life's "hold tight until 20 mph" (their 2018-2022 sample had 48 games over 20 mph; they measured PROE -2.6 points and CPOE -1.6 points there). Advanced Football Analytics (2012) found moderate wind costs 0.7% completion rate and 0.13 yards per attempt, which is a gradient too. Both schools agree passing suffers and rushing does not. - Temperature, outdoor games. Level model: +0.0105 pts per degree F (CI 0.0055-0.0155, n 22,258, held-out 2025 +0.016). Joint ratio model with wind held fixed: +0.002 ratio per degree F (se 0.0006) for QB/WR/TE, -0.0015 (se 0.001, null) for RB. Ratio by temperature, WR: <=25F 0.77 (n 92), 25-40 0.84, 40-55 0.88, 55-70 0.94, 70-85 0.90; QB: 0.85 (n 53), 0.95, 0.98, 0.99, 1.02. But in calm cold (<=32F, wind <=10 mph) QB 0.96 vs 1.02 in calm mild air, WR 0.91 vs 0.93, RB 0.93 vs 0.95: the cold-bin deficit is mostly wind (sub-freezing games average 9.5 mph). 4for4 (1994-2015, 3,935 outdoor games) found game totals flat across temperature (R-squared 0.008; <=20F games averaged 42.8 points vs 42.1 league-wide), consistent with cold being a passing-vs-rushing mix shift rather than a scoring drop.
- Dome / closed roof. Level model +0.36 pts (CI 0.23-0.50, n 39,574 incl. K, held-out +0.33). Ratio: QB 1.07 vs 0.99 outdoors, WR 0.97 vs 0.89, TE 0.90 vs 0.84, RB 0.95 vs 0.94.
- Implied team total. Elasticity check: slope 0.150 (CI 0.069-0.232, n 14,309, floor 0.032; QB
0.15, RB 0.17, TE 0.27, WR 0.09). Check
analysis/checks/theory-situational-factors-implied-total.pypassed 2026-09-11. Ratio by implied total: <=17 pts 0.85, 17-20 0.91, 20-23 0.94, 23-26 0.95, 26-29 0.98, 29+ 0.94; by implied total relative to team norm: <0.85 0.89, 0.85-0.95 0.92, 0.95-1.05 0.94, 1.05-1.15 0.96, >1.15 0.95. Level model on the game total: +0.023 pts per total point (CI 0.008- 0.039). Betting-markets outlets (Sharp Football, Fantasy Footballers) use implied totals as a directional prior and publish no elasticity; the proportional rule of thumb is not measured anywhere we found. - Spread. Sign convention first: nflverse
spread_lineis positive when the home team is favored (correlation with the home margin +0.45, 2019-2025; games withspread_line> 3 have a mean home margin of +7.9).harness.data.effectsbuildsteam_spreadpositive = favored but labels it "negative = favored", so the rendered block below reads "+0.023 pts per point of underdog spread" when it is +0.023 per point of favorite spread (WO-0020 fixes the label). Both models agree. Joint starters' ratio model (game total, outdoor wind, dome and favorite points; n 14,309): +0.0042 ratio per favorite point (CI 0.0027-0.0058, shuffle floor 0.0008; train 2019-2024 0.0043, held-out 2025 0.0039), QB 0.0044, RB 0.0039, WR 0.0037, TE 0.0068. Checkanalysis/checks/theory-situational-factors-spread-direction.pypassed 2026-09-11. Mean ratio by favorite-points bin: under -7 0.89, -7 to -3 0.91, -3 to 0 0.91, 0 to 3 0.93, 3 to 7 0.96, over 7 0.97; by position, QB 0.98 -> 1.07, RB 0.91 -> 0.96, WR 0.85 -> 0.94, TE 0.83 -> 0.93 from the heaviest underdogs to the heaviest favorites. With form, home and opponent also controlled the favorite slope is +0.0064 and the total slope +0.0043 per point; since implied = (total + favorite points) / 2, the total's coefficient is the implied-total share and the excess (+0.002 to +0.003 per point pooled; RB +0.003, WR +0.003, TE 0.000, QB between -0.006 and +0.002 depending on controls) is the pure game-script residual. The level-of-total coefficient is cross-team and weaker than the within-team elasticity in rule 1, so read the split as approximate. Garbage-time split (play-by-play, PPR rebuilt per play; garbage time = 4th quarter at 17+ or win probability outside 5-95%, or 3rd quarter and later at 25+): 12.1% of pass/run plays, 10.2% of starters' points, 14.1% of non-starters' points. Level slope on favorite points with form, home, opponent, total, wind and dome fixed, starters (n 15,478): non-garbage points +0.113 per point (CI 0.093-0.132, floor 0.007; QB 0.26, RB 0.07, WR 0.10, TE 0.13), garbage-time points -0.021 (CI -0.029 to -0.014; WR -0.034, TE -0.026, QB -0.017, RB +0.004). Non-starters (form under 8 PPR) have an all-points slope of -0.010: the pooled model was diluted by low-usage players whose points are garbage time on trailing teams. Checkanalysis/checks/theory-situational-factors-spread-garbage-time.pypassed 2026-09-11. Other schools agree on direction and add the mechanism: Fantasy Footballers (nflfastR, half-PPR since 2008) find favorites win 66% of games (85% at 10+), RBs on underdogs hold a slightly larger share of team yards (33.5% vs 34% is their phrasing, underdog share is the higher one) but the favorite's pie is 27 yards larger, WRs on 7+ point underdogs own 66.5% of team yardage vs 66% on favorites and still score less, conclusion "target points, not game script". FantasyPros (nflfastR since 2013) rejects "trailing teams throw more" because trailing teams are the ones that could not move the ball, and finds QB the only position where the spread reads cleanly on its own. Neither publishes a per-point slope; ours is the first. QB split with the implied total fixed (A-59; level model on the QB's own trailing-3 form of each outcome, home, opponent leave-one-out QB points allowed, outdoor wind, dome, implied team total = (total + favorite points) / 2, favorite points; starters = trailing PPR >= 8 and trailing attempts >= 15; n 3,190 QB-games 2019-2025): total PPR -0.098 per favorite point (CI -0.175 to -0.020, shuffle floor 0.017; train 2019-2024 -0.085, held-out 2025 -0.116); passing points -0.077 (se 0.035), rushing points -0.031 (se 0.018), attempts -0.238 per point (se 0.048), carries -0.015 (se 0.013), yards per attempt +0.005 (se 0.010, null). Without the opponent control the slope is -0.111 (what a bare trailing average should carry); without any total it is +0.200 per favorite point (the raw favorite lean, all implied total and then some). Mobile QBs (trailing carries 5+) -0.152, pocket -0.100. Checkanalysis/checks/theory-situational-factors-qb-spread-split.pypassed 2026-09-11. Read it as: at a fixed own implied total, a bigger opponent total means a higher-scoring, longer game in which the QB throws more; the same coefficient is +0.10 PPR per point of the opponent's implied total. Sharp Football's correlation table (pass attempts per game r 0.29 with QB fantasy points, far below yards and TDs) is the same lesson from the other side: attempts move points only a little. FantasyPros' "QB is the one position where the spread reads cleanly" is the raw +0.20 lean, which is implied total, not game script; once the total is fixed the sign flips. - Rest. Short-rest gap (<=4 days minus 6-8 days) +0.029 ratio (CI -0.013 to 0.071, n 934 vs 11,226); QB +0.077, TE +0.085, WR -0.007, RB +0.015. Check `analysis/checks/theory-
…(truncated)
Claim
Expected touchdowns (xTD) are the sum over a player's carries and targets (for a QB, his pass attempts and own carries) of the league-wide TD rate for that field position and, for throws, depth. Touchdowns over expected (TDoE) are almost entirely noise for RB, WR and TE: the next season keeps about 10% of the excess rate (RB 0, WR 0.12, TE 0.18; y/y r of TDoE 0.04-0.13), and top-decile over-performers score fewer TDoE the next year in 88-96% of cases. Within a season the odd/even-week r of TDoE rate is 0.05. Quarterbacks are the exception in degree, not kind: passing TDoE keeps 0.35 of the excess rate (y/y r 0.22), rushing TDoE 0.31 (r 0.25), top-quintile QBs regress 77% of the time, and in season the passing TDoE split-half r is 0.12 against 0.32 for the field-position mix. Opportunity volume, not TD rate, predicts next season's TDs everywhere; a QB's rushing TDs are the stickiest TD stat in football (y/y r 0.67) because they ride on designed goal-line carries (carries -> next rush TD r 0.56). Making xTD richer does not help: adding down, distance to go and score state to the yardline and depth buckets moves r with next-season TDs by at most 0.01 at every position (measured null), and raw TDs still beat any xTD at TE (0.22 vs 0.17) and QB passing (0.41 vs 0.35). What raw TDs know there is not field position; it is a player- or offense-level TD-rate prior, and the 0.18 (TE) and 0.35 (QB) carry-over terms already price it.
Evidence for
- Measured, RB/WR/TE (2019-2025 REG pbp, consecutive-season pairs, min opps RB100/WR60/TE40 both
seasons). Check
analysis/checks/theory-td-regression.pypassed 2026-09-11 (0.928, n 69 top-decile pairs, floor 0.5). The 2026-09-10 run (0.724, "keep 30%") paired every season with every other season, not consecutive ones; repaired 2026-09-11 and the numbers below replace it.
| pos | n pairs | r TDoE y/y | r TD y/y | r xTD -> next TD | top decile TDoE -> next | share regressing | next TD rate keeps of prior excess |
|---|---|---|---|---|---|---|---|
| RB | 226 | 0.04 | 0.26 | 0.26 | 5.4 -> 0.2 | 96% | -0.05 |
| WR | 296 | 0.13 | 0.31 | 0.31 | 4.7 -> 0.3 | 93% | 0.12 |
| TE | 153 | 0.13 | 0.22 | 0.17 | 3.9 -> 0.8 | 88% | 0.18 |
Bottom-decile under-performers recover symmetrically (93-100% score more TDoE next year).
- Measured, QB (same pbp; 200+ non-sack attempts both seasons, n 155 pairs; rushing lines need
30+ carries both seasons, n 90). Check analysis/checks/theory-td-regression-qb-season.py passed
2026-09-11 (0.774, CI 0.63-0.92, n 31 top-quintile pairs, floor 0.5).
| component | r TDoE y/y | r TD y/y | r xTD -> next TD | r volume -> next TD | top quintile TDoE -> next | share regressing | keeps of prior excess rate |
|---|---|---|---|---|---|---|---|
| passing | 0.22 | 0.41 | 0.34 | 0.29 (att) | 8.0 -> 1.8 | 87% | 0.35 |
| rushing | 0.25 | 0.67 | 0.66 | 0.56 (carries) | 3.3 -> 1.1 | 83% | 0.31 |
| total | 0.25 | 9.2 -> 3.2 | 77% |
League pass TD per non-sack attempt 0.048. For passing, raw TDs beat xTD as a next-season
predictor (0.41 vs 0.34): part of a QB's TD rate is his own or his offense's skill.
- Measured within season, RB/WR/TE: split-half r of TDoE rate RB 0.00, WR 0.08, TE 0.08 while
xTD rate split-half r is 0.20-0.27. Check analysis/checks/in-season-td-rate-noise.py passed
2026-09-10 (0.052, n 1035, floor 0.21).
- Measured within season, QB (100+ non-sack attempts per half, n 233 QB-seasons): passing TDoE
split-half r 0.12 (CI -0.01 to 0.25), raw TD-rate r 0.37, xTD-rate r 0.32; rushing (15+ carries
per half, n 149) TDoE r 0.10, raw TD-rate r 0.21, xTD-rate r 0.17. Check
analysis/checks/in-season-td-rate-noise-qb.py passed 2026-09-11 (0.123, floor 0.324).
- Measured null, richer xTD (same pbp, sacks excluded; consecutive-season pairs, min opps
RB100/WR60/TE40/QB200; n 830). Rich xTD = leave-one-season-out TD rate by yardline x air-yards
x down x distance-to-go (0-2, 3-6, 7+) x score state (down 8+, within 8, up 8+), shrunk k=30
toward the coarse cell, so no player's own season feeds his expectation. Check
analysis/checks/theory-td-regression-rich-xtd.py passed 2026-09-11 (mean r gain over raw TDs
-0.03, CI -0.08 to +0.02, floor 0).
