Research notes
Sticky stats: what actually carries over from one season to the next.
July 11, 2026 · 7 min read
Every August argument is secretly a bet about repeatability. He’s a touchdown machine. He’s due for regression. He’s always hurt. Each of those claims assumes that what a player did last season tells you something about this season. For a few stats that’s true. For most of the ones people argue about, it barely is. We measured it, and the answer shapes every projection on our board.
The measurement
Take every pair of back-to-back seasons from 2015 through 2025. For every player who cleared a volume floor in both years (100 pass attempts for QBs, 50 carries for RBs, 30 targets for WRs and TEs), line up his stat in year one against the same stat in year two and compute the correlation. That gives 300 to 1,400 player-season pairs per stat, depending on position.
The correlation (r) runs from 0 to 1: at 1.0, last season tells you everything; at 0, it tells you nothing. The honest way to read it is to square it, which gives the share of next season’s number that this season explains. A stat at r = .72 explains about half of next year. A stat at r = .12 explains about 1.5%, which is to say the argument it powers is a story, not a forecast.
Pearson r between a player’s stat in one season and the same stat the next season. Longer bar = more of it carries over.
Full table with sample sizes
| Stat | Pos | r | Explains | Pairs |
|---|---|---|---|---|
| Target share | WR | 0.72 | 52% | 820 |
| Carries per game | RB | 0.65 | 42% | 469 |
| Target share | TE | 0.65 | 42% | 302 |
| PPR points per game | QB+RB+WR+TE | 0.61 | 37% | 1,159 |
| Target share | RB | 0.54 | 29% | 259 |
| Pass attempts per game | QB | 0.50 | 25% | 311 |
| PPR points, season total | QB+RB+WR+TE | 0.47 | 22% | 1,159 |
| Completion rate | QB | 0.44 | 19% | 311 |
| Catch rate | WR | 0.43 | 18% | 820 |
| Yards per attempt | QB | 0.42 | 18% | 311 |
| Yards per target | TE | 0.38 | 14% | 302 |
| Pass TD rate | QB | 0.37 | 14% | 311 |
| Catch rate | TE | 0.30 | 9% | 302 |
| Games played | QB | 0.27 | 7% | 304 |
| Yards per target | WR | 0.22 | 5% | 820 |
| Yards per carry | RB | 0.20 | 4% | 469 |
| Interception rate | QB | 0.17 | 3% | 311 |
| Receiving TD rate | TE | 0.17 | 3% | 302 |
| Games played | TE | 0.14 | 2% | 222 |
| Rushing TD rate | RB | 0.12 | 1% | 469 |
| Receiving TD rate | WR | 0.12 | 1% | 820 |
| Games played | WR | 0.12 | 1% | 687 |
| Games played | RB | 0.07 | <1% | 451 |
Opportunity is the signal
The stickiest thing in fantasy football is not a talent stat. It’s a role. A wide receiver’s share of his team’s targets correlates at .72 year over year, explaining about half of next season’s share all by itself. Carries per game for running backs is close behind at .65. Coaches keep feeding the same players until a departure, an injury, or a draft pick forces them not to, and the data shows it: usage is a decision that organizations repeat.
Points per game sits at .61, and it inherits nearly all of that from volume. Note the gap to season-total points at .47: a chunk of what looks like players “breaking out” or “falling off” between seasons is nothing more than games played moving around, which we’ll get to.
Efficiency is mostly noise
Now the stats people actually argue about. Yards per target: .22, explaining 5% of next season. Yards per carry: .20, explaining 4%. A running back’s yards per carry this year is very close to useless for predicting his yards per carry next year, which is why “efficient runner” is one of the least bankable labels in the sport. Catch rate (.43) and completion rate (.44) hold up better, real skill diluted by circumstance.
Touchdown rates are the bottom of the table. A receiver’s touchdowns per target correlates at .12 year over year; a running back’s touchdowns per carry, the same. Squared, each explains about 1.5% of next season. The player who scored 12 times on 90 targets had a real season, and it counted. It just isn’t evidence about next year. Passing TD rate is the sturdiest of the family at .37, which makes sense: a quarterback’s sample is five times anyone else’s. It still explains only 13%. Interceptions, at .17, are nearly as random as touchdowns.
The Burrow problem
This is where the finding gets uncomfortable, so let’s put the hardest case on the table: Joe Burrow, and why our board is lower on him than the market.
| Season | Games | Attempts | Pass TD | TD rate | League |
|---|---|---|---|---|---|
| 2020 | 10 | 404 | 13 | 3.2% | 4.9% |
| 2021 | 16 | 520 | 34 | 6.5% | 4.5% |
| 2022 | 16 | 606 | 35 | 5.8% | 4.2% |
| 2023 | 10 | 365 | 15 | 4.1% | 4.1% |
| 2024 | 17 | 652 | 43 | 6.6% | 4.6% |
| 2025 | 8 | 259 | 17 | 6.6% | 4.8% |
Burrow’s touchdown rate has beaten the league by 30 to 45 percent in four of his last five seasons. He is exactly the player a regression-based model handles worst, because the model’s job is to ask: across every quarterback in the sample, how much of an elevated TD rate survives to next season? The answer, fitted on 311 quarterback season-pairs, is “not much.” So the board pulls his rate most of the way back toward the league and projects him at 24 passing touchdowns, not the 33 to 35 his personal history implies. Add that he contributes almost nothing as a runner (about 180 projected rushing yards, against roughly 340 for Josh Allen or Lamar Jackson) and the top-five QB case collapses.
Is the model wrong about him? Maybe. Sustained elite TD rates exist: the Rodgers and Mahomes primes were real, and Burrow’s five-year record looks more like a skill than a streak. But for every quarterback who sustains one, the sample holds several who gave it all back the next year, and a model that chases the Burrows also chases the mirages. We backtested that tradeoff over four seasons and betting on regression wins on average, so that’s the bet the board makes, with the cost named out loud: when a true exception comes along, we will be low on him. If you believe Burrow is the exception, the board is showing you exactly where your edge over the model is, and drafting him is a coherent, priced decision instead of a vibe.
The least sticky stat we measured
One row of the chart deserves its own headline: games played. For running backs, the year-over-year correlation is .07. Squared, that explains half of one percent of next season. “Injury prone” as a draft-day concept is built almost entirely on a stat with no year-to-year signal at the two positions where it drives the most decisions. Quarterbacks are the partial exception (.27), and that nuance shows up in how we handle them.
This finding is load-bearing for the whole board: it’s the reason we project established starters at a healthy season instead of discounting them by their injury history. That decision gets its own research note.
What the model does with this
None of this is commentary bolted onto the projections after the fact; it’s the architecture. The model spends its effort on the sticky layer, projecting team volume and each player’s share of it, and regresses the noisy layer toward position norms. How hard each rate gets regressed isn’t chosen by feel: a shrinkage constant per stat is fitted by backtest, and the fitted constants independently agree with the table above. Yardage rates keep a meaningful share of a player’s own history. Touchdown rates keep very little, because every time the backtest was allowed to choose, it chose to trust them less. The full system, and its scorecard against the expert consensus, is on the methodology page.
Fine print: correlations are pooled over the ten consecutive-season pairs from 2015-16 through 2024-25, full PPR where points are involved. A pair requires the player to post qualifying volume in both seasons, so retirements and washouts drop out; the games-played rows require the floor only in year one. Both choices flatter the correlations if anything, since the players who vanish entirely are the extreme case of non-repeatability. Derived from nflverse play-by-play aggregates via the same pipeline that builds our projections.