A player database reaching back to 2002 in the NFL and spanning twelve college seasons and 60,987 college athletes. 349,000 graded consensus lines across the last two NFL and college seasons. Every stat category swept, every popular belief tested, every result replicated across seasons or thrown out. This is the complete methodology behind everything GWTTKB publishes — including the studies that failed.
Most fantasy analysis has a dirty secret: nobody checks whether it was right. A take gets published in July, the season happens, and by the time anyone could grade it, there’s a new take. We built this site on the opposite premise. Every claim below is a measured result — a correlation, a hit rate, a mean-absolute-error — computed on our own database and required to replicate before we’d repeat it. When something failed, it’s in here too, in a section we call the graveyard, because the failures are half the value.
One thing this article is not: betting advice. We grade consensus lines because they’re the closest thing football has to an honest forecast, not because we want you wagering. The question here is narrow and useful — what actually predicts what a player does next — and it turns out the betting market is one of the best instruments ever built for answering it.
Before any finding gets used in our projections, rankings, or the Coach, it has to pass a replication bar most fantasy content never faces: the effect must show the same sign in every season tested (up to five, 2021–2025), with a meaningful magnitude in at least three.[1] We swept roughly 70 stat fields per position — every counting stat, every share, every efficiency metric, trailing levels and short-term trends both — against what actually mattered: the player’s PPR scoring over the next four weeks, controlling for his current baseline. That target matters. It’s not “who was good last year.” It’s the redraft question: given everything I can see right now, who scores going forward?
One deliberate choice worth explaining, because it’s the opposite of a limitation: the replication sweeps run on the most recent five seasons of weekly data, on purpose. Football changes — the 2003 passing game does not predict the 2026 one, and a factor that needs twenty-year-old seasons to look significant is a factor that stopped working. The deeper history isn’t wasted, though: two-plus decades of NFL careers and a dozen college seasons power everything longitudinal in this article — the draft-class outcome distributions, the career-stage expectation curves, the comp engine, the college-to-NFL crosswalks, the injury spell base rates. Depth for the slow questions, recency for the fast ones.
With that many tests, some junk sneaks through by chance — expect roughly 5% of nothing-signals to pass replication on luck alone. So we hold the line at the strong stuff: results with |r| ≥ 0.08 in multiple seasons are findings; borderline passes are candidates and get labeled that way.[2]
Here’s the single most important chart we’ve ever produced. It’s the full-database sweep for wide receivers — 3,222 qualifying player-stretches — showing which trailing stats predicted the next month of fantasy scoring, and which predicted the opposite.
Read the bottom three rows again, because they’re the part nobody wants to believe. A receiver coming off a hot stretch of yards-per-target, receiving EPA, or touchdowns doesn’t just fail to keep it up — he is measurably more likely to decline than a comparable player who wasn’t hot. Negative, every season, five seasons running. Efficiency heaters are borrowed points. The market for “he’s just better per touch” is a market for regression.
Meanwhile the top of the ladder is boring, and boring is the point. Target share is the best forward predictor of WR scoring we can find in seventy fields. Same story at tight end, where it’s even louder — target share hits +0.337 on 1,266 TE stretches, with WOPR right behind at +0.326, while catch rate (−0.146) and yards-per-target (−0.147) replicate negative. Running backs follow the pattern with their own vocabulary: snap share (+0.198) and carries (+0.125) are the sticky signals, and a hot stretch of receiving efficiency (−0.143) is the trap. Rising short-term usage trends — targets, target share, WOPR ticking up over the last three games versus the last five — all replicate positive for RBs too. Role changes are real. Efficiency changes are weather.[3]
And quarterbacks? Quarterbacks are humbling. Across 74 fields and 1,151 stretches, almost nothing about a QB’s own recent play replicates as a positive forward predictor. Recent completions and attempts actually replicate negative (−0.101, −0.089) — garbage-time-inflated passing stretches mark the top, not the takeoff. What predicts QB fantasy is the environment (team implied total, the one situation factor that survived all five seasons) and the legs: red-zone and goal-line carry share, rushing first downs. Passing efficiency — CPOE, completion percentage, passer rating — predicts nothing forward. A passer-rating heater is a sell signal wearing a suit.
Here is the number that should reframe how you read every fantasy article you’ll see this August. We tested 272 stat fields across the four positions. Exactly 41 replicated. An 84.9% failure rate. The overwhelming majority of the numbers available to you — including most of the ones sold as “advanced” — carry no repeatable forward signal at all. Not weak signal. None that survives five seasons of asking.
| Position | tested | survived | the signal that predicts | the signal that fades |
|---|---|---|---|---|
| WR | 70 | 14 | target share +0.217, WOPR +0.192, targets +0.191 | rec TDs −0.128, rec EPA −0.124, yds/tgt −0.109 |
| TE | 60 | 12 | target share +0.337, WOPR +0.326, targets +0.274 | rec TDs −0.184, yds/tgt −0.147, catch rate −0.146 |
| RB | 68 | 11 | snap share +0.198, carries +0.125, rising targets +0.086 | receiving EPA −0.143 |
| QB | 74 | 4 | nothing positive replicated | completions −0.101, attempts −0.089 |
Read the QB row one more time. Out of seventy-four measured things a quarterback did last month, not one predicted his next month positively in a way that held up across five seasons. Everything that replicated ran the wrong direction. That’s not a gap in our data — it’s the finding. Quarterback fantasy scoring is an environment-and-rushing story, and the entire industry of passing-efficiency analysis is, for forecasting purposes, decoration.
Notice too what’s identical across the three receiving positions: the same three stats predict and the same three stats fade, in the same order, at WR, TE, and RB. Volume on top, efficiency and touchdowns on the bottom, every time. Four independent sweeps, four different sample pools, one answer. When a rule reproduces itself that cleanly across positions that share almost no context, you’re not looking at a quirk of the data. You’re looking at how football works.
The last decade handed fantasy a wave of genuinely sophisticated metrics, and it would be dishonest to wave at them generically. So here they are by name, with what each one did in our sweeps. Every number below is the same test as everything else: does knowing this about a player help you forecast his next four weeks?
| Receiving metrics (WR / TE) | result | verdict |
|---|---|---|
| Average separation | +0.012 | no signal |
| Average cushion | −0.019 | no signal |
| YAC above expectation | −0.012 | no signal |
| RACR | −0.033 WR / −0.055 TE | no signal |
| Share of intended air yards | +0.052 | no signal |
| Catch rate | −0.080 WR / −0.146 TE | fade it |
| Receiving EPA | −0.124 WR / −0.145 TE | fade it |
| Air-yards share | +0.122 WR / +0.231 TE | real, replicated |
| WOPR | +0.192 WR / +0.326 TE | real, replicated |
| Red-zone target share | +0.187 WR / +0.185 TE | real (ceiling) |
| Rushing metrics (RB) | result | verdict |
|---|---|---|
| Rush yards over expected | 0.000 | exactly nothing |
| NGS rushing efficiency | +0.005 | no signal |
| Yards per carry | −0.027 | no signal |
| Time to line of scrimmage | −0.097 | no signal |
| Eight-plus defenders in box % | −0.081 | no signal |
| Rushing EPA | −0.082 | no signal (leans fade) |
| Receiving EPA | −0.143 | fade it |
| Snap share | +0.198 | real, replicated |
| Carries | +0.125 | real, replicated |
| Red-zone / goal-line carry share | +0.133 / +0.103 | real (ceiling) |
| Passing metrics (QB) | result | verdict |
|---|---|---|
| CPOE | −0.096 sweep / −0.113 ceiling | actively negative |
| Completion percentage | −0.074 | no signal |
| Passer rating | −0.029 | no signal |
| Yards per attempt | −0.009 | no signal |
| Aggressiveness | −0.046 | no signal |
| Time to throw | −0.003 | no signal |
| Intended air yards | +0.012 | no signal |
| Completed air yards | +0.013 | no signal |
| Air yards to sticks | +0.052 | no signal |
| Passing EPA | +0.051 | no signal |
| Sacks / sack yards | 0.000 | exactly nothing |
| Interceptions | 0.000 | exactly nothing |
| Red-zone carry share | +0.175 | real (ceiling) |
| Rushing first downs | +0.154 | real (ceiling) |
| Goal-line carry share | +0.148 | real (ceiling + only floor factor) |
Start with the running backs, because that table contains the single most brutal number in this entire study. Rush yards over expected — the NFL’s own flagship metric for isolating a runner’s talent from his blocking — returns exactly 0.000. Not 0.03. Zero. It is a genuinely brilliant measurement of who is good at running the football, and it tells you precisely nothing about who will score fantasy points next month. Its companions die alongside it: NGS rushing efficiency (+0.005), time to line of scrimmage (−0.097), eight-plus defenders in the box (−0.081), and plain yards per carry (−0.027). Every metric designed to separate a back from his situation fails, and the two that work — snap share (+0.198) and carries (+0.125) — are just counting how often he’s on the field. At running back, the situation is the player, at least as far as your lineup is concerned.
The quarterback table is the same story told through the passing game, and it ends with three metrics returning literal zeroes: sacks, sack yards, and interceptions. A quarterback’s turnovers and pressure tell you nothing about his fantasy future. Neither does how far he throws it (intended air yards +0.012, completed air yards +0.013, air yards to sticks +0.052), how fast he releases it (−0.003), how aggressively he attacks coverage (−0.046), or how efficiently he moves the ball (passing EPA +0.051, yards per attempt −0.009). Then there’s CPOE, which manages something rarer than uselessness: it’s actively backwards. A quarterback coming off a high completion-percentage-over-expected stretch is less likely to post a ceiling game next month (−0.113, replicated all five seasons). It’s a legitimate measure of passing skill and a fantasy sell indicator at the same time, because efficient passing at low volume is how you get a quarterback who wins games and loses fantasy weeks. The only things that survive at the position are the rushing lines: red-zone carry share (+0.175), rushing first downs (+0.154), goal-line carry share (+0.148, and the only replicated floor factor at quarterback). Law #3 in one table — QB fantasy is environment plus legs, and the arm is decoration.
Which brings us back to the receiving table, and the least comfortable finding in this article. Separation, cushion, and YAC above expectation — the crown jewels of the player-tracking era — carry essentially zero forward fantasy signal. Not because the metrics are badly built; they measure exactly what they claim to. They just measure things that either don’t persist or don’t convert into fantasy points once you know a player’s volume. Separation is a skill that earns targets, and once you’ve counted the targets, the separation has already told you everything it’s going to.
The pattern across all three tables is worth stating plainly, because it explains why our engine is built the way it is: every advanced metric that survived is a measure of opportunity, and every one that died is a measure of efficiency. WOPR, air-yards share, snap share, and red-zone carry share survive because they count chances. RACR, EPA, separation, CPOE, and rush yards over expected die because they are rate stats — and a rate, however elegantly measured, is the part of football that regresses. The tracking era didn’t fail — it just confirmed, with much better instruments, the least glamorous thing anyone could have told you.
Touchdowns are the most valuable fantasy event and the least stable fantasy stat, and the resolution to that paradox is one of the cleanest findings in the whole program: the count reverts, but the role that produces the count persists.
Last season’s TD total is a negative forward predictor at every position — for WRs, trailing receiving TDs comes in at −0.128, one of the strongest fade signals on the whole board. But red-zone target share — the share of your team’s trips inside the 20 that come with your name on them — is the #1 ceiling factor for wide receivers in our boom study (+0.187, replicated all five seasons) and simultaneously one of the strongest bust-protection factors (−0.187 against bust risk). End-zone target share replicates right behind it. For RBs, it’s red-zone and goal-line carry share doing the same double duty; for QBs, goal-line carry share is the only replicated floor factor at the position, full stop.
Fade last year’s touchdown count. Pay for last year’s red-zone role. When those two point in opposite directions on the same player — big TD total, thin red-zone share — that’s the single most reliable “regression incoming” profile we can construct.
A thing you’ll hear in every draft room: “he’s got a safe floor but no ceiling.” We went looking for that tradeoff in 17,000+ scored player-stretches and it is not there. Volume raises the ceiling and the floor at the same time. The ceiling factors and the floor factors are the same list — target share, WOPR, snaps, red-zone role — wearing different signs. What actually creates fragility is TD-dependent and efficiency-dependent production: the profiles scoring on a high yards-per-target or a fat catch rate bust at elevated rates when the efficiency inevitably normalizes.
One real exception, and it is a positional split rather than a league law. At RB, WR, and TE, boom rate and bust rate are strongly negatively correlated (−0.65, −0.72, −0.59) — the players who boom often are the same players who bust rarely, so a high boom rate marks a consistent producer, not a volatile one. At quarterback the sign flips positive (+0.356), and the same reversal shows up on week-to-week volatility versus boom rate. “High ceiling” means something genuinely different at quarterback than everywhere else: at QB it really does come with a floor cost, and at every other position it does not. If you are paying a volatility tax for a high-ceiling receiver, you are paying for a risk that does not exist.
Where does explosive upside genuinely live, then? In lower-baseline players carrying elite usage indicators. A flex-scoring WR (5–9 PPG baseline) boomed in 24.7% of his next-month windows against a 10.8% boom rate for established studs — because studs are already priced at their level and the flex guy with a rising role has room to double. Same shape at RB: 26.6% boom rate in the flex band versus 9.6% for studs. The bust rates barely differ. Upside isn’t a personality trait. It’s a low baseline plus a real role.
Now the part this whole article has been building toward. For two full seasons we converted consensus closing player lines — cross-market medians, receptions plus receiving yards plus rushing and passing lines plus anytime-touchdown probability with the margin haircut applied — into an implied fantasy score for every player, every week: 7,141 player-weeks of “what does the market think this guy scores.”[4] Then we graded it against reality, and against the thing every fantasy manager actually uses instead: the player’s own recent scoring.
It wasn’t close.
The market beat the box score at every single position, overall MAE 4.96 to 5.54. And weekly accuracy is the market’s weakest event. The rest-of-season result is the one that should change how you play:
A player’s consensus-implied scoring through the season’s first four or five weeks predicts his rest-of-season output better than his actual scoring does. Correlation with weeks 5–18 PPG: 0.827 in 2024 and 0.827 in 2025 for the early-season market number, against roughly 0.73–0.75 for prior-year production. Not one season, both seasons, same coefficient to the third decimal, which frankly spooked us too. When we fit the optimal blend between market and prior production for rest-of-season projection, the answer came back 85–90% market.[5] From roughly week 3 onward, your rest-of-season projections should be market-first, and everything on this site is.
Which sets up the actually useful question. If the market and the box score both track rest-of-season — what happens when they disagree about a player?
Take every player with lined games in weeks 1–5 and at least eight games after. Compute the gap: market-implied PPG minus actual PPG over that early stretch. A positive gap means the market kept pricing a player above what he’d shown — the lines “saw more.” A negative gap means he was outscoring his own lines — the market wasn’t buying the hot start. Then just… watch what happened.
| 2024 — market saw more (top 20 names) | implied | actual (wk1-5) | ROS PPG | vs. early |
|---|---|---|---|---|
| Matthew Stafford | 15.9 | 10.3 | 14.8 | +4.5 |
| Will Levis | 13.3 | 7.8 | 11.1 | +3.3 |
| Patrick Mahomes | 19.2 | 14.2 | 19.3 | +5.1 |
| Jalen Hurts | 21.4 | 16.9 | 22.5 | +5.6 |
| Caleb Williams | 18.0 | 13.4 | 15.6 | +2.2 |
| Sam LaPorta | 11.2 | 6.7 | 12.3 | +5.7 |
| Mark Andrews | 8.5 | 4.4 | 13.9 | +9.5 |
| Keenan Allen | 10.2 | 6.0 | 13.9 | +7.8 |
| Tank Dell | 10.9 | 7.4 | 11.0 | +3.6 |
| Travis Kelce | 12.4 | 9.4 | 13.5 | +4.1 |
| Bijan Robinson | 16.4 | 13.5 | 22.8 | +9.3 |
| Tyreek Hill | 14.9 | 12.1 | 13.1 | +1.0 |
| Josh Allen | 21.4 | 18.8 | 25.9 | +7.2 |
| Jared Goff | 17.0 | 15.5 | 20.2 | +4.7 |
| Rachaad White | 12.1 | 9.7 | 15.1 | +5.4 |
| Breece Hall | 16.5 | 14.3 | 15.4 | +1.1 |
| Justin Herbert | 13.0 | 10.4 | 18.8 | +8.4 |
| CeeDee Lamb | 17.7 | 15.3 | 18.7 | +3.4 |
| Trey McBride | 11.3 | 9.4 | 17.2 | +7.8 |
| Josh Jacobs | 13.0 | 11.9 | 19.5 | +7.6 |
All twenty improved rest-of-season. These are the twenty biggest fantasy names on the 2024 divergence board (the full-sample math, deep-roster guys and all, is at the bottom of this section — the signal doesn’t need the curation). Andrews was being outright dropped in shallow leagues at 4.4 PPG while his lines never moved — the market refused to believe the slump and was right by nine and a half points a game. The quarterback wave is its own lesson: Josh Allen (+7.2), Justin Herbert (+8.4), Jared Goff (+4.7), Mahomes, Hurts, and Stafford were all priced above their sleepy Septembers, and every one delivered the second-half surge the market was quietly promising. And notice CeeDee Lamb, Trey McBride, and Josh Jacobs living at the bottom of the list with small divergences and huge payoffs — for elite players, even a two-point market vote of confidence was worth seven points a game. Flip it over:
| 2024 — market didn’t buy it (top 20 names) | implied | actual (wk1-5) | ROS PPG | vs. early |
|---|---|---|---|---|
| Kenneth Walker III | 12.5 | 22.4 | 14.3 | −8.1 |
| Derrick Henry | 12.4 | 22.0 | 18.9 | −3.2 |
| Saquon Barkley | 15.3 | 24.5 | 21.4 | −3.1 |
| Josh Downs | 5.4 | 14.4 | 12.7 | −1.7 |
| Jauan Jennings | 7.0 | 15.5 | 13.3 | −2.3 |
| Malik Nabers | 14.4 | 22.9 | 16.5 | −6.4 |
| JuJu Smith-Schuster | 1.5 | 9.9 | 2.9 | −7.0 |
| Jayden Reed | 10.3 | 18.3 | 8.8 | −9.5 |
| Alvin Kamara | 16.1 | 23.8 | 16.2 | −7.6 |
| Tee Higgins | 8.7 | 16.1 | 19.3 | +3.3 |
| Drake London | 12.8 | 17.1 | 16.3 | −0.8 |
| Darnell Mooney | 8.6 | 15.0 | 10.7 | −4.3 |
| Brian Thomas Jr. | 9.9 | 16.2 | 16.9 | +0.7 |
| David Montgomery | 11.1 | 17.4 | 15.2 | −2.2 |
| Ja’Marr Chase | 15.6 | 21.7 | 24.6 | +2.9 |
| Jonathan Taylor | 13.3 | 18.2 | 17.2 | −0.9 |
| Kyren Williams | 14.8 | 19.4 | 15.9 | −3.5 |
| Brian Robinson Jr. | 10.8 | 15.5 | 9.2 | −6.3 |
| Zach Charbonnet | 9.8 | 14.4 | 9.6 | −4.9 |
| Wan’Dale Robinson | 9.0 | 13.8 | 9.5 | −4.3 |
17 of 20 declined. Jayden Reed was WR-heaven through five weeks and the lines never repriced him like one; sell-high windows don’t announce themselves any louder than that. The whole 2024 RB hot-start class is here — Kyren, Jonathan Taylor, Brian Robinson, Charbonnet, Montgomery — every one of them regressed to within shouting distance of his implied number. The three misses stay in the table, uncut: Tee Higgins ran straight through his fade, rookie Brian Thomas kept his hot start, and Ja’Marr Chase turned a −6 divergence into a triple crown — the reminder that this is a probability engine, not a prophecy. And read the top of the table carefully, because it teaches the signal’s most important nuance: for true outliers, “sell” means reverts toward the market number, not off a cliff. Saquon fell from a 24.5 pace to 21.4 — still elite, exactly as his 15.3 implied said he’d land closer to. The catastrophic drops live in the middle tiers: Reed, JuJu, Charbonnet, Brian Robinson — players whose entire early-season case was the part the lines refused to price. Then 2025 ran the identical experiment and got the identical answer:
| 2025 — market saw more (top 20 names) | implied | actual (wk1-5) | ROS PPG | vs. early |
|---|---|---|---|---|
| Joe Flacco | 15.5 | 6.5 | 13.4 | +6.9 |
| Geno Smith | 17.2 | 12.5 | 11.1 | −1.3 |
| A.J. Brown | 13.3 | 8.9 | 17.6 | +8.7 |
| Cam Ward | 11.8 | 7.4 | 12.5 | +5.1 |
| Ladd McConkey | 13.8 | 9.7 | 12.1 | +2.4 |
| Jameson Williams | 9.8 | 7.9 | 15.0 | +7.2 |
| Brian Thomas Jr. | 13.6 | 9.6 | 10.1 | +0.5 |
| Tee Higgins | 12.2 | 8.2 | 17.1 | +8.9 |
| Jerry Jeudy | 10.9 | 6.9 | 7.2 | +0.2 |
| Bryce Young | 16.5 | 12.7 | 14.1 | +1.4 |
| Chase Brown | 13.6 | 9.9 | 19.4 | +9.5 |
| Xavier Worthy | 12.9 | 9.4 | 7.4 | −2.0 |
| Jakobi Meyers | 13.8 | 10.8 | 11.1 | +0.3 |
| Tyrone Tracy Jr. | 9.7 | 6.7 | 11.7 | +5.0 |
| Colston Loveland | 5.3 | 2.4 | 12.1 | +9.7 |
| Brock Bowers | 13.4 | 10.9 | 16.6 | +5.7 |
| Trevor Lawrence | 17.4 | 15.0 | 21.9 | +6.9 |
| Trey McBride | 14.8 | 12.5 | 21.1 | +8.6 |
| Derrick Henry | 13.6 | 11.9 | 18.3 | +6.5 |
| Michael Wilson | 6.0 | 3.8 | 16.8 | +12.9 |
| 2025 — market didn’t buy it (top 20 names) | implied | actual (wk1-5) | ROS PPG | vs. early |
|---|---|---|---|---|
| Puka Nacua | 18.0 | 26.8 | 21.9 | −4.9 |
| Javonte Williams | 12.0 | 20.7 | 12.7 | −8.0 |
| Emeka Egbuka | 11.9 | 20.5 | 7.8 | −12.7 |
| Jonathan Taylor | 15.6 | 24.3 | 20.1 | −4.2 |
| Rome Odunze | 12.0 | 19.9 | 8.3 | −11.6 |
| Quentin Johnston | 10.0 | 17.3 | 10.6 | −6.7 |
| Amon-Ra St. Brown | 15.3 | 22.5 | 17.6 | −4.8 |
| Tre Tucker | 7.9 | 14.9 | 7.3 | −7.6 |
| Quinshon Judkins | 8.8 | 15.5 | 10.8 | −4.7 |
| Jake Ferguson | 10.9 | 17.2 | 8.5 | −8.7 |
| Kenneth Gainwell | 7.2 | 13.2 | 13.0 | −0.3 |
| James Cook | 13.2 | 19.2 | 17.2 | −2.0 |
| Deebo Samuel | 12.0 | 17.7 | 9.1 | −8.7 |
| Christian McCaffrey | 19.1 | 24.8 | 24.4 | −0.4 |
| Patrick Mahomes | 19.3 | 23.1 | 18.9 | −4.1 |
| Baker Mayfield | 17.6 | 21.1 | 13.9 | −7.3 |
| Keenan Allen | 11.2 | 15.2 | 8.9 | −6.3 |
| Michael Pittman Jr. | 11.1 | 15.5 | 10.4 | −5.0 |
| George Pickens | 12.5 | 17.7 | 16.9 | −0.8 |
| Bijan Robinson | 17.0 | 22.1 | 21.7 | −0.4 |
All twenty declined. A.J. Brown at 8.9 PPG through five weeks with the whole industry writing Philadelphia obituaries — and his lines barely budged. The market held at 13.3, he delivered 17.6 the rest of the way, and the cheapest elite-WR trade window of the season belonged to whoever read the lines instead of the timeline. Chase Brown, Trey McBride, Jameson Williams, and rookie Colston Loveland were the same trade, and the deepest cut on the whole board was Michael Wilson: a 6.0 implied against a 3.8 start, then 16.8 PPG the rest of the way — a +12.9 swing on a player available in every league in America. On the sell side, Egbuka and Odunze were 2025’s loudest early-season stories, both outscoring their own lines by 8 points a game, both cut nearly in half the rest of the way — the market called both mirages in real time, in public, for free. The QB fades landed too: Mahomes and Baker were both outrunning their own passing lines through five weeks, and both gave it back (Baker by 7.3 a game). And the superstar nuance held again in year two: CMC, Bijan, Puka, and St. Brown all “declined” only back toward their (still elite) implied levels, while the mid-tier hot starts — Egbuka, Odunze, Deebo, Ferguson, Keenan — fell through the floor. The signal doesn’t tell you a great player stopped being great. It tells you what part of a hot streak was ever real.
The group-level math, so you know this isn’t twenty cherry-picked names: computed on the complete sample — every lined player, deep-roster guys included — players the market priced 2+ points above their early production improved by +3.2 PPG rest-of-season in 2024 and +4.6 PPG in 2025. Players the market priced 2+ points below their early scoring fell by −1.2 and −1.8. The effect is monotone across every quintile of disagreement, in both seasons, and it is the closest thing to a free lunch that exists in redraft.[6]
Fair objection: maybe this only works in September, when the market still remembers August and the box scores are five games of noise. So we moved the window and ran the whole experiment again — divergence measured over weeks 5–9, graded against weeks 10–18, the exact stretch where trade deadlines and playoff pushes get decided. Roster construction is settled by then, everyone’s “seen enough.” The market still knew things the box score didn’t:
| 2024 mid-season — market saw more (wks 5–9) | implied | actual | wks 10–18 | vs. mid |
|---|---|---|---|---|
| Travis Etienne | 10.8 | 5.6 | 8.4 | +2.8 |
| C.J. Stroud | 17.4 | 12.7 | 11.3 | −1.4 |
| Deebo Samuel | 13.6 | 9.1 | 9.0 | +0.0 |
| Marvin Harrison Jr. | 11.4 | 7.4 | 12.0 | +4.6 |
| Tyreek Hill | 13.9 | 10.1 | 14.4 | +4.3 |
| Jayden Daniels | 21.4 | 17.6 | 21.8 | +4.2 |
| Sam Darnold | 18.1 | 14.4 | 19.1 | +4.7 |
| Jerome Ford | 7.2 | 3.7 | 10.2 | +6.5 |
| Jordan Love | 18.3 | 15.0 | 14.3 | −0.7 |
| Jaylen Waddle | 9.8 | 6.7 | 12.0 | +5.3 |
| Kyler Murray | 19.6 | 16.6 | 18.3 | +1.7 |
| Malik Nabers | 15.2 | 12.4 | 18.1 | +5.7 |
| Michael Pittman Jr. | 10.9 | 8.3 | 12.5 | +4.2 |
| Aaron Jones | 14.9 | 12.7 | 13.2 | +0.5 |
| Xavier Worthy | 9.6 | 7.5 | 13.9 | +6.4 |
| Jerry Jeudy | 9.6 | 7.4 | 19.8 | +12.3 |
| Jayden Reed | 13.0 | 11.1 | 8.0 | −3.1 |
| Mike Williams | 4.7 | 1.7 | 4.7 | +3.0 |
| Antonio Gibson | 6.6 | 3.8 | 6.7 | +2.9 |
| Devin Singletary | 6.0 | 3.6 | 4.8 | +1.3 |
16 of 20 rose, one held flat, three missed — and the single biggest call of either season is sitting in this table. Jerry Jeudy was scoring 7.4 a game through week 9 of 2024, an afterthought on waiver wires, and the market kept a 9.6 on him — a quiet two-point vote. He averaged 19.8 the rest of the way, the most famous second-half heater of the year, and the lines whispered it first. Marvin Harrison Jr.’s rebound, Waddle’s, Worthy’s, Nabers’ second gear — all flagged mid-season by the same gap. The misses are honest and instructive: Stroud and Love were market votes that never cashed, and Reed made the buy list mid-season after topping the sell list early — the market chased him back at 13 and he kept falling, its worst two-window read in the sample.
| 2024 mid-season — market didn’t buy it | implied | actual | wks 10–18 | vs. mid |
|---|---|---|---|---|
| Garrett Wilson | 12.8 | 22.6 | 11.9 | −10.8 |
| George Kittle | 11.5 | 20.8 | 13.8 | −7.0 |
| Joe Mixon | 14.2 | 23.4 | 14.0 | −9.4 |
| Tee Higgins | 12.6 | 20.7 | 20.1 | −0.6 |
| Derrick Henry | 14.1 | 21.6 | 16.9 | −4.7 |
| Rhamondre Stevenson | 10.4 | 17.4 | 8.5 | −8.9 |
| Jonnu Smith | 5.6 | 12.3 | 17.4 | +5.0 |
| Rachaad White | 9.9 | 16.7 | 13.4 | −3.3 |
| Rico Dowdle | 10.5 | 17.2 | 12.3 | −4.9 |
| Jalen Hurts | 20.9 | 27.4 | 19.7 | −7.7 |
| Zay Flowers | 12.4 | 18.7 | 9.4 | −9.3 |
| Jakobi Meyers | 10.0 | 16.3 | 15.2 | −1.1 |
| Mark Andrews | 6.7 | 12.9 | 14.0 | +1.1 |
| D’Andre Swift | 12.6 | 18.8 | 10.1 | −8.7 |
| Darnell Mooney | 11.2 | 17.3 | 9.1 | −8.2 |
| Kareem Hunt | 12.2 | 18.3 | 9.0 | −9.3 |
| Jahmyr Gibbs | 13.8 | 19.8 | 23.6 | +3.9 |
| Bo Nix | 15.6 | 21.1 | 20.8 | −0.3 |
| Baker Mayfield | 17.8 | 23.3 | 20.4 | −2.9 |
| Demarcus Robinson | 7.0 | 14.7 | 4.4 | −10.2 |
17 of 20 declined. Garrett Wilson was averaging 22.6 through his hot mid-season stretch and the market flat refused to move past 12.8 — he scored 11.9 the rest of the way, a −10.8 collapse the lines called to the decimal. Mixon, Rhamondre, Zay Flowers, Kareem Hunt: every one a trade-deadline temptation, every one repriced back to earth on schedule. The misses are the fun kind of honest: Jonnu Smith’s Miami takeover and Gibbs’ December were real role changes that outran the lines, and Andrews slipped the fade too — three genuine escapes out of twenty attempts.
| 2025 mid-season — market saw more (wks 5–9) | implied | actual | wks 10–18 | vs. mid |
|---|---|---|---|---|
| Jakobi Meyers | 13.0 | 7.1 | 12.0 | +4.8 |
| George Kittle | 11.9 | 7.1 | 18.3 | +11.2 |
| Jerry Jeudy | 8.9 | 4.1 | 8.1 | +4.0 |
| Darnell Mooney | 9.1 | 4.8 | 5.9 | +1.1 |
| Geno Smith | 15.1 | 10.9 | 10.5 | −0.4 |
| Bryce Young | 14.6 | 11.0 | 15.3 | +4.3 |
| TreVeyon Henderson | 8.9 | 5.6 | 17.6 | +12.0 |
| Puka Nacua | 20.3 | 17.4 | 23.7 | +6.3 |
| Baker Mayfield | 18.2 | 15.4 | 14.6 | −0.8 |
| Cam Ward | 11.6 | 8.8 | 14.2 | +5.4 |
| Xavier Worthy | 11.9 | 9.2 | 6.7 | −2.4 |
| Kenneth Walker III | 9.8 | 7.3 | 12.1 | +4.8 |
| Deebo Samuel | 13.0 | 10.7 | 10.1 | −0.6 |
| Matthew Golden | 9.2 | 6.8 | 2.7 | −4.1 |
| Nico Collins | 14.8 | 12.4 | 16.9 | +4.5 |
| Jauan Jennings | 10.0 | 7.9 | 14.0 | +6.2 |
| Travis Etienne | 12.0 | 10.3 | 16.2 | +5.9 |
| Hunter Henry | 8.9 | 7.3 | 11.1 | +3.8 |
| Adonai Mitchell | 2.4 | 0.9 | 8.2 | +7.3 |
| Cole Kmet | 3.9 | 1.9 | 5.4 | +3.5 |
Fifteen rose outright, three held roughly flat, two missed (Worthy and rookie Matthew Golden, whose situations genuinely broke). The headline is rookie TreVeyon Henderson: 5.6 PPG through week 9, an 8.9 implied — then 17.6 a game down the stretch. The market flagged the breakout three weeks before the box score confirmed it, the exact profile our belief-state study says to chase. Puka’s monster December, Kenneth Walker’s revival, Nico’s, Etienne’s, Jennings’ — all sitting in a two-to-three-point gap anyone could read on a Thursday.
| 2025 mid-season — market didn’t buy it | implied | actual | wks 10–18 | vs. mid |
|---|---|---|---|---|
| Rico Dowdle | 11.7 | 22.1 | 10.6 | −11.5 |
| Jonathan Taylor | 17.2 | 26.9 | 17.3 | −9.7 |
| D’Andre Swift | 11.4 | 20.5 | 13.0 | −7.5 |
| Ja’Marr Chase | 17.1 | 26.1 | 18.0 | −8.1 |
| Jaxon Smith-Njigba | 17.7 | 26.4 | 20.3 | −6.1 |
| Jaxson Dart | 15.4 | 22.9 | 17.9 | −4.9 |
| DeVonta Smith | 11.9 | 19.0 | 9.7 | −9.3 |
| Oronde Gadsden II | 7.8 | 14.8 | 5.5 | −9.3 |
| Tee Higgins | 10.8 | 17.7 | 15.6 | −2.1 |
| Pat Freiermuth | 4.7 | 11.6 | 6.7 | −4.8 |
| Christian McCaffrey | 20.4 | 27.0 | 23.2 | −3.9 |
| Josh Jacobs | 14.5 | 21.1 | 11.5 | −9.6 |
| Josh Downs | 8.3 | 14.6 | 6.8 | −7.8 |
| Colston Loveland | 5.7 | 11.3 | 12.5 | +1.2 |
| Dallas Goedert | 9.5 | 14.9 | 10.5 | −4.3 |
| Matthew Stafford | 17.2 | 22.3 | 21.0 | −1.3 |
| A.J. Brown | 12.4 | 17.1 | 16.7 | −0.4 |
| Travis Kelce | 10.5 | 15.2 | 9.8 | −5.4 |
| Kareem Hunt | 6.2 | 10.9 | 8.1 | −2.8 |
| Harold Fannin Jr. | 7.8 | 13.1 | 12.6 | −0.5 |
19 of 20 declined — the lone escape is Colston Loveland, who’d already proven the market right once from the buy side and simply kept climbing. Rico Dowdle is this study’s running gag: flagged as a mid-season sell in both seasons, declined on cue both times (−4.9, then −11.5). And George Kittle deserves his own footnote in market efficiency: mid-2024 the lines faded his hot streak and he fell seven points; mid-2025 they backed his cold streak and he rose eleven. Same player, opposite reads, right both times. But the best single illustration in either window is Jalen Hurts’ 2024: an early-season buy at a 21.4 implied when he was scoring 16.9, a mid-season sell at 20.9 when he was scoring 27.4 — and he finished the back half at 19.7. The number was his level the entire time. The box score just orbited it.
Full-sample group math for the mid-season window, same construction as before: players priced 2+ points above their weeks 5–9 scoring gained +2.7 PPG over weeks 10–18 in 2024 (n=28) and +3.5 in 2025 (n=21); players priced 2+ points below fell −1.8 (n=101) and −2.4 (n=97). The sell-side effect is actually stronger mid-season than early — which makes sense, because by November a hot streak has had time to get famous, and fame is exactly what the lines don’t pay for. This matches the checkpoint result from the yearly harness: re-measured at week 10, the market number still out-predicts trailing production for the rest of the season. The signal doesn’t expire after September. It re-arms every week there’s a slate.
Same instrument, sharper question. When a player spikes — last-3 scoring at 135%+ of his baseline — or craters to 70% of it, the market posts a fresh number on him within days. We call the ratio of that fresh number to his new level the belief state, and across 273 spike-and-slump events over two seasons it separates real from fake better than anything else we’ve tested:
A breakout the market repriced upward sustained 44% of the time; a breakout it faded, 19% — and in 2025 the split went fully feral: 54.5% versus 4.8%. One in twenty. The slump side is the one that saves seasons, though. A slumping player whose number held recovered 54% of the time. A slumping player whose number got cut recovered 12.5% of the time. That’s the difference between “hold, he’s fine” and “the people who watch this for a living just told you he isn’t” — and it’s checkable every Thursday for free.
Three fast structural facts before we move on. First, at the team level the market is boringly efficient: across 100+ games per franchise, teams beat their implied totals at 51–53% — a coin flip. A team total is a forecast, not an opinion, and you should treat it as one. Second, team implied totals mean-revert hard (persistence r = −0.37), which produces our most counterintuitive replicated team rule: players on offenses whose implied totals just rose 2.5+ points decline about −1.1 PPG over the next month, negative in four of five seasons. The hype offense is the fade; and no, the collapsing offense is not the buy — that side tested out neutral.[7] Third, the one honest inefficiency: after a genuine shock performance, prices react within a week (r = 0.41 first week) with zero pre-game anticipation — but the association keeps drifting into weeks two through four (r = 0.16, then 0.08). The market underreacts to real role changes. That drift window — days two through twenty after the shock, once the first reprice already happened — is the only door we’ve found that stays open.
Everything above is aggregate. This is what it looked like week to week, on real players, in the season that just ended — and the sequence is the same every time. Role moves first. The market reprices second. The box score confirms third. Managers who wait for step three are buying at the top.
| Week | snap share | implied | actual | what it meant |
|---|---|---|---|---|
| Wk 7 | 14% | 9.2 | 0.5 | rock bottom, unstartable |
| Wk 8 | 21% | 5.5 | 5.5 | market gives up too |
| Wk 9 | 75% | 11.7 | 12.7 | role explodes, line doubles |
| Wk 10 | 84% | 12.4 | 28.0 | box score finally catches up |
| Wk 11 | 88% | 14.4 | 32.3 | league-winner territory |
Read week 9 carefully, because that single row is the entire thesis of this article. His snap share went from 21% to 75%, and his implied number more than doubled from 5.5 to 11.7 — before he had scored a single big fantasy game. The market saw the role change and priced it inside seven days. Anyone reading his box score was looking at a back who’d scored 0.5, 5.5, and 12.7 in his last three — noise. Anyone reading his line was looking at a player the market had just revalued by 113%. The 28- and 32-point weeks arrived two and three weeks after the signal was already public.
| Week | targets | implied | actual | what it meant |
|---|---|---|---|---|
| Wks 1–10 | 2–7 | 5.5–6.6 | 1.5–9.1 | flat line, flat production |
| Wk 11 | 18 | 9.7 | 33.5 | line jumps 49% at kickoff |
| Wk 12 | 15 | 13.2 | 21.8 | market doubles down |
| Wk 14 | 16 | 14.6 | 37.2 | fully repriced, still delivering |
Wilson’s implied number sat between 5.5 and 6.6 for ten straight weeks — the market’s way of saying “nothing has changed here.” Then in week 11 it jumped to 9.7 before kickoff, a 49% revaluation on a player nobody was discussing. He caught 18 targets for 33.5 points that afternoon. The line moved first, on information about his role that was knowable before the game started, and it kept climbing while he posted 21.8 and 37.2. He finished as the single largest positive divergence in our entire 2025 sample.
The flip side of catching a breakout is refusing to panic-sell a star, and 2025 gave us two textbook cases. A.J. Brown opened the season with 1.8 points on one target in week 1, then 2.7 in week 4, and the discourse declared the Eagles’ passing game broken. His implied number never left the 11.9–14.9 band all season. Not once. The market’s position was that nothing was actually wrong, and it was right — he posted 28.1, 25.0, and 35.2 in the back half. Trey McBride ran the same play: a 9.1-point dud in week 5 with his line holding at 14.7, then 21.2 and 29.4 in the two weeks that followed. In both cases the box score screamed sell and the line said hold, and the line won.
Rookie tight end Colston Loveland is the third flavor — the slow-build. His snap share climbed 42% to 67% to 81% across weeks 6 through 8, and his implied number crept from 3.9 to 4.3 to 8.2 in step with it. Then came the 29.8-point week 9. The role was visible for three weeks in both the snap counts and the lines before it ever showed up in a box score.
In all five cases the actionable moment was a week-over-week jump in the implied number that wasn’t yet justified by production — Henderson 5.5→11.7, Wilson 6.5→9.7, Loveland 4.3→8.2. That jump is the market pricing a role change it can see and you can verify: snap share, target count, depth-chart news. When the line moves and the box score hasn’t, that’s the window. Once the box score confirms, you’re bidding against everyone.
Dynasty asks a different question — not “what does he score next month” but “what is he worth” — so we ran the market study on prices instead of points. Across 16,532 player-games we measured the relationship between a production surprise and the player’s dynasty value movement, in windows before and after the game.
Three findings stack up here, and the order matters. First, the dynasty market does not anticipate anything. The pre-game correlation is essentially zero at every position (QB +0.003, RB +0.011, WR −0.023, TE −0.056). That flat bar is a clean negative control — it proves the rest of this isn’t just “hot players trend up.” Nobody is front-running breakouts. Second, it reprices fast: within seven days, correlation jumps to +0.25 (RB) through +0.41 (QB). That part is the market working, and it is not an edge — by the time you’ve read the takes, the price has moved.