| pos | n | r raw TD -> next TD | r coarse xTD -> next | r rich xTD -> next | r rz opps -> next | r gl opps -> next | TDoE y/y r coarse / rich | keeps of excess coarse / rich | xTD y/y r (rich) |
|---|---|---|---|---|---|---|---|---|---|
| RB | 226 | 0.26 | 0.26 | 0.26 | 0.27 | 0.21 | 0.04 / 0.04 | -0.05 / -0.05 | 0.29 |
| WR | 296 | 0.31 | 0.31 | 0.32 | 0.26 | 0.22 | 0.13 / 0.11 | 0.12 / 0.11 | 0.43 |
| TE | 153 | 0.22 | 0.16 | 0.17 | 0.18 | 0.11 | 0.13 / 0.10 | 0.18 / 0.17 | 0.23 |
| QB pass | 155 | 0.41 | 0.34 | 0.35 | 0.30 | 0.24 | 0.22 / 0.20 | 0.36 / 0.35 | 0.40 |
Down, distance and score state explain the same-season TD total no better either (r 0.69 vs 0.68 at WR, 0.87 vs 0.86 at QB). Raw red-zone and goal-line opportunity counts are weaker than either xTD. xTD is stickier year to year than TDs at WR (0.43 vs 0.31) and about equal elsewhere, which is PFF's (0.60 vs 0.52) and Fantasy Points' (r-squared 0.38 vs 0.28) finding in direction. The Fantasy Footballers model (nflfastR, down + distance + depth + direction, seasons since 2010) publishes the same regression shares we see (WR over-performers decline 89.7%, QB 80.1%) and no r, so no public richer model shows a gain over yardline and depth. - League-wide TD rate by yardline is steep and stable: rushes from the 1 score 58%, from 3-5 yards 25%, from 10-15 yards 6%, beyond 20 under 2%; targets from inside the 5 score 44-57%, from 10-15 17%, from 20-30 6%. Red-zone and goal-line share, not total touches, set xTD. - Outside studies agree on direction and on QBs being stickier. Footballguys (236 QBs, 2005-2015, 250+ attempts): TD rate above 9.5% regressed in 26 of 27 (average -3.0 points, "7-10 TDs"), below 5% recovered in 15 of 17; INT rate regresses even harder (27 of 27, 22 of 23). PFF (2008-2010, TD per completion, 25 QBs): 67% of high and 80% of low rates regressed, 72% pooled against 90% for RB/WR/TE. Fantasy Classroom (2012-2022 starters, per game): passing TDs y/y r 0.34 against 0.61 for attempts and yards, 0.25 for QBs under 30. Pitcherlist and FantasyPros build QB xTD from passing yards and rush attempts (r-squared 0.54 and 0.59) or from xTD% and name candidates without publishing a hit rate. PFF's skill-position alerts hit 86%; Fantasy Points reports xTD stickier than raw TDs (r-squared 0.38 vs 0.28); FantasyPros, ESPN (Clay) and FanDuel publish xTD-style models built the same way.
Evidence against
- Our xTD ignores play-caller, personnel and the player's own history. Down, distance and score state were tried and add nothing (measured null above), but raw TDs still beat xTD at TE (0.22 vs 0.17) and QB passing (0.41 vs 0.35): a player- or offense-level prior exists that field position cannot see. The carry-over terms (0.18 TE, 0.35 QB) are the current stand-in for it (A-55, A-43).
- Bins are coarse and the shrinkage constant (k=30) is a choice; a formation- or personnel-aware model (nflverse has no route or alignment field on disk) could still differ. A tracking-data xTD is out of reach for this harness.
- QB samples are thin: 155 pairs, 31 in the top quintile, 90 rushing pairs. The 0.35 and 0.31 carry-over coefficients have wide intervals; the direction (QB stickier than RB/WR/TE) is what four sources and our data agree on, the size is ours alone.
- Population is survivors with enough volume in both seasons; role losers who fell below the volume floor are excluded, which flatters every y/y correlation here.
- The QB carry-over may belong to the offense (play-caller, pass catchers) rather than the QB; a team-change split would separate them (A-43).
- Regression is a population statement applied to individuals; a few red-zone specialists and elite-QB pass catchers do sustain positive TDoE, and the RB coefficient of -0.05 says the group as a whole does not.
How we use it
- Season prior for TDs, RB/WR/TE: projected TDs = xTD rate at last season's location mix x projected opportunities + 0.1 x (last season's TDs - last season's xTD). Use 0 for RB, 0.12 WR, 0.18 TE when the position matters. Replaces the 2026-09-10 rule's 0.3, which came from the mis-paired check. Apply at every magnitude, both directions.
- Season prior for TDs, QB: passing TDs = league rate by field-position and depth mix x projected attempts + 0.35 x last season's passing excess rate x projected attempts; rushing TDs = xTD rate on his own carries x projected carries + 0.3 x last season's rushing excess. Project QB rushing TDs from carries and goal-line role first (r 0.56-0.67), rate second.
- Never rank on raw TDs: when two sources disagree because one leans on last season's TD total, weight the opportunity-based one. A TD total is a volume signal first and an efficiency signal only through the 0.1 (RB/WR/TE) or 0.35 (QB) share.
- In season, ignore TD rate for RB/WR/TE (split-half r 0.05). Project rest-of-season TDs from xTD per opportunity times expected opportunities. A player whose red-zone or goal-line share rose is a buy; one whose TDs rose without it is not.
- In season, QB TD rate gets a quarter weight: after 8+ games shrink a QB's observed passing TDoE rate by 0.75 toward zero (split-half r 0.12, CI to 0.25) and let the offense's field-position mix (r 0.32) carry the rest. Do not sell a QB for a cold TD-rate start or buy one for a hot start alone; do move on a change in designed goal-line carries.
- Trade and waiver pricing: a RB/WR/TE is over-priced by the market roughly in proportion to TDoE x 0.9 x points per TD; a QB by TDoE x 0.65 x points per TD. Use that as the discount when comparing against opportunity peers.
- Re-run all five checks each August when the new season's pbp lands (agenda A-03, A-17, A-18).
- Keep xTD coarse. Yardline (and depth for targets) is the whole field-position signal; do not spend effort or tokens on down, distance or game-script terms, and treat any source's "richer xTD" as equivalent to ours unless it publishes a next-season r gain. Red-zone and goal-line opportunity counts are a weaker substitute (r 0.11-0.27), use them only when pbp is unavailable. At TE and QB, where raw TDs beat xTD, rules 1-2 (xTD plus the 0.18 / 0.35 carry-over) already recover what raw TDs know; do not add raw TDs on top.
Open questions
- What do raw TDs know at TE and QB that field position does not: a multi-season player TD-rate prior (two prior seasons of TDoE, or career TDoE rate) or the offense (team TD rate per red-zone trip, play-caller)? Test a 2-season TDoE prior against the 1-season carry-over. (A-55)
- Is the QB carry-over the QB's or the offense's? Split QBs who changed teams from those who did not; if it travels with the QB, price it as skill. (A-43)
- Does INT rate regress as hard in our data as Footballguys' 2005-2015 sample says (27 of 27)? Matters for leagues that penalize INTs. (A-42)
- Does the 10% RB/WR/TE carry-over concentrate in identifiable players (goal-line backs, big-slot TEs), or is it spread thinly?
