Clutch is
real. And
irrelevant.

Backlog #47: “The 'clutch gene' — do some players hit better in the clutch? [NULL/noise] · `Retrosheet`.” The cited source, Retrosheet's raw play-by-play, was substituted with the same play-by-play already aggregated into FanGraphs' own public Clutch leaderboard — a keyless JSON API backing the site's own leaderboard pages, no login wall, no key. Clutch, as FanGraphs defines and publishes it, is Tom Tango's formula: a player's actual win-probability contribution (WPA, adjusted for leverage) minus what his own context-neutral batting line would have produced in the same spots — a measure of whether a hitter did better or worse than himself when it mattered most, not whether he simply hit well. It requires play-by-play, which is why FanGraphs' own Clutch column is null before 1974 (confirmed by probing 1950, 1960, and 1973 — all null — against 1974, populated, before writing the fetch) — the first season of complete Retrosheet-derived win-probability coverage on this site.

Every qualified-hitter season (≥300 PA, FanGraphs' own standard “semi-regular” cutoff) from 1974 through 2024 was pulled: 12,711 player-seasons. 2020 contributes zero rows — the 60-game pandemic schedule made 300 PA unreachable for anyone, not a row dropped by hand. Consecutive-season pairs for the same player — year Y's Clutch score against year Y+1's — give 9,182 pairs across 1,799 players, 1974-2024.

The backlog called this one noise. It isn't, quite. Regressing Clutch(Y+1) on Clutch(Y): naive slope +0.0559, 95% CI [+0.0356, +0.0763] — excludes zero, p=7.7e-08. This is not a fragile result: it survives standard errors clustered by player (1,799 players each contributing several overlapping pairs, CI [+0.0353, +0.0766]), a 4,000-draw bootstrap resampling players with replacement (CI [+0.0355, +0.0764], 100% of resampled slopes positive), a 4,000-draw permutation test shuffling which player's next season gets paired with which (p=0.000 — not one of 4,000 shuffles produced a slope this extreme), trimming the 25 pairs with |Clutch|>3 in either season (CI [+0.0338, +0.0752]), and a precision-weighted refit using harmonic-mean plate appearances (CI [+0.0338, +0.0746]). Spearman agrees: ρ=0.0563, p=6.7e-08.

One real confound, checked and surviving. Clutch is supposed to net out a player's own overall quality by construction, but it doesn't do so completely: same-season Clutch correlates with wOBA at r=-0.1151 (p=9.1e-39) — better hitters score slightly less clutch by this measure, plausibly because pitchers work around them more carefully in the highest-leverage spots. Since wOBA is itself a famously persistent, stable skill, this raised the obvious worry: is the Y-to-Y+1 Clutch persistence actually just borrowed from stable hitting talent? Adding year-Y wOBA directly to the regression shrinks the Clutch(Y) coefficient to +0.0474 — 85% of the naive size — but its CI, [+0.0270, +0.0679], still excludes zero. Some of the persistence is talent-adjacent; not all of it.

But the size is the story. R² on the naive fit: 0.0031. Knowing a hitter's Clutch score this year explains 0.31% of the variance in his own score next year. A two-years-apart check (year Y vs. Y+2) finds a weaker but still zero-excluding echo, slope +0.0281, CI [+0.0063, +0.0500], consistent with a small, decaying, real signal rather than a one-lag artifact. Splitting the window at 2000 finds both eras individually clearing zero — 1974-99 slope +0.0349 (CI [+0.0059, +0.0638]), 2000-24 slope +0.0750 (CI [+0.0460, +0.1039]) — which looks like the effect more than doubled. The direct test of that says otherwise: the era-interaction term, the gap between the two slopes fit in one model, is +0.0401, CI [-0.0005, +0.0806] — contains zero, p=0.053. Two individually-significant subgroups do not make a significant difference between them; this is the textbook trap, named rather than walked into.

For color, not for the fit: among hitters with at least 8 qualifying seasons, Marcus Semien posted the lowest career-average Clutch (-1.156 over 9 seasons); Jose Vizcaino the highest (+0.831 over 8 seasons). The bottom of the list leans toward well-known sluggers — Mike Schmidt, Sammy Sosa, Giancarlo Stanton among them — consistent with the wOBA finding above, not with any of them having choked for a decade.