Third, and this is the only genuinely tradeable finding in dynasty we’ve been able to verify: the reaction doesn’t finish. The association persists into week 2 (+0.09 to +0.16) and is still alive at weeks 3–4 (+0.04 to +0.08), same sign at all four positions. That’s textbook post-event drift — underreaction. The market moves most of the way in week one and the rest of the way over the following month. The window is days 7 through 28 after a breakout game, once the obvious reprice has already happened and everyone has stopped talking about it. Not the day after. The month after.
In percentage terms, measured as excess over non-boom games by players at the same position: a single boom game moves dynasty value +4.7% (QB), +5.5% (RB), +5.3% (WR), +6.1% (TE) within seven days, then drifts another +1.0% to +1.6% over days 7–14, and stays slightly positive through day 28. The excess-versus-control framing is what makes those numbers honest — raw medians go negative in the later windows simply because the entire board drifts down over a season, which would have made a genuine drift look like a decline.
Two more dynasty results worth having. Rookie picks have a calendar. Future firsts are cheapest in June (−1.5% against their own annual median), with April and May close behind, and most expensive in February (+3.3%) and January (+2.2%) — a 4.8-point swing between the cheap end and the dear end of the year. Buy in the post-draft lull, sell into the pre-draft hype cycle. It costs nothing to verify and it repeats annually.
And the structural one: the superflex quarterback premium has collapsed. The median gap between a quarterback’s superflex value and his 1QB value has fallen from roughly 2,529 in May 2020 to about 532 by late 2025 — a 79% decline over six years, close to monotonic, with the non-quarterback spread drifting slightly negative as the expected mirror image. Six years of the market steadily deciding quarterback scarcity was worth less than it thought. If your superflex valuations still carry 2020’s instincts, they are two full eras out of date.
Matchup analysis is the most-consumed content in fantasy football. Every week, every platform, every podcast: who’s got the good matchup. So we built the defensive side properly — 24,924 player-weeks, with opponent profiles constructed leakage-free by design. For a week-9 game, the opponent’s profile uses only weeks 1 through 8. That detail sounds pedantic and it is the whole ballgame: using a defense’s full-season numbers would include the very performance you’re trying to predict, which is the single easiest way to manufacture a defensive signal that isn’t there.[10]
We tested five defensive aspects. Exactly one survived, and it’s the crudest one on the list.
| Defensive aspect | partial r | verdict |
|---|---|---|
| Opponent PPR allowed to the position | +0.021 to +0.075 | small but real |
| Opponent man-coverage rate | −0.044 to +0.035 | noise |
| Opponent pressure rate | −0.010 | noise |
| Opponent defenders in box | +0.023 | noise |
| Opponent pass-rusher count | ≈ 0.00 | noise |
The one that works is the dumb one: how many fantasy points a defense has allowed to that position. Moving a player from a bottom-third matchup to a top-third matchup is worth about +1.50 PPR at quarterback, +0.84 at running back, +0.84 at tight end, and +0.38 at wide receiver. Positive at all four positions, so the effect is real. It is also, at receiver, under half a point. That belongs in a projection as a small adjustment. It does not belong in a start/sit argument, and it certainly doesn’t justify benching a good player for a worse one.
Everything else — the detailed scheme data everyone assumes is the sophisticated version — is noise at the player-week level. Man rate, pressure rate, box count, rusher count: partial correlations between −0.044 and +0.035, with tercile spreads that flip sign across positions (running backs score 0.25 points fewer against man-heavy defenses while quarterbacks score 0.51 more). The data is real and validated cleanly against known defensive identities. It just doesn’t move fantasy outcomes.
This is the claim you hear every Sunday morning, so we tested it as a genuine player trait. Split each player’s games at his own median opponent man-rate — so a receiver who never faces man-heavy defenses isn’t being compared to one who always does. Compute his man-minus-zone split over the first half of his sample. Then check whether it predicts the second half. If beating man coverage is a real skill, it has to persist.
| Position | players | first-half split predicts second half | verdict |
|---|---|---|---|
| WR | 161 | +0.080 | no |
| RB | 97 | +0.027 | no |
| TE | 69 | −0.133 | reverses |
| QB | 38 | −0.416 | actively reverses |
It does not persist at any position, and at quarterback it inverts hard. A passer who crushed man coverage in the first half of his sample tends to do the opposite in the second half — that −0.416 is about as close to a pure mean-reversion signal as this database produces. The practical translation is uncomfortable but clean: a player’s apparent man/zone split is noise, and the splits that look most impressive are the ones most likely to regress. When you see a graphic showing a receiver’s massive edge against man, you are looking at a sample, not a skill.
We should be precise about what this does not rule out, because the honest version matters. It doesn’t rule out individual defender matchups — a true shadow corner traveling with a receiver is a different question, and the 26,710 defender-weeks of coverage data needed to answer it are ingested but not yet joined to the receivers they covered. It doesn’t rule out extreme outliers rather than terciles, the truly elite and truly hopeless units. And it doesn’t rule out effects on volume rather than points — a defense may change how a team attacks without changing the points that result. Those are the three open doors, and we’ll say so until we’ve tested them.
Here is where we have to be careful, because there are two different claims wearing the same jersey and only one of them dies.
Claim one: you can plan a season around strength of schedule. That one is dead. We tested whether upcoming-schedule softness predicts a player’s next-month scoring, controlling for the player, and at every position across five seasons the sign flips between seasons. Soft schedules “worked” some years and inverted in others — the statistical signature of noise. Playoff-schedule columns in August are astrology, for a reason that’s obvious once stated: you don’t know what the spreads will be in December. Teams you pencil in as soft are 8–3 by then.
Claim two: this week’s game script changes how a team plays. That one is real, and it is more commonly asserted than measured. We joined every player-week since 2021 to that game’s closing spread and bucketed by projected margin:
| Game script | RB PPR | WR PPR | TE PPR | QB PPR |
|---|---|---|---|---|
| Big favorite (7+) | 8.81 | 8.02 | 6.35 | 15.14 |
| Favorite (3–7) | 8.32 | 7.58 | 5.80 | 14.68 |
| Close (<3) | 7.66 | 7.27 | 5.35 | 13.82 |
| Underdog (3–7) | 7.41 | 7.29 | 5.13 | 12.88 |
| Big underdog (7+) | 6.97 | 6.28 | 4.98 | 10.78 |
Perfectly monotone at every position. Favorites outscore big underdogs by 1.84 PPR at running back, 1.74 at receiver, 1.37 at tight end, and a massive 4.36 at quarterback. So being favored is worth a ton?
No. That table is mostly a measurement of team quality, not game script. Good teams are favored, and good teams have good players. To separate the two you have to compare a player to himself: take each player-season and compare his own games as a favorite against his own games as an underdog. Same player, same offense, same year. Only the script changes.
| Position | player-seasons | PPR as favorite | volume change | % of players better favored |
|---|---|---|---|---|
| RB | 212 | +0.99 | +0.64 touches | 57% |
| TE | 171 | +0.57 | 0.00 targets | 56% |
| WR | 361 | −0.50 | −0.47 targets | 43% |
| QB | 76 | −0.55 | −2.43 dropbacks | 47% |
The 4.36-point quarterback gap evaporates completely — it flips to −0.55 once you hold the player constant. Those raw numbers were Mahomes and Allen being favored every week, not game script doing anything. Same story at receiver: within-player, wideouts are slightly better as underdogs (−0.50 PPR, −0.47 targets as favorites, and only 43% of players score more when favored). Trailing teams throw. That’s the classic garbage-time effect, and it’s real enough to cancel the favorite bump entirely.
The running back effect is the one that survives. Within-player, backs gain +0.99 PPR and +0.64 touches as favorites, and 57% of individual players score more in their own favored games. Favored teams do feed the run — that part of the folk wisdom holds up. What doesn’t hold up is the size of it. Two-thirds of a touch. It is a real, replicable, directionally-correct effect worth about one PPR point, which makes it a tiebreaker between similar flex options and nothing more. It is not a reason to start a bad back over a good one, and the raw table’s 1.84-point gap oversells the true effect by nearly half.[11]
Stack this with the defensive finding from the previous section and you get the complete matchup picture, honestly sized. A soft positional matchup is worth about +0.84 PPR to a back. Being favored is worth about +0.99. Both are real, both replicate, and both are fractions of a point next to the 5-to-10-point swings that come from role changes and market divergence. That’s the correct hierarchy: role first, market second, matchup and script as tiebreakers. Use game script on Sunday morning to break a coin-flip flex call. Never use a projected schedule in August to draft.