Claim
A team change costs a player usage share in two ways: the group's share level falls (movers arrive to a smaller role, on average 8-17% smaller), and less of any above-average prior share survives (the slope of next-season share on prior share falls from about 0.8 to 0.75 for WR targets, 0.62 for TE targets, 0.56 for RB carries, 0.52 for RB snaps, and 0.32 for RB targets). Together, a top-quartile prior share projects at about 0.87x (WR), 0.85x (TE), 0.80x (RB carries, snaps) and 0.67x (RB targets) of what the same player would project to on his old team. A change of primary passer on the same team is a measured null for share persistence (WR target share r 0.80 vs 0.82, TE 0.78 vs 0.82, RB carry share 0.79 vs 0.76): the QB moves efficiency and points, not the share of the pie. Measured on nflverse 2019-2025, regular season, per-game means over active weeks, 8+ games in both seasons, team = modal team that season, primary QB = most pass attempts on that team that season.
Evidence for
Season-to-season OLS slope and level ratio (mean share in s+1 divided by mean share in s), 8+ active games both seasons, same team vs new team:
| metric | same team: slope, y/x (n) | new team: slope, y/x (n) | top-quartile prior: new/same projection |
|---|---|---|---|
| WR target share | 0.82, 1.00 (477) | 0.75, 0.88 (155) | 0.87 |
| WR snap share | 0.72, 1.01 (477) | 0.63, 0.90 (155) | 0.87 |
| TE target share | 0.81, 0.99 (268) | 0.62, 0.92 (65) | 0.85 |
| RB carry share | 0.77, 0.99 (282) | 0.56, 0.85 (93) | 0.81 |
| RB snap share | 0.71, 1.02 (282) | 0.52, 0.85 (93) | 0.80 |
| RB target share | 0.69, 1.00 (282) | 0.32, 0.83 (93) | 0.67 |
The RB carry-share slope gap (0.56 vs 0.77, CI on the new-team slope 0.40-0.72) is the check. Among established starters (50+ targets or 100+ carries both seasons) the gaps are larger and noisier: WR target share slope 0.59 vs 0.70 (n 56), TE 0.27 vs 0.72 (n 33), RB carry share 0.22 vs 0.62 (n 32), RB target share 0.07 vs 0.52 (n 26). Shuffle noise floors are 0.01-0.17 at these n.
Same team, primary passer changed vs not (Pearson r, 8+ games): WR target share 0.800 vs 0.819 (gap -0.02, 95% half-width 0.19, n 183 / 294); TE target share 0.78 vs 0.82; RB carry share 0.79 vs 0.76; RB snap share 0.75 vs 0.70. Shares do not care who throws. Points do, a little: WR PPR/g r 0.73 vs 0.80 (8+ games) and slope 0.575 vs 0.715 among 50+ target starters, so a new QB shrinks a receiver's above-average points about 15% more while leaving his share projection alone. This matches the in-season QB-change null in theory/ball-share-shifts (share gap +0.003 over 76 switches).
PFF (players who changed teams since 2010, top-36 RB / top-24 QB, TE ADP): RB beat ADP 6 of 35 and improved PPG 12 of 31 healthy; WR beat ADP 35% and improved PPG 31%; TE beat ADP 0 of 15 and matched or improved PPG 7 of 15. The market already discounts movers and still overrates RB and TE movers, which is the same ordering as the slopes above (RB and TE lose more than WR).
Evidence against
- Selection, not just disruption: movers had lower prior shares than stayers (WR 0.148 vs 0.164), and the level drop mixes players cut for cause with stars traded for. A receiver a team paid to acquire may sit above the group line; the split by mover type is A-44.
- Modal team per season misclassifies mid-season trades (a week-9 move looks like a partial-season mix) and treats the rare QB-plus-receiver package move as a plain team change.
- Survivors only (8+ games both seasons). The starter subsets are thin (n 26-56) and their slopes have wide intervals; the 8+ games population drives the rule.
- 4for4's predictability tables (2017-2023, 30+ targets) do not filter by team, so public stickiness numbers already blend movers and stayers; the rules-of-thumb claim that "a new signal caller disrupts the entire distribution hierarchy" (Fantasy Strategy Guide) cites no data and our data rejects it for share, though not for points.
- Play-caller change is not separated here (WO-0014, A-24); some of the new-team loss is a new offense rather than a new roster.
How we use it
- Same team, same primary QB: project share from prior-season share at the usage-stickiness slope (about 0.8 for WR/TE target share and RB carry share, 0.7 for RB target and snap share).
- Same team, new primary QB: no discount to the share projection. For 50+ target starters only, shrink the player's above-average PPR/g excess by an extra 15% (multiply the excess over positional mean by 0.85); at 8+ games the points dent is inside noise (residual -0.22 PPR/g, CI -0.67..+0.25). The shrink does not scale with who the new QB is: the new passer's prior-season EPA/dropback predicts none of the receiver's next-season points (slope 0.03 PPR/g per 0.1, floor 0.34, n 183; theory/qb-quality-and-receiver-points). Let the QB's own projection carry the efficiency change. Do not treat a QB change as a reason to fade a share.
- New team: project share as the new-team line,
new-team mean + new-team slope x (prior - mover prior mean). The shortcut, applied to the same-team projection for an above-average prior: WR target and snap share x0.87, TE target share x0.85, RB carry and snap share x0.80, RB target share x0.67. For a below-average prior use the level ratio alone (x0.88 WR, x0.92 TE, x0.85 RB). This replaces the old "shrink 25% toward the mean" rule, which came out near x0.96 and undershot. - These are pre-season and weeks 1-4 priors only. Once the in-season window (in-season/usage-window) has 3-4 games on the new team, observed share replaces the prior and the discount is spent.
- RB receiving work on a new team is close to unprojectable from history (slope 0.32, 0.07 for 30+ target backs); use the new team's prior-season RB target share and the depth chart instead.
- Re-run both checks each August after the data refresh; the falsifier resets
last_verified.
Open questions
- Does the level drop depend on why the player moved (traded for, signed as a starter, cut) or on the vacated share at the destination? (A-44.)
- Answered (A-45): the points-only QB effect is not a quality effect priceable from the new passer's prior season (ex-ante slope 0.03 PPR/g per 0.1 EPA/dropback, a measured null); realized quality is worth about 0.25 per 0.1 for changed and same-QB pairs alike (theory/qb-quality-and-receiver-points). Whether rookie/backup vs veteran passers split it is A-25.
- Play-caller continuity split (WO-0014, A-24).
