How many games a statistic needs before it describes a player rather than the noise around him. Snap share gets there in one game. Yards per carry needs more than a season. Touchdowns per target never arrive.
39 metrics, 1,000 bootstrap resamples, measured at exactly n games rather than extrapolated from one half. As far as we can find, this is the first split-half stabilization table published for NFL player metrics — the data has been free for a decade and nobody had done it.
Every number here says how consistently a statistic describes the same player withina season. It does not say the statistic predicts anything. We measured that separately: a quarterback's passing touchdown rate is the more reliable of two scoring metrics and was worth nothing at ranking the following week. A stat can be measured perfectly and still tell you nothing about what happens next.
Share of his team's offensive snaps, from PFR via nflverse. A coaching decision rather than a play outcome — the closest football has to being written into the lineup.
INFERRED, and an upper bound: pass plays he was on the field for, which counts anyone who stayed in to block. It is here because YPRR's reliability cannot be read without its denominator's.
Inferred denominator, per wr_routes_pg. Deflated against a true route count by the same factor as YPRR.
Inferred: the denominator is pass snaps, not charted routes, so every value sits BELOW its true YPRR — Ja'Marr Chase 2024 reads 2.49 here against ~2.85 charted. A level to be compared within this study, never quoted beside a published YPRR.
Expected touchdowns from field position, down, distance and where the ball was thrown, with the player removed. Model xtd-v1; the air-yards term is what a field-position lookup cannot have.
21 seasons (1999–2025, excluding 2003–2008) — nflverse carries no targets for 2003–2008 (confirmed upstream)
Share of his team's offensive snaps. For a back this is the committee question stated directly.
Expected touchdowns from field position, down and distance, with the player removed. Model xtd-v1, calibration on held-out seasons in src/data/nfl-xtd-model.json.
27 seasons (1999–2025) — carries and rushing yards are complete back to 1999
Expected touchdowns from his own carries. ⚠️ A game with no carries is a real zero, not a missing value — he had no rushing scoring chances, which is exactly the thing being measured.
Where his runs happen rather than how many. Nearly orthogonal to carries per game, which is why it is measured separately.
The expected touchdowns his throws created, credited to him as well as to his receivers. Never summed with rec_xtd — see migration 079.
27 seasons (1999–2025) — carries and rushing yards are complete back to 1999
Every published NFL reliability figure we could find is a year-over-year correlation. That instrument confounds three different things: measurement noise, real change in the player, and real change in his environment. A receiver whose target share moved because he changed teams is not evidence that target share is a noisy statistic.
Splitting a single season holds the environment fixed. Measured on the same players, both ways, and projected to the same length: split-half reads higher on 39 of 39 metrics, by as much as +0.307 (QB rushing xtd per carry). The gap is a measure of how much of a player's year-to-year change is the player, and how much is everything around him.
Yards per carry is the clearest case. Its published year-over-year R² of ~0.11 is real — but split-half puts it at 0.500 at season length. It is measured far better than year-over-year implies, and still needs a full season before it says anything.
Raw touchdowns are the least stable thing in football, so we replaced them with expected touchdowns — the chance each carry or target had of scoring, from where and when it happened, with the player removed. Each one is measured against the raw rate it replaces, on identical rows.
It works for receivers and running backs, roughly halving the sample needed. It fails for a quarterback's throws: his raw passing touchdown rate is more reliable than the expected version. A passer's touchdown rate carries skill that field position cannot see — which is exactly what the expected version removes by design.