We built spell-level return curves from every skill-position injury absence in the database — how many games each injury type actually costs, and what the player scores relative to his pre-injury self in each game back:
| Injury | spells | median missed | miss 3+ | game 1 back | game 2 | game 3 |
|---|---|---|---|---|---|---|
| Hamstring | 198 | 1 | 14.1% | −0.6 | +2.0 | +4.8 |
| Ankle | 219 | 1 | 15.5% | −0.7 | −1.2 | +0.4 |
| Knee | 214 | 1 | 13.1% | −0.8 | −1.9 | −0.6 |
| Concussion | 174 | 1 | 9.8% | −0.1 | +0.4 | +1.7 |
| Groin | 58 | 1 | 10.3% | −0.2 | −2.4 | +5.0 |
Two things worth tattooing somewhere. Hamstrings are the market’s favorite overreaction — the first game back is soft, and then the player comes back above his prior level by game three. Knees are the quiet opposite: no bounce, three straight games below baseline, the one soft-tissue-adjacent tag where the discount is real and lasting. And the boring headline over all of it: the median absence for nearly every non-surgical tag is one game. The dead-roster-spot panic cut is almost always the wrong move by the base rates.
When a target-earner goes down, the standard analysis sprays his vacated share across the whole depth chart. The database says otherwise. Across 165 absence events, the median share of the total scoring gain captured by the single top beneficiary is 63% — and in 55% of all cases, one player takes over 60% of everything that was vacated. Absence value doesn’t distribute. It concentrates. The whole game is identifying which one — and the answer, per everything above, is the guy whose usage indicators (snap share, route participation, red-zone role) were already rising, not the biggest name on the depth chart.
We run the same replication program on college football — twelve seasons of it, 2014 through 2025, across a database of 60,987 college athletes — and on the college-to-NFL pipeline, and the latest round of studies produced findings on both sides of the bridge.
First: college production is more year-over-year predictable than the NFL. The same factor sweeps that top out around r ≈ 0.2–0.25 for NFL skill players run meaningfully hotter on college data, for a structural reason that’s obvious once you say it out loud — college rosters have wider talent gaps, returning-role continuity is stronger, and elite usage concentrates harder. The volume-persists / efficiency-reverts law holds there too; it just holds louder. If you play CFF, last year’s target share is even more of a cheat code than it is on Sundays.
Second, and this one hurt: the college-to-NFL projection problem is mostly already solved by draft night, and not by analytics. We built a full ridge-regression rookie model on 405 drafted skill players — college production profiles, combine data, departure-based landing-spot vacancy, everything knowable in April — targeting peak PPG in NFL years two and three. Then we graded it against the dumbest possible baseline: just carry the rookie season forward. The naive baseline is unbeaten. Not one model configuration cleared it decisively at any position, and the cross-validation kept asking for more shrinkage — the math itself saying “collapse me into the baseline, the extra features add nothing.”[8] Landing-spot vacancy added a real but small lift at RB and TE and none at WR.
So what does matter before a rookie plays a snap? Draft capital, brutally. Our rookie comp engine — outcome distributions for every drafted player near a given pick — puts the cliff in plain numbers: RB prospects taken in the top handful of picks hit a startable year-one season ~76% of the time with an 18% bust rate. A WR taken around pick 90: 6% hit, 68% bust. Around pick 140: 0% hit, 82% bust in the sample. The romance of the late-round sleeper is running head-first into an order-of-magnitude base-rate wall.
Here’s a brand-new study that quantifies a bias we’d only suspected. Take every drafted skill player since 2020, control for draft capital (which dominates everything), and ask two separate questions: which college stats predict how the player gets priced right after the draft — and which college stats predict how that price performs over the next two years? For wide receivers (n=158), the answer is a sentence we’ll be repeating for years: the market pays for college points and should be paying for college target share. Career college PPG is the single strongest driver of a rookie WR’s post-draft price (partial r = +0.42 beyond capital) and it predicts negative two-year returns (−0.11) — you pay full freight for gaudy college scoring and it doesn’t cash. Meanwhile final-season college target share is barely priced at entry (+0.19) and is the best two-year return predictor in the feature set (+0.16). Volume-over-efficiency isn’t just an NFL law — it’s true before the player is even drafted, and the crowd prices the wrong one. A schedule note that falls out of the same table: production padded against cupcake opponents carries a mild negative return signal. Points scored against air were never information.
The newest addition to the program: 3,580 player-seasons organized by career year, asking what a player’s prior season tells you about clearing a startable bar at each stage of his career. The spreads are enormous — these are the base rates that should anchor every “is the breakout coming” conversation:
| Career stage | observable | above threshold | below threshold | base rate |
|---|---|---|---|---|
| Year-2 WR | rookie PPG > 9.9 / ≤ 5.5 | 80.0% | 3.8% | 39.5% |
| Year-3 RB | prior PPG > 14.2 / ≤ 7.9 | 92.9% | 0.0% | 44.4% |
| Year-3 WR | prior tgt share > 20.1% / ≤ 13.7% | 82.6% | 16.7% | 45.7% |
| Year-4 WR | prior tgt share > 22.7% / ≤ 13.1% | 89.5% | 10.0% | 50.0% |
| Year-5+ WR | prior tgt share > 22.8% / ≤ 14.2% | 89.8% | 4.0% | 48.0% |
| Year-5+ TE | college final-season PPG > 11.3 / ≤ 8.1 | 89.3% | 42.4% | 53.5% |
| Year-5+ QB | college final-season PPG > 30.3 / ≤ 25.4 | 100% | 61.1% | 76.9% |
Sit with the bottom two rows. A tight end’s final college season still separates hit rates by 47 points in his fifth NFL year and beyond — and in this sample, not one year-5+ QB who’d cleared 30 fantasy points a game in college has missed the startable bar. College production doesn’t stop mattering when the rookie season starts; it echoes half a decade into a career.[9] The year-2 WR row, meanwhile, is the cleanest breakout gate we have: an 80% hit rate above a 9.9 PPG rookie season versus 3.8% below 5.5. If a rookie receiver couldn’t crack six points a game, the “wait for year two” story is a 1-in-26 prayer.
And the sharpest cut of the whole study — the part built to kill circular reasoning — is the incremental test: does a role metric tell you anything beyond last season’s points? For WRs and TEs, emphatically yes, and the effect peaks exactly where it’s most useful: in the middle production band, where players look identical on a stat line, prior target share separates next-year hit rates by up to 49 points. Among 10-PPG receivers, the one with the 24% target share and the one with the 16% target share are entirely different assets wearing the same box score. For RBs, the same test fails — every role metric flips sign across production bands. A running back’s future is his scoring plus his draft pedigree, full stop, which is why “his usage profile is quietly elite” is a real argument for a receiver and a mirage for a back.
Two honest closers on the college side. Our college weekly boom model — the machine behind CFF start/sit probabilities — beats every baseline on a blind holdout season, but modestly (AUC 0.63–0.70), and we’ll say the quiet part loudly: weekly booms are substantially irreducible noise, and anyone claiming a large edge on them is selling something. What the model genuinely delivers is calibration — when it says 30%, reality lands within about two points of 30%, and adding consensus-line features lifts WR ranking accuracy meaningfully (AUC 0.667 → 0.712), the market proving useful even inside a college model. And the calibration porn holds at the source: across 5,816 college player-weeks, consensus anytime-TD probabilities tracked actual scoring rates bucket-for-bucket — players implied at 20–35% scored 24.8% of the time, 35–50% scored 37.9%, 50–70% scored 50.3%. The lines are honest all the way down to Tuesday-night MACtion. The margin baked into them is the price of admission; the forecast underneath is real.
And once the rookie year exists, it immediately becomes the best forecast of the player — better than his college profile, better than his combine, better than our model. Which loops all the way back to the market anchor: by a rookie’s second October, his consensus lines have digested all of this for you.
Everything above describes what we’ve learned. This is what we do with it in July, when there’s no season to react to and every projection in the industry is somebody’s opinion wearing a decimal point.
Our engine blends four signals, and we’ll name all four plainly:
1. The comp signal. For every player we find his closest historical analogues — not by name or narrative, but by statistical fingerprint. A receiver is matched on eleven dimensions at once: targets, receptions, yards, touchdowns, target share, air-yards share, WOPR, RACR, separation, scoring rate, and snap share. We take that player’s ten nearest neighbors in league history and ask what those players actually did the following season. It answers “what usually happens to guys shaped like this.”
