What doesn’t.work

Evidence

• First registered run: 445 calls over Weeks 4-18 of 2025, 375 offense-games in the test. Calls hit 55.2%, control 51.2%: +4.0 points, p = 0.079, 95% interval -0.5 to +8.4. The pre-registered bar was p < 0.05
• Second registered run, after a review tightened the rule for which slices count as signal (amendment 1, registered before it was built): 356 offense-games, calls 55.3%, control 51.5%: +3.8 points, p = 0.106
• Both runs carry the same label: no edge demonstrated
• Descriptive only: pass slices about +5 points in both runs, run slices nothing; the top third by expected advantage hit 61-65%
• Each week's calls used plays from earlier weeks only, and persistence was refitted on 2022-2024 so the engine never saw the season it was tested on

Sample

375 offense-games, 2025 regular season, Weeks 4-18

Method

Pre-registered before the script existed. A call says the offense's EPA per play on its top-ranked slice will be at or above the league average for that slice. The control is the same statement about every other slice in the same game. Paired difference per offense-game, two-sided t-test.

Caveat

One season. The verdict word here is the registry's nearest fit: the pre-registered label was NO EDGE DEMONSTRATED. A second season is the test that matters, and 2026 is being graded as it is played.

Evidence

• EPA allowed per dropback through Week 3 against Weeks 4-18: r = 0.10. Per designed run: r = 0.05. The threshold for telling a correlation from zero on 128 team-seasons is 0.17
• Cut 36 ways (play-action, blitz, depth, direction, target position, run gap, box count), 2 of 36 defensive slices carry any signal after three games
• The same test on offenses: EPA per dropback r = 0.40, and 18 of 36 slices carry signal
• By Week 8 a defense's EPA allowed per dropback reaches r = 0.28

Sample

32 teams over four seasons, 2022-2025, 128 team-seasons

Method

For each team-season, the number through Week 3 against the same team's number over Weeks 4-18, centered on that season's league mean so a league-wide shift cannot pose as persistence, pooled over four seasons. Repeated for windows through Week 12.

Caveat

Raw EPA, not adjusted for the three opponents faced. Adjusting may recover some signal and has not been tested yet. A defense can also change mid-season through injury, which this cannot see.

Evidence

• Yards per target allowed through Week 3 against the rest of the season: cornerbacks r = 0.10, safeties r = 0.00, linebackers r = 0.19
• Through Week 9 the same three are r = 0.14, 0.06 and 0.03
• Share of a defense's targets that go at each group: cornerbacks r = 0.55, safeties r = 0.52, linebackers r = 0.64 after three games
• Pressures per 100 snaps: edge rushers r = 0.26, interior line r = 0.28 after three games, interior line r = 0.37 by Week 9

Sample

32 teams over four seasons, 2022-2025, 126-128 team-seasons per cell

Method

Targets, yards, pressures and missed tackles charged to each defender by Pro Football Reference's charting, summed by position group per team, early window against the rest of the season, centered per season and pooled.

Caveat

Charted targets are assigned to the nearest defender by a human charter and miss zone handoffs. A group also changes when a starter is hurt; position labels here are end-of-season roster labels.

Evidence

• 35 of 36 offensive tendencies carry signal after three games
• Runs from under center r = 0.83, motion before the throw r = 0.75, runs inside r = 0.66, targets to tight ends r = 0.61, play-action r = 0.46
• A defense's own choices are stable too: blitz rate r = 0.59
• Even the strongest keeps only about two thirds of its distance from league average, so every tendency on a sheet is pulled toward the league before it is used

Sample

32 teams over four seasons, 2022-2025, 128 team-seasons

Method

Each tendency as a share of its base population through Week 3 against the same share over Weeks 4-18, centered per season and pooled. The regression slope is the share of a deviation from league average that survives.

Caveat

Measured within a season on a continuing play-caller. A mid-season change of play-caller or quarterback resets it, and the sheet's coaches table has to catch that.

Evidence

• Year-over-year persistence r = 0.08 at QB and r = 0.05 at WR, against a significance threshold of 0.20
• RB (r = 0.31) and TE (r = 0.25) are real but weak
• Measured on 96 team-season transitions, drift-adjusted so a league-wide trend cannot be mistaken for a team

Sample

32 teams over four seasons, 2022-2025, 96 transitions

Method

Correlate a defense's PPR allowed per game to a position in season Y against the same figure in Y+1, after removing league drift. Compare the predictive error of carrying the number forward against simply using the league average.

Caveat

This measures the season-to-season persistence of the stat, not whether a defense is weak in the week you are looking at. A within-season signal may exist; we have not tested it.

Evidence

• Only pre-snap motion persists strongly (r = 0.74); shotgun 0.59, blitz 0.57, red-zone rush 0.55, no-huddle 0.51
• For play-action, deep rate, screen rate and the coverage families, a team's own prior number predicts next season WORSE than the league average does
• Raw carry-forward loses to the league mean on 7 of 11 rates; shrinking toward the mean by measured persistence beats it on all 11

Sample

32 teams over four seasons, 2022-2025, 96 transitions

Method

Compute each team's rate per season from play-by-play, participation and FTN charting; correlate Y against Y+1 after removing league drift; compare mean absolute error of carry-forward, league mean, and a shrunk estimator.

Evidence

• Blitz splits: the split predicts next season at r = 0.854, but the player's overall rate predicts it at r = 0.876
• Game-script splits: 0.837 against 0.866, the same way round
• Both come out negative against the player's own baseline; only target depth carries information beyond the player, at +6.5%

Sample

785 to 1,173 player-season pairs depending on split type, players with at least 50 opportunities in both seasons

Method

For each player and split bucket with n >= 15 in consecutive seasons, test whether the split value predicts next season's split better than the player's overall PPR per opportunity does.

Evidence

• League man-coverage rate reads 28.8%, 41.3%, 48.4%, 30.9% across 2022-2025 — defenses do not move like that
• Cover-1 and Cover-3 swing 15 percentage points in exact mirror image, the signature of a classifier moving the boundary between two lookalike single-high shells
• Published 2023 figures put Cover-1 near 18.7% and Cover-3 near 35.7%; our feed reads 34.8% and 19.9% — the two families near-inverted, against a documented league trend toward zone

Sample

Four seasons of scrimmage dropbacks, roughly 19,000 per season, participation join rate 100% throughout

Method

Compare league-level coverage distributions across seasons on a fixed denominator, check feed schema changes and label completeness, and compare against published external charting.

Caveat

Charting providers define coverage families differently, so the external comparison is corroboration rather than proof. It does not need to be proof to stop us citing the numbers.

Evidence

• 136 blind paired picks: our call finished top-3 32 times against the consensus's 40
• Outright #1s tied 12-12; average finish 16.14 against 12.54
• All four configurations of the engine lose on the blind seasons, and the one chosen beforehand on principle turned out to be the worst of the four

Sample

2022 and 2023, 17 weeks x 4 positions x 2 seasons = 136 paired picks, neither season run before the rule was fixed

Method

Pre-register the configuration, endpoint, test and labels in a committed file; run the blind seasons once; report McNemar's exact test on podium rate and paired tests on finish.

Caveat

136 paired picks can only detect a large effect. This is not proof that no edge exists; it is proof that none has been demonstrated.

Evidence

• Expected calibration error 0.0024 over 5,866 forecasts, base rate 1.09%
• Bias +0.0007 — we promise very slightly more winners than arrive
• Paired t on the residuals gives t = 0.53, p = 0.598, so the gap from honesty is not distinguishable from zero
• Equal-width bins instead of population bins give ECE 0.0015, so the binning choice is not carrying the result

Sample

5,866 forecasts, one per active player per position-week, 2025 weeks 2-18

Method

Bin the published pTop by population, compare predicted frequency against observed frequency, and decompose the Brier score into reliability, resolution and uncertainty.

Caveat

Calibration is not accuracy. A forecast that says 1% for everybody would also be near-calibrated on a 1% base rate; the resolution component (0.00030) is what says we are doing more than that, and it is small.

Evidence

• Kaggle API returns 401 without a token; the CLI refuses before sending a request
• One 2017 game is genuinely public on the NFL’s own GitHub: 316,025 rows, 23 tracked objects, 10 Hz, verified by downloading it
• The 2026 release covers 2023 and 2024 — exactly the two seasons whose coverage labels we quarantined — with a coverage call from a different charting source

Sample

Eight competition releases surveyed; one 2017 game opened and measured directly

Method

Probe every documented endpoint with curl and record real HTTP status codes; download and count the one public file rather than citing its documentation.

Caveat

The licence terms were read from pages that resist automated fetching, so the quoted clauses should be re-checked in a browser before any commercial decision rests on them.

Evidence

• Across 32 testable metric-weeks in 2025: current season best 15 times, prior season 6, league mean 11
• Blitz: current-season tape wins 10 of 16 weeks and every week from 12 onward bar one (week 18: 6.84pp error against the prior season's 8.43pp)
• Pressure: no crossover exists — the league mean wins 7 weeks, and the spread between all three predictors is under a point most weeks

Sample

32 defences, 2025 weeks 3-18, two tendencies whose feed is comparable across the season boundary

Method

For each week, predict a defence's rate from (a) its current season to date, (b) all of its prior season, (c) the league to date, and compare mean absolute error against what it actually did that week.

Caveat

Only blitz and pressure could be tested. Coverage, man/zone and box — the tendencies a film room most wants — cannot cross this boundary because 2024's participation labels are quarantined. Enough to retire the blanket rule, not enough to claim a crossover week for anything else.

Evidence

• Correlation between games played in year Y and Y+1: r = 0.05 (RB), 0.10 (QB), 0.17 (WR), 0.29 (TE), against the rCrit = 0.20 bar this project uses elsewhere
• Carry-forward mean absolute error LOSES to the group mean at running back (3.42 games against 3.36)
• Nobody plays a full season either: players who finished top-12 at their position averaged 14.2 games the next year and only about half reached 16

Sample

624 player-seasons, 2021-2025, players finishing inside the drafted universe at their position

Method

Rank every player within position by PPR total in season Y; count regular-season games played in Y+1; correlate, and compare the predictive error of carrying a player's own games-played forward against simply using the group mean.

Caveat

This measures season-to-season persistence of games played, which mixes injury with losing a job. It is the right quantity for a draft projection and the wrong one for a medical question.

Evidence

• Pooled rank correlation RB 0.408 against 0.287 for the better baseline; the margin widens each season (0.26, 0.42, 0.55)
• WR 0.338 against 0.300; beat both baselines in 2024 and 2025, tied in 2023
• QB 0.097 and TE 0.050 — near zero for every method tested, ours included
• Top-12 hit rate is 0.42-0.50 for OUR board and 0.42-0.50 for the baselines: nobody picks the eventual top twelve better than a coin flip

Sample

three preseason boards (2023, 2024, 2025) graded on the union of each method's top N and the actual top N

Method

Rebuild the board with no knowledge of the target season — no game logs, the season's first depth-chart snapshot rather than its newest, no betting lines — then Spearman against the realised finish, plus top-12 and top-24 hit rates.

Evidence

• First-round backs opening as the listed starter averaged 14.1 carries and 4.7 targets per team game in year one; sixth- and seventh-rounders behind two backs averaged 1.8 and 0.4
• First-round receivers as the listed WR1 averaged 6.9 targets per team game against 0.06 for undrafted receivers behind two others
• Measured per TEAM game, so games a rookie misses are already inside the rate rather than applied twice

Sample

2022-2025 rookie seasons at QB/RB/WR/TE, roster-joined to the season's opening depth chart

Method

Collapse each rookie's usage vector to one PPR-weighted scalar; fit a capital effect and a depth effect on the position's drafted-rookie base, each multiplier shrunk toward 1 by its own sample size.

Caveat

Cells are thin at the top of the board — two to six first-round backs across four seasons — which is exactly why the multipliers are shrunk rather than read raw.

Evidence

• Blitz: +4.7% in quarterback residual SD per 1 SD of opponent blitz rate, p=0.130, and 0 of 13 tests survive BH FDR at q=0.10. Minnesota 2023 blitzed on 52.2% of dropbacks, the heaviest defence-season in four years, and NARROWED quarterback outcomes (residual SD 6.12 against a 7.12 league figure); the three lightest-blitzing defences all sat above league dispersion
• The killer number: a defence's tendency to widen quarterback outcomes carries year to year at r=+0.026 across 96 transitions (p=0.80) against this project's rCrit=0.20 — while the blitz rate itself carries at r=0.576. A permutation test says there is no such defence trait to persist: between-defence variance 0.270 against a shuffled null of 0.248, p=0.234
• Game script: Justin Jefferson's outcome SD was 9.78 as a 6+ point favourite and 9.62 as a 3+ point underdog, a ratio of 1.02. Across 12 spread and total specifications, 0 survive FDR and 0 reach even uncorrected p<0.05 in the 2025 holdout
• Red-zone defence: a TE-only effect clears FDR in-sample and then fails out of sample, and the trait cannot persist anyway — the league's best red-zone defence fell to 13th, 18th, 11th and 16th the following year, four seasons running
• Pace, coverage shell and funnel defences: all refuted, mean- and volume-controlled

Sample

2021-2025 play-by-play, participation and FTN charting; fit 2021-2024, held out on 2025

Method

Model log residual-variance with the player's own leave-one-out mean and weekly volume as covariates (and in the stricter form, player-season fixed effects), cluster-robust on defence-season, with Benjamini-Hochberg FDR across the whole test family and a permutation null.

Evidence

• Controlled for mean, volume and games, year-over-year persistence of volatility is r=+0.086 at WR and +0.135 at RB — both under the rCrit=0.20 bar; TE +0.229 with a confidence interval straddling zero; only QB (+0.320) clears cleanly
• UNCONTROLLED the same quantity looks compelling and would have shipped: RB r=+0.700, TE +0.572, QB +0.570, WR +0.470
• Zack Moss was the most volatile qualifying back in the league in 2022 (100th percentile) on 5.0 points a game as a backup; given a real role in 2023 his volatility fell to the 76th. The boom/bust was his snap count, not his running style
• Tank Bigsby, same shape: 100th percentile volatility in 2023 on 1.9 points a game, 87th in 2024 on a real role

Sample

625 player-seasons across four transitions, 2021-2025, with a 2024-25 holdout

Method

Coefficient of variation and log-SD residualised on the player-season mean, opportunity volume and games played, z-scored within position-season, drift-adjusted, correlated Y against Y+1.