Claim
Usage share is the most persistent thing about a fantasy player, and how persistent depends on the horizon and the position. Season to season, per-game target share (WR, TE) and carry share (RB) carry r of roughly 0.73-0.80; snap share about 0.70; per-game touchdowns about 0.45-0.50; fantasy points per game sit between. Week to week, snap share and RB carry share are the stickiest (r about 0.74-0.78), target share is moderate (WR 0.60, TE 0.56, RB 0.42), and touchdowns are close to noise (r under 0.17). A team change costs a share metric 0.1 to 0.4 of its persistence, worst for RB receiving work. Measured on nflverse 2019-2025, regular season, per-game means over active weeks (snaps or touches > 0).
Evidence for
Season-to-season r (per-game means, 8+ active games in both seasons, our data; n in parentheses):
| metric | QB | RB (375) | WR (632) | TE (333) |
|---|---|---|---|---|
| target share | 0.66 | 0.79 | 0.78 | |
| carry share | 0.86 (157, carries/g) | 0.73 | ||
| snap share | 0.27 | 0.67 | 0.70 | 0.68 |
| WOPR | 0.65 | 0.80 | 0.81 | |
| air-yards share | 0.35 | 0.78 | 0.82 | |
| PPR points/g | 0.54 | 0.70 | 0.77 | 0.75 |
| TDs/g | 0.69 | 0.48 | 0.51 | 0.44 |
Range restriction: the check's population (50+ targets in both seasons, n 347) gives WR target share r 0.67 (CI 0.61-0.73; last pair 2024-2025 r 0.67, n 60); TE 0.58 at 50+ targets; RB carry share 0.57 at 100+ carries. Volume filters remove the low-share players that carry most of the spread, so the 8+ games table above is the right prior for a full roster and the filtered number is the right one for choosing between established starters. Both clear the noise floor by 10x.
Week-to-week r (consecutive active weeks, same season): snap share RB 0.74, WR 0.78, TE 0.75; RB carry share 0.75; target share RB 0.42, WR 0.60, TE 0.56; PPR points RB 0.44, WR 0.39, TE 0.37, QB 0.30; TDs 0.09-0.17. Shuffle noise floor for every metric is under 0.06.
Same team vs new team, season-to-season r: WR target share 0.81 vs 0.72 (n 473 / 159); TE target share 0.81 vs 0.61; RB carry share 0.77 vs 0.58; RB snap share 0.72 vs 0.50; RB target share 0.74 vs 0.33.
Public studies agree on the ordering: 4for4 (2018-2023, 100+ attempts) finds RB receptions/g 0.70, routes/g 0.64, rush attempts/g 0.61, TD rate 0.05; Sharp Football (10 seasons, R-squared) finds WR targets/g 0.54 and TD/target 0.01, TE targets/g 0.60 and TDs/g 0.24; 4for4 and Sumer report target share and WOPR above 0.70 for WRs; Fantasy Classroom finds carries 0.76 and YPC under 0.30 (2012-2022).
Evidence against
- All stickiness is measured on survivors: players who played 8+ games in both seasons. Injured or benched players are exactly the ones a projection most needs, and they are absent from the pairs.
- Per-game means hide within-season role changes (a week-1 backup who starts from week 9 looks average).
- Public numbers vary by filter and by r vs R-squared; treat cross-source comparisons as ordinal.
- Routes run are not in the loaded data (WO-0012); the TE literature says routes beat snaps there.
How we use it
- Rank projection inputs by measured persistence: snap share and carry share, then target share, WOPR and air-yards share, then points per game, then TDs last. Efficiency stats are not inputs.
- Prior-season share is a strong prior for WR/TE target share and RB carry share (r about 0.75-0.80). Use it at full weight when the player stayed on the same team.
- Team-change discount: when a player changed teams, use theory/team-change-discount (A-16): the same-team projection x0.87 for WR target/snap share, x0.85 TE target share, x0.80 RB carry and snap share, x0.67 RB target share. The old "shrink 25% toward the mean" rule undershot (about x0.96) because it ignored the level drop movers take on arrival. A same-team change of primary passer is a measured null for share and gets no discount.
- Weekly lineups: a one-week spike in targets is weak evidence (r 0.42-0.60 to next week); a one-week spike in snap share or RB carry share is strong evidence (r 0.74-0.78). See in-season/usage-window.
- Never project TDs from last week's TDs; they are noise at the weekly horizon and only half-persistent at the season horizon (theory/football-probability handles the regression rule).
- Re-run the check each August after the data refresh; the falsifier resets
last_verified.
Open questions
- QB split done (theory/team-change-discount, 2026-09-11): a same-team QB change leaves share persistence intact (WR r 0.80 vs 0.82) and only dents points persistence. The play-caller split waits on a coaching table (WO-0014, A-24; theory/coaching-change-share-reset holds the interim rule).
- Snap share for RBs is stickier week to week than carry share is season to season; is snap share the better prior for RB touches once routes are available? (WO-0012.)
Claim
A trade helps only if it raises our rest-of-season starting-lineup projection (and, late in the year, our playoff-weeks projection) by more than the depth it costs. Consolidation 2-for-1 trades are the most reliable way to do that in H2H, because only starters score. Buy low and sell high are about the gap between usage (targets, snaps, red-zone touches) and recent results (touchdowns, big plays), not about hot and cold streaks.
Evidence for
- Every school agrees: value the trade by rest-of-season projection, not season-to-date points, and count the bench player you drop when accepting a 2-for-1.
- Usage-driven buy-low signals (strong target share, poor TD luck) precede rebounds; TD-driven sell-high signals precede regression (see theory/football-probability.md).
- Bye-week arbitrage: trading a player with a bye ahead for an equal one whose bye has passed adds a usable game (rules-of-thumb).
Evidence against
- Trade value charts are one analyst's opinion and swing weekly; using them as truth violates PG-3.
- League mates are not rational; "fair" trades by projection often fail because the counterparty anchors on name value. Success requires modeling the other manager, which we cannot do well yet.
- Consolidation trades increase concentration risk: one injury undoes the gain.
How we use it
- Evaluate any trade with
team-range-of-outcomessimulation before and after: accept only if P(championship) rises by at least 1.0 percentage point. Before week 8, also accept for a P(playoffs) rise of 3+ points even if P(championship) falls (Q-0005: maximize championship odds, then playoff odds). After week 8, if P(playoffs) is below 0.45, flip the priority and evaluate by P(playoffs) instead (team-range-of-outcomes.md rule 3). - Always include the dropped bench player's replacement value in a 2-for-1 or 3-for-1.
- Buy-low list: players whose target share or snap share is top-24 at position but whose fantasy points rank 12+ spots worse. Sell-high list: the inverse, or TD rate more than 1.5x expected.
- Prefer trading for players whose playoff-weeks (15-17) matchups are favorable once past week 9 (in-season/bye-and-playoff-planning.md).
- Never propose a trade the harness would not accept from the other side by these same rules with a 20% discount (keeps offers credible and reduces spam to league mates).
- Every proposal and response becomes a decision record; REFLECT scores it against re-simulated odds 4 weeks later.
Open questions
- Do we value future draft picks at all? Assume no (redraft) until the league type is confirmed.
Resolved 2026-09-11 (Q-0005, owner, school: league-owner-preference, weight: 1.0): maximize
P(championship) first, then P(playoffs), re-evaluated as the season progresses; rule 1 above now
states the threshold directly instead of pointing at the open question.
Claim
Weekly matchup selection at DST and K reaches most of what the season's best units score, at near-zero draft cost, and beats chasing recent form by a wide margin.
- DST: the opponent's Vegas implied team total is the signal. Each point of opponent implied total is worth -0.37 DST points (CI -0.41 to -0.33, n 3,742 team-games 2019-2025, ESPN default scoring without yards-allowed tiers). Mean DST points by opponent implied total: <=17: 8.7, 17-20: 7.5, 20-23: 6.4, 23-26: 5.3, 27+: 4.1. That is a 4.5-point spread between the best and worst matchup buckets against a weekly SD of 5.3. Points allowed is the component the line predicts (r 0.40); sacks are second (r -0.20); interceptions, fumbles and TDs are barely predicted (|r| < 0.09). A unit's own trailing four-week scoring has r 0.11 with next week and adds only 0.08 points per trailing point once the implied total is in the model; the prior season's mean has r 0.10. Dome is a null for DST (r -0.05).
- DST streaming comparison, per-season averages 2019-2025: the five DSTs with the lowest opponent implied total each week averaged 8.3, the hindsight season top five 8.7, the five best by trailing four-week points 7.1, and all units 6.1. Matchup streaming reaches 95% of a comparator nobody can pick in August, and beats form-chasing by 1.2 points per week.
- K: own implied team total plus dome. Each point of own implied total is worth +0.14 K points (CI 0.10 to 0.18, n 3,676 kicker games), holding spread and roof fixed. A dome adds +0.8 points (se 0.16), the same as six points of implied total. Mean K points by own implied total: <=17: 6.6, 17-20: 7.2, 20-23: 7.8, 23-26: 8.3, 27+: 8.6. The spread is a null (-0.04 points per point of |spread|, inside its error), the game total on its own is a null (r 0.03), and a kicker's trailing four-week points are worthless (r 0.04).
- K streaming comparison: the five highest implied totals each week averaged 8.7, the hindsight season top five 9.8, all kickers 7.9. The matchup captures less of the gap at K than at DST because the kicker himself matters (FG accuracy r 0.44 with points per game, offense TD rate 0.43 per Winks); implied total is a +0.8 edge over an average kicker, not a +1.9 one.
- K weather: wind is a small gradient, cold is a small real cost, neither works through misses. Outdoor games 2019-2025 (n 2,216 kicker games, box-score wind and temperature), holding own implied total, |spread| and temperature fixed: -0.035 K points per mph of wind (CI -0.071 to +0.002, shuffle floor 0.009), half the QB/WR/TE gradient (-0.068). Mean K points by wind: 0-5 mph 7.7, 5-10 7.5, 10-15 7.2, 15-20 7.4, 20+ 6.1 (n 42). FG% is flat across wind bins (0.82-0.88) and misses do not rise (+0.002 per mph, null): teams manage wind by attempting less (1.57 FG attempts per game at 20+ mph vs 1.9) and the offense scores fewer TDs (PAT made -0.015 per mph, se 0.006). Temperature: +0.0136 K points per degree F (se 0.0055), so a 32F game costs about 0.4 points against a 60F one. Calm cold (<=32F, <=10 mph) averages 7.05 vs 7.73 in calm mild air (n 79 vs 1,206; implied total explains 0.1 of the gap). FG% is unchanged in the cold (0.836 vs 0.841); the cost is fewer attempts (+0.0033 per degree) and fewer 40+ yard makes (+0.0024 per degree). Dome or closed roof averages 8.5 vs 7.45 outdoors, the +0.8 in the joint model above.
- DST weather is a measured null. Same window (n 2,264 outdoor DST team-games), holding the opponent's implied total fixed: wind +0.007 points per mph (CI -0.035 to +0.049), temperature +0.006 per degree F (se 0.006). Sacks, turnovers and defensive TDs do not move with either; points allowed falls 0.08 per mph of wind (se 0.04) but the ESPN points-allowed tiers barely register it (+0.012 per mph). The bins wander without a trend (15-20 mph +0.8 above the implied-total fit, n 156; 20+ mph -1.3, n 42; calm cold -0.6, n 82) and none clears its noise.
- TE streaming is viable only when the elite tier is gone; a streamed TE loses to an elite TE but not to a mid-round one. Unquantified here (in-season/usage-window covers the target-share rule).
Evidence for
- Two checks on 2019-2025 data:
analysis/checks/waivers-streaming-dst-opp-implied.py(slope -0.37, floor 0.015, passed) andanalysis/checks/waivers-streaming-k-own-implied.py(slope 0.14, floor 0.008, passed). Both run directly; the sandboxed runner is blocked by WO-0018. - Two weather checks on the same window, run through
harness.checks2026-09-11 (the runner works for these):analysis/checks/waivers-streaming-k-wind-cold.py(wind -0.035, bounded claim, passed) andanalysis/checks/waivers-streaming-dst-wind-cold.py(wind +0.007, null holds, passed). - StatsbyLopez (analytics-projection, GAM on 2000-2014 attempts): cold costs long field goals most (50-yarders 70% -> 50%, 30-yarders 94% -> 89%); a -25F wind chill is worth about 10 yards of distance. Advanced Football Analytics (2012): FG success is flat across wind bins (82% under 6 mph and over 25 mph) because teams choose not to attempt in wind, while a 52-yard try at <=30F succeeds about 30% vs 55% in moderate air. Our data agrees on the mechanism: the fantasy cost of wind and cold shows up as fewer and shorter attempts, not as misses.
- Sharp Football (betting-markets): a 3% FG drop at 15 mph and scoring -5% between 25F and 50F; directionally consistent with the -0.4 points cold cost and the attempt decline.
- Subvertadown (analytics-projection, 2017-2019): default DST scoring self-correlates at only 0.145 week to week and yards allowed is the most consistent component (0.175); a projection model reaches 0.37-0.42. Our r 0.11 for trailing form agrees; our r -0.27 for the implied total alone is most of what their model gets.