2. The naive signal. Last season, carried forward. It sounds primitive. It is astonishingly hard to beat, and we’ve published our own failure to beat it in the rookie section above. Any engine that can’t articulate why it’s deviating from last year’s number is worse than the number.
3. The market environment. The player’s team’s implied point total, measured against league average — and once real lines for the upcoming season post, we swap in the live number. This is where every offseason event you’ve been arguing about on the timeline gets priced automatically: the coaching change, the quarterback trade, the line rebuild, the schedule. We don’t need a subjective take on whether an offense improved. The market already made one, and the Vegas section above is why we trust it.
4. The touchdown-luck correction. We estimate how many touchdowns a player should have scored from his red-zone and goal-line opportunity, compare it to what he actually scored, and pull the projection back toward the opportunity. This is Law #5 — TDs revert, trips persist — wired directly into the math. It’s the single mechanism that stops the engine from paying full price for a fluke touchdown season, which is the most common way preseason rankings go wrong.
Those four get blended with weights tuned separately for each position, because the positions genuinely differ — and the tuning happens on one set of seasons and gets validated on a completely different, held-out season the model never saw. That train/test discipline is the whole ballgame. It’s the difference between a model that fits the past and a model that forecasts. Anyone can build the first one.
What we’ll say about the weights themselves: at running back the historical-comp signal carries less of the load and the touchdown correction carries more — backs churn, and their touchdown totals are the noisiest in football. At receiver and quarterback the comp fingerprint does most of the work. At tight end the market environment matters least, because target share inside an offense dominates everything else at that position. The exact coefficients stay in-house; the structure doesn’t.
And the part almost nobody else ships: every projection comes with a distribution, not just a number. We publish a 10th, 25th, 50th, 75th, and 90th percentile outcome for each player, built from the empirical residuals of how wrong this engine has actually been on players like him. When we say a back projects at 22.8 PPG, we also tell you his realistic floor is near 15 and his ceiling is near 29 — and the calibration check says our 80% band catches the true outcome 87.9% of the time. Slightly conservative, which is the direction we’d rather err. A point estimate with no band is a guess with good posture.
Two honest limitations. Rookies are the known gap — no NFL fingerprint exists to comp them, so they route through the draft-capital base rates in the college section rather than this engine, and we don’t pretend otherwise. And when the upcoming season’s lines haven’t posted yet, the market term temporarily runs on the prior year’s team environment, which we flag on a per-player basis in the output rather than quietly papering over.
“He always beats his line.” Player line-beating persistence across 58,890 graded lines: −0.055. Not zero — negative. The market over-corrects toward recent line-beaters, making them slightly worse bets going forward. There is no such thing as a chronic over-performer.
“He owns this team.” Player-versus-opponent history weight on the same 58,890 lines: −0.008. Revenge games and owner-of-the-Broncos narratives are content, not information.
Season-long schedule planning. Sign-flipped between seasons at every position — you cannot know December’s spreads in August. Note this kills season-long schedule planning only; this-week game script is real but small, and we size it honestly in the game-script section.
“Buy the collapsing offense.” Players on offenses whose implied totals slid 2.5+ points tested out roughly neutral going forward (−0.2 PPG) — not the discount-bin bounce the old heuristic promised. We used to believe this one ourselves. Five seasons said no.
Alternate lines as “value.” Across 291,000 graded outcomes, every alternate-line rung prices 2.3–6.3 points above measured reality. The longshot ladder is a margin machine at every rung. This is also why we treat only main consensus lines as forecasts anywhere in our models.
Pre-draft rookie models beating the rookie season. Ours included. See above — we published the loss.
“He eats man coverage.” Tested as a player trait with a stability check across 553 players: first-half man/zone splits do not predict second-half splits at any position, and at QB the relationship is strongly negative (−0.416). The splits that look most impressive regress hardest. See the defense section above.
Scheme-level matchup analysis. Opponent man rate, pressure rate, defenders in box, and pass-rusher count are all noise against weekly fantasy outcomes (partial r between −0.044 and +0.035). Only crude PPR-allowed survives, and at WR it is worth under half a point.
Our own dynasty age curves. We tried to measure the age cliff and got a nonsense result — median RB production appeared to rise with age (6.25 PPG at 22, 11.95 at 29). That is survivorship, not a discovery: backs who decline stop being priced and leave the dataset, so only the survivors remain to be measured. We are not publishing an age-cliff chart, because the honest answer is that this method cannot produce one. Fixing it means modeling exit as an outcome rather than as missing data.
QB efficiency as a forward signal. CPOE, passer rating, completion percentage: nothing replicates. Environment and legs or nothing.
Every entry in that box was, at some point, something we either believed or wanted to believe. The reason you can trust the green findings in this article is that the red ones are sitting right next to them.
Everything above, compressed into the rules we actually run:
1. Rank by role, not results. Target share, WOPR, snap share, red-zone share — in that order of trust. A rising role with lagging production is a buy; the reverse is a sell.
2. Fade every efficiency heater and every TD spike that isn’t backed by red-zone share. Pay for trips, never for the count.
3. From week 3 on, weight the market ~85/15 over the box score for rest-of-season. When they disagree by 2+ points per game, side with the market — that gap alone was worth +3 to +5 PPG of rest-of-season edge in each of the last two seasons, and it found Andrews, Bijan, A.J. Brown, Chase Brown, and Loveland while the box score was telling you to cut half of them.
4. Grade every breakout and slump by the belief state within a week. Repriced up: mostly real. Faded: mostly mirage. Number held through a slump: hold. Number cut: move on.
5. Fade offenses the market just fell in love with; team totals mean-revert. Use this-week game script as a tiebreaker — favored backs gain about a point, underdog receivers gain a little volume — but never draft off a projected schedule.
6. On injuries, play the base rates: median absence is one game, hamstring returns finish above baseline, knees don’t. On absences, find the one beneficiary with the pre-existing usage trend — the median winner takes 63% of the pie.
7. On rookies: respect the draft-capital cliff, ignore anyone’s model (including ours) that claims to out-predict it pre-snap, and let the rookie season overwrite everything the moment it exists.
Anyone can tell you target share matters. What you’re rarely shown is the sweep that puts a number on it, the seasons it had to survive, the market study with the named receipts, and the shelf of our own dead ideas next to it. This page is the standard every future GWTTKB piece is accountable to — if we ever publish a claim you can’t trace back to a study like these, call us on it.
Cite this report: GWTTKB Research. “What Actually Predicts Fantasy Football: Everything We’ve Measured.” 2026. https://www.gwttkb.com/research/what-actually-predicts-fantasy-football.html
Nothing here is betting advice. Consensus lines are studied as forecasts of player production, full stop. Signals, not guarantees. 21+, play responsibly.
Sources: the GWTTKB player database (NFL history to 2002; college 2014–2025, 60,987 athletes; 12 draft classes); factor sweeps (7,661 NFL player-stretches, 60–74 fields per position, replicated 2021–2025); boom/bust and ceiling/floor studies; consensus-line fantasy studies (7,141 implied player-weeks, 2024–2025; 39,693 weekly consensus line rows; two-checkpoint divergence studies at weeks 1–5 and 5–9; 349k+ graded outcomes across all line studies); breakout/falloff belief-state study (n=273 events); team-environment ledger and implied-total cascade (5 seasons); injury spell database (863 skill-position spells across tracked tags); absence-beneficiary study (n=165); rookie ridge model and comp engine (405 drafted players, 12 college crosswalk files); college factor and calibration studies (5,816 college player-weeks); career-stage expectation curves (3,580 player-seasons, 862 transitions, with incremental-information controls); capital-controlled college-pricing study (draft classes 2020+, WR n=158); college weekly boom model (blind season holdout, ECE 0.013–0.024); the game-script study (all player-weeks 2021–2025 joined to closing spreads, with within-player controls); the defensive lab (24,924 player-weeks, leakage-free trailing opponent profiles, 553-player man/zone stability test); player trait profiles (boom/bust and volatility correlations by position); dynasty lead-lag and event studies (16,532 player-games, pre-game negative control, excess-vs-control boom framing), pick seasonality and superflex-premium time series; the preseason projection engine (four-signal blend, per-position weights tuned on 2021–2023 transitions and validated on a held-out 2024 transition, empirical residual bands). All computations reproducible from the GWTTKB research database.