Do player situational splits repeat? (2026-10-20)

Question. Highlight clips love a stat like "QB X has +0.40 EPAexpected points addedHow much a play changed the number of points its team should expect to score on that drive. A better measure of a play than raw yards./dropback against the blitz." Is a split like that a trait of the player, or a description of one season?

Data. nflverse play-by-play, regular seasons 2016–2025, with FTN charting (play action, pocket, blitz, box count, screens, motion, RPO; 2022 on) and NGS participation data (pressure, coverage, time to throw, personnel; 2016–2025). Built by scripts/nfl/build_player_plays.py; computed by build_repeat() in [redacted]; live at /nfl/stats.

Method. For each player-season with at least half a season's qualifying volume (7 dropbacks, 3.1 carries or 0.94 targets per team game × 17), and at least 30/20/12 plays both in the split and outside it:

excess = (player EPA/play in the split − player EPA/play in the rest of that split's plays)
       − (the same gap for the whole league that season)

Then correlate a player's excess in season t with his excess in t+1, pooled over all consecutive pairs. Removing the league gap means a split doesn't "repeat" just because everyone is worse under pressure; subtracting the player's own rest-of-split EPA means it doesn't repeat just because good players are good everywhere.

Result

Almost nothing repeats. The median year-to-year r is +0.01 for passers (78 split values), +0.02 for rushers (53) and +0.02 for receivers (67). Of ~200 split values, nine have a 95% interval clear of +0.03 with 50+ pairs, and they are mostly mechanical:

Splits that don't repeat (95% interval includes zero) include blitzed vs not (−0.15, CI −0.35 to +0.06), play action (−0.03), in/out of pocket (+0.06), pressure (+0.04), man vs zone (+0.01), shotgun vs under center (0.00), home/away, every weather split, quarter, two-minute, score and win probability. A couple lean negative (3rd & 7+ −0.22, leading by 1–8 −0.21), which is what regression to the meanregression to the meanExtreme results tend to be followed by more ordinary ones, purely by chance. Mistaking this for a real decline is a classic error. looks like after selecting extreme seasons.

Rushing with 8+ in the box looked sticky (r +0.34) but has only 41 pairs (CI +0.03 to +0.58); worth rechecking as FTN seasons accumulate.

What this means for the site

• /nfl/stats highlights show "Splits that hold up" only for splits that clear this test, and tag every "Biggest swing" with its split's r ("mostly noise" vs "repeats").
• The swings list also prints how many |z| ≥ 2.5 gaps chance alone would produce. 2026 through week 6: 31 found from 3,073 checks, ~38 expected by chance. 2025 full season: 194 from 11,121, ~138 expected, so a modest excess, most of which is still noise.
• Practical reading: a player's overall EPA/play, sack rate and depth profile carry forward; his "against the blitz" or "in the 4th quarter" number mostly does not.

Caveats

• FTN splits have only three season pairs (2022→23→24→25), so their intervals are wide.
• Pairs require the same player to qualify in both seasons, which drops injured and benched seasons; that selection would, if anything, inflate repeatability.
• One-season-ahead only; no attempt to model aging or scheme changes.

Full table

Pairs = player-season pairs; holds = 95% CI95% confidence intervalThe range the true value is plausibly in. If this range includes zero, we cannot rule out that the real effect is nothing at all. lower bound > +0.03 with 50+ pairs. Two-value splits are listed once (the gap is the same story from either side).

Passing (EPA/dropback)

Rushing (EPA/designed run)

Receiving (EPA/target)