Evidence

• Head to head on teammate pairs, 2021-2025: plain volume share ranked next season correctly 68.7% of the time (268/390), inside-the-ten share 62.1%, goal-to-go 63.7%, red-zone 'leverage' 54.4% — a coin flip
• On points per game the gap is wider: RB volume 79.2% against 72.2% for inside-the-ten and 50.0% for leverage; WR volume 78.0% against 66.5% and 51.0%
• It never beats what the board already knows from volume, and when it disagrees with volume it is backwards

Sample

390 teammate pairs at the same position, 2021-2025

Method

Compute inside-20, inside-10 and goal-to-go opportunity shares from play-by-play; rank teammate pairs by each metric and by plain volume share; score each against the pair's actual next-season finish.

Evidence

• Within-team first differences: +1 takeaway per game buys +2.82 yards of starting field position (p=1.1e-12) and +0.593 offensive drives (p=3.8e-5) — but only +0.61 plays (p=0.40), because plays per drive falls 0.245 (p=0.0033). A short field means fewer snaps to finish the drive, and the two effects cancel
• Persistence, drift-adjusted, n=128, against rCrit=0.20: takeaways/g r=+0.157, turnover margin +0.150, offensive drives +0.165, plays +0.158, and drive-start field position r=-0.039 — the one thing takeaways genuinely buy is the flattest series in the study
• The lagged link a board would actually use is nil: last season's takeaways predict this season's drives at r=-0.008, plays +0.064, skill opportunities +0.06
• Mean reversion is the strongest single result: r=-0.649 between a season's takeaway z and the change into the next (p=1.1e-16). Top-8 takeaway defences gave back 74% of their turnover margin
• Jacksonville 2025, the case that prompted this: best starting field position in football (own 32.9, 1st) off the 2nd-most takeaways, and it produced the 13th-best fantasy offence. Drift-adjusted they gained +5.85 plays a game and the takeaway model explains 13% of them — 87% came from the offence, not the defence

Sample

160 team-seasons, 2021-2025, with a 2025 holdout; 61 tests, Benjamini-Hochberg at q=0.10

Method

Team-season panel of drives, plays, drive-start field position and takeaways from play-by-play; pooled, within-team first-difference and drift-adjusted persistence estimators, controlling for prior point differential and the preseason market line.

Caveat

The user's mechanism was half right and it is worth saying which half: the defence did buy Jacksonville the field position and the possessions. It did not buy the plays, and the scoring it did buy went to the quarterback rather than to anyone else on the roster.

Evidence

• Zero of 18 mean tests on offensive volume and shape clear even raw p<0.05: plays/g +0.01 SD (p=0.98), drives/g +0.02 (p=0.51), pace +0.08 (p=0.78), pass rate -0.12
• On 599 stayer player-seasons the standard-deviation ratio of change in opportunity share under a new staff is 1.040 (Levene p=0.21); target share 1.069, snap share 1.020, points per game 1.014 — no measurable scrambling
• The only thing that reliably predicts a player's role change is mean reversion of his own prior share, slope -0.23, and it is identical under a new staff (-0.231) and a continuing one (-0.205)
• The real driver of forecast error is a bad prior season: bad-prior teams that KEPT their staff carried their offensive profile at r=0.274 against 0.473 for good-prior continuity teams
• One genuine year-one effect exists and is fantasy-irrelevant: a new head coach roughly doubles the dispersion of shotgun rate (variance ratio 2.15 raw, 1.94 quality-matched)

Sample

coaching changes 2022-2025 with a 2025 holdout; 599 stayer player-seasons

Method

Compare year-one change in team volume and in player opportunity share under new staffs against continuity teams, with Levene tests on dispersion and controls for prior point differential and the preseason market line.

Caveat

Twenty-one of thirty-two offences changed hands for 2026. A tag that fires on two thirds of the league was never going to separate anything.

Evidence

• 0 interceptions logged for NYJ across 17 regular-season games and 1,419 defensive plays, in both play_by_play_2025.csv and stats_player_week_2025.csv
• The league minimum excluding NYJ is 6, and 380 interceptions were thrown league-wide in 2025
• A zero-interception season on 1,419 defensive snaps is not a possible outcome

Sample

all 32 defences, 2025 regular season

Method

Count interceptions and fumble recoveries credited to defteam in play-by-play, cross-checked against the player-week feed and against league-wide interceptions thrown.