- ESPN (rules-of-thumb): DSTs facing the five best matchups each week beat the season top five by about a point per game in each of four seasons (10.4 vs 9.6). Our replication with a cleaner comparator (hindsight top five, points-allowed-only scoring) finds streaming 0.4 short of hindsight rather than ahead; the direction of the ESPN claim (streaming ~ top five) holds, the margin does not.
- Winks (analytics-projection, three seasons): K1 11.0, K2 10.3, replacement 8.2 points per game; trailing per-game average has almost no correlation with the next week; FG% r 0.44 and offense TDs r 0.43 with points per game. Our r 0.04 for trailing form agrees.
Evidence against
- 4for4 (analytics-projection): kickers scored 10+ "more than twice as often" at implied total 27+ than at 26 or less. Falsified at ESPN default scoring: 37.8% vs 33.4%. The implied total is a real but modest kicker signal; weight 0.3 for that source.
- Our DST points omit yards-allowed tiers (ESPN default has them; this league's items are unverified, A-41). Yards allowed is the most self-consistent component per Subvertadown, so the implied-total slope is a floor under yards-allowed scoring, not a ceiling. Sack/turnover-heavy scoring would weaken it.
- Hindsight season top five is a biased comparator (selected on the outcome); an ex-ante draft pick lands in it about 4% of the time per ESPN, so the streaming gap to a drafted top DST is smaller than 0.4 and probably negative.
- Streaming requires a roster spot's worth of churn and a weekly decision.
- FantasyLabs (analytics-projection, no n) reports DST +1.05 plus/minus at <=30F and a non-monotone wind pattern (+0.56 at 10+, -0.34 at 15+). The cold premium does not replicate here (calm cold -0.6 below the implied-total fit, n 82); their wind pattern is the same bin noise we see. Weight 0.2.
- The kicker wind CI touches zero (+0.002 at the top). The claim is a bound (between zero and the QB effect), not a point; its confidence is 0.6 on its own. The 20+ mph cell is 42 games.
- nflverse
windandtempare one box-score number per game, not the kickoff-hour forecast the weekly decision uses; WO-0017's backfill (A-36) would re-test both checks on kickoff-hour values.
How we use it
- Draft: DST and K in the final two rounds, choosing by week 1-3 matchups, not season rank (roster-building.md rule 4).
- Weekly DST: rank available DSTs by opponent implied team total, ascending; 0.37 points per point of implied total. Ignore the unit's last four weeks except as a tiebreaker (worth 0.08 points per trailing point). Dome, wind and cold are not factors (measured nulls): do not move a DST for a forecast. Stream when the best available faces an opponent implied total 4+ points lower than our DST's opponent (about 1.5 points).
- Weekly K: rank by own implied team total (0.14 points per point) plus +0.8 for a dome. Ignore the spread, the game total on its own, and the kicker's recent scoring. Stream when the best available beats ours by 1.5+ points on that scale (about 10 implied points, or a dome plus 4). 3b. Weekly K weather (outdoor, forecast at kickoff): subtract 0.035 points per mph of sustained wind above 8 mph (a 20 mph forecast is -0.4, 25 mph -0.6; ignore forecasts under 10 mph) and 0.4 points when the kickoff temperature is at or below 32F (0.2 between 33F and 45F). Both are small next to the implied-total and dome terms: a cold, 20 mph game is worth about 5 implied points, so it breaks ties and rarely moves a streaming call on its own. Never fade a kicker for weather when his implied total is 3+ points higher. As a multiplier for the simulator: K x(1 - 0.0047 x (wind - 8)), floor 0.90, and x0.95 at or below freezing.
- TE: if we have no top-5 TE, rank available TEs by target share and red-zone targets over the last three weeks; stream when the best available beats ours by 2+ projected points.
- Never hold two DST or two K. Hold two TEs only when both project as top-12 or one is a stash for a known injury return.
- Playoff weeks: lock in DST/K choices for weeks 15-17 by week 13 if a clear multi-week matchup run exists.
Open questions
- How does this league score DST (yards-allowed tiers, sack and turnover weights) and K (distance
tiers, miss penalties)? Re-run both checks under the league's
scoringItems(A-41). - Opponent sack rate allowed as a second DST sort: unmeasured; sacks are the only other component the line predicts.
- Kickoff-hour wind and precipitation for K (rain and snow are unmeasured; situational-factors rule 8 holds K at x0.90 for a heavy-precipitation forecast): re-run both weather checks once WO-0017's historical weather backfill is on disk (A-36).
- Is the DST null real or masked by the points-allowed tiers? Under this league's yards-allowed and sack weights (A-41) wind may register through points and yards allowed; the raw points-allowed slope is -0.08 per mph here.
Claim
Waivers are where H2H leagues are won after the draft. The best adds are opportunity-driven (a starter got hurt, a depth chart changed, snaps or targets jumped) rather than production-driven (a big box score on a normal role). FAAB is a season-long budget: overspending in September forfeits the November league-winner. In rolling-priority leagues, priority is spent only on adds that project as multi-week starters.
Evidence for
- Every school agrees that role change predicts future points better than one big week (film-and-usage; theory/football-probability.md stickiness).
- FAAB guides converge on percentage bands: 40-70% for a true league-winner (backup RB into a full starting job), 15-30% for a new weekly starter, 5-12% for flex upside, $1-4 for streamers; odd-number bids beat round-number crowds.
- $0 bids on speculative players are free options.
- What the "backup RB into a starting job" add is actually worth (nflverse 2019-2025, n 172 lead-back absences, lead back = 0.40+ trailing carry share; the promoted back is the top remaining RB by prior carry share; share mechanics in theory/ball-share-shifts):
| promoted back | first-game PPR (from his trailing 4) | >= 12 PPR | >= 16 PPR |
|---|---|---|---|
| all (n 172) | 12.5 (from 6.9); median 10.8, SD 8.7 | 0.45 | 0.30 |
| committee partner, prior share 0.20+ (n 104) | 14.0 (from 8.7) | 0.51 | |
| pure backup, prior share < 0.20 (n 68) | 10.2 (from 4.3) | 0.37 | |
| second absence game (n 88) | 12.0 | 0.42 |
So the promoted back is a mid-RB2 in the mean, an RB1 three games in ten, and under 6 PPR in 28% of games. The league-winner FAAB band fits a committee partner or a backup on a run-heavy team, not every handcuff. - How long the job lasts, and what survives the starter's return (same events, nflverse 2019-2025, analysis/checks/waivers-handcuff-hold-four-games.py and waivers-handcuff-hold-after-return.py; an inactive game counts as zero share):
| question | all | committee partner (pre share 0.20+) | pure backup (< 0.20) |
|---|---|---|---|
| absence lasts 4+ games (n 35): 0.50+ carry share in 3 of the first 4 games | 0.37 (game-1 rate 0.57) | 0.44 | 0.29 |
| same, all 4 games | 0.17 | 0.22 | 0.12 |
| carry share games 1-4 | 0.52 / 0.46 / 0.41 / 0.45 | 0.58 / 0.49 / 0.48 / 0.56 | 0.45 / 0.42 / 0.33 / 0.33 |
| PPR games 1-4 (4-game mean, >= 12) | 13.7 / 10.9 / 8.7 / 10.6 (11.0, 0.49) | 15.3 / 11.9 / 10.9 / 13.9 (13.0, 0.56) | 12.0 / 9.9 / 6.4 / 7.1 (8.8, 0.41) |
| active in all 4 games | 0.74 | 0.89 | 0.59 |
| lead back's first game back (n 105): promoted share vs his pre-absence share | 0.22 vs 0.25 (absence 0.51) | 0.27 vs 0.31 | 0.13 vs 0.13 |
| same, PPR vs pre | 6.4 vs 7.3 | 7.4 vs 8.7 | 4.5 vs 4.5 |
| P(promoted share >= 0.30 in the return game) | 0.30 | 0.39 | 0.14 |
Half the absences last one game (78 of 172), a quarter 2-3 games, a fifth 4+. Game-1 share predicts the hold: a promoted back who took 0.55+ of the carries in game 1 averages 0.54 over games 2-4 (n 17), one who took under 0.40 averages 0.30 and never holds 0.50 for three straight (n 9). After the return the promoted back goes straight back to his old role in the mean (share -0.03 vs pre, CI -0.06..+0.01, against +0.25 if the promotion had stuck), and the returning lead back opens at 0.87 x his share, 0.86 x his snap share and 0.88 x his PPR. The one exception is earned: a promoted back who averaged 14+ PPR during a 2+ game absence holds +0.19 share above his pre-absence level in return games 2-3 and is at 0.30+ share 64% of the time (n 14); under 14 PPR he is at -0.07 and 15% (n 34). Thin, so a lean not a rule. No public source (ESPN, Footballguys, RotoWire, PFF, Aug 2025-Aug 2026) publishes a hold rate or a return-game split; all treat the handcuff as an insurance ranking and argue from single fill-in games.
Evidence against
- Bands are conventions, not measured optima; the right bid depends on league mates' spending, which we must learn.
- The hold and return numbers pool injuries with healthy scratches and suspensions, and the 4+ game sample is 35 events (17 pure backups), so the role split there is a direction, not a rate; the earned-slice fork rests on 14 events. Absences that ran to season end count toward the hold rate but have no return game, so the return sample leans to shorter absences (48 of 105 are one game). The lead back's own return numbers count a re-injury as zero share; in-season/injury-processing rule 4b is the cleaner source for his ramp.
- Speculative stashes occupy bench slots that otherwise hold streaming flexibility; there is a real opportunity cost.
How we use it
- Each waiver window, score every free agent on two axes: role delta (change in projected snaps/targets/carries from last week) and ROS starter probability (from team-range-of-outcomes simulation). Rank by role delta first.
- FAAB bids (if FAAB): league-winner 45-60% of remaining budget, new weekly starter 15-25%, flex upside 5-10%, streamer 1-3%. Bid odd amounts. Never exceed 60% before week 10 unless the add is a top-12 RB/WR by ROS projection.
- Rolling priority (if not FAAB): use priority only for an add with ROS starter probability > 0.5; otherwise add as free agent after waivers clear.
- Keep one bench slot as a flex slot for streaming and speculative adds; never fill the roster with 2 QBs, 2 TEs, 2 DST, or 2 K.
- Drop order: lowest ROS starter probability first, never a player whose role delta was positive in the last two weeks.
- Speculative stash rule: a backup RB behind an RB1 with 2+ multi-game absences in the last two seasons is worth a $0-3% bid any week. 6a. Promoted-back bids: when a 0.40+ carry-share lead back is ruled out for 3+ games, project the top remaining RB at his trailing PPR plus 6 (theory/ball-share-shifts rule 7), which lands a committee partner (prior share 0.20+) around 14 PPR and a pure backup around 10. Bid the new-weekly-starter band (15-25%) for the committee partner, the flex band (5-10%) for the pure backup, and the league-winner band only when the absence is season-long and the promoted back's projection clears the top-12 RB line. A one-game absence is a streamer bid, not a starter bid. Price his TDs from the lead back's goal-line share, not a discount of it: the promoted back inherits the goal-line job whole (0.58 of team carries inside the 5 vs 0.56 for a healthy lead back, n 81; theory/ball-share-shifts rule 7c), so do not drop him a band because "a bigger back will get the goal line" unless the team has named one. 6b. Hold horizon: a promoted back is a hold for exactly as long as the lead back is out, plus nothing. Project his role at the promotion level for every absence game (theory/ball-share-shifts rule 7) but assume it lasts the announced absence, not longer; the job survives a month only about a third of the time (0.37 for 3 of 4 games), and a pure backup is out of the lineup or hurt himself in 4 of 10 four-game absences. Re-check after his first game: 0.55+ carry share in game 1 means he is a weekly start while the lead back is out; under 0.40 means the committee formed and he is a flex at best, so drop-priority rises even before the starter returns. 6c. Return: the week the lead back is back, project the promoted back at his pre-absence baseline (share and PPR), not a blend; he keeps no residual slice in the mean. A pure backup goes back to about 4 PPR and is a drop for any add with a positive role delta. A committee partner is what he was before (about 7-9 PPR). The only stash-through-return case is a promoted back who averaged 14+ PPR across a 2+ game absence: hold him one more week at half weight and re-check his share in the return game, since two-thirds of those keep 0.30+ share afterwards (n 14).
- Every claim is a decision record; REFLECT scores adds by starter-weeks delivered.
Open questions
- Whether free agency is continuous outside the weekly processing window changes rule 3 timing.
- Is the earned-slice fork (14+ PPR during the absence -> +0.19 share after the return) the backup's doing or the lead back coming back diminished? Needs the lead-back threshold lowered to 0.35 for a larger pool and the returning back's own snap ramp held fixed (agenda A-60).
Resolved 2026-09-11 (Q-0001): this league is rolling waiver priority, no FAAB (mSettings
FAAB=no). Rule 2 (FAAB bid bands) is dormant for this league; rule 3 (rolling priority) is the
one in force. Claims are filed Tue 18:00 ET, ESPN processes waivers Wed ~03:00 ET
(harness/schedule.py calendar), which fixes the "waiver processing day" half of the old
question.