calibrated SPORTS
NFLAnalytics

What holds,
and what doesn’t

Every published measure of a player, each with its estimate, its interval and the games it rests on, and whether it carries from one week to the next. Pick a question below.

Measures
76
in 8 families
Published files
124
each with interval and n
Seasons
1999
to 2026, where the data reaches
BEST OF · USAGE: WHAT HOLDS

Carry share carries. Target share doesn’t.

Week-to-week r for carry share, [0.216, 0.258], n 1,101. Target share: −0.014, [−0.024, −0.003].

Open the measure
+0.237
Carry share
Target share
ZERO MARKED · A FILLED GREEN DOT CARRIES · A HOLLOW DOT DOES NOT
MOVES EVERY FEW SECONDS · PAUSES ON HOVER

The usage board has one home now: Fantasy.

Find a measure

76 pages live in 8 families; 3 more families are built and waiting to be published.

ReturnsReturn share and yards per opportunity.—

Returns: planned; no export is built for it yet.

How this is computed

Pick a unit, then the question you are asking. Every live family opens to its pages; the rest say what they are waiting for. Share of the shelf is a family’s pages over every live page.

The one that matters

Only carry share and carries clear zero; target share does not, and of the passing Next Gen Stats, 0 of 7 do. For the rest, this week’s surprise says nothing about next week’s: that is a finding, not a gap.

Most of what a stat line shows does not carry from one week to the next. 22 of 24 week-to-week measures published do not carry over. This is how much each measure carries over, with its interval: wholly right of zero it carries; reaching zero or left of it, it does not.

USAGE · 2 OF 6 CARRY

  • Carry sharen 1,101 · 1999–2026+0.237interval 0.216 to 0.258. Carries over
  • Carriesn 1,101 · 1999–2026+0.114interval 0.094 to 0.135. Carries over
  • Target sharen 1,540 · 2009–2026−0.014interval −0.024 to −0.003. Does not carry over
  • Targetsn 1,540 · 2009–2026−0.015interval −0.026 to −0.004. Does not carry over
  • Reception sharen 2,135 · 1999–2026−0.040interval −0.048 to −0.032. Does not carry over
  • Receptionsn 2,135 · 1999–2026−0.043interval −0.051 to −0.034. Does not carry over

PASSING NEXT GEN STATS · 0 OF 7 CARRY

11 more: receiving next gen stats, rushing next gen stats · 0 of 11 carry

RECEIVING NEXT GEN STATS · 0 OF 5 CARRY

RUSHING NEXT GEN STATS · 0 OF 6 CARRY

How this is computed

Each row is the lag-1 autocorrelation of a player’s weekly deviation from his own season mean, pooled over every player-season the file covers, with its interval. An interval wholly right of zero carries; one that reaches zero, or sits left of it, does not. A small negative r is not called a reversal: demeaning by the player’s own mean pulls it below zero by itself.

22 of 24 week-to-week measures published do not carry over.

USAGE · TARGET SHARE, BETWEEN PLAYERS55%of the variance in target share sits between player-seasonsinterval 53% to 57% · n 1,5402009–2026share of total variance sitting between player-seasons rather than within them; high means the metric separates playersMARKET · THE OVER AT THE CLOSE−2.43ppThe over is overpricedrealized minus priced · game-block 95% interval −3.13pp to −1.72pp · n 44,198 · 814 gamesresearch/market_calibration.jsonMARKET · LONGSHOT PROPS0.090R01: Do sportsbooks overprice longshot player props? Register verdict: retired.realized rate in the 0.125-0.150 price bucket (priced 0.1388) · interval 0.055 to 0.143 · n 167research/longshot.py

Inside a measure

Every page opens on its finding, then a chart of active players with the figure each rests on.

Air-yard bins, passer · 0–4 air yardsTOP 8 AND BOTTOM 3 OF 50 ACTIVE QBS
· · · 39 MORE · · ·

11 of 50 shown, the rest under the disclosure. 2006–2026 pooled · analytics/nfl/air_yards.bins.passer.json · generated Tue, Oct 27, 9:32 AM ET

WHAT EVERY PAGE CARRIES

  1. 01 The finding in one line, and whether the metric reaches the season in progress.
  2. 02 Active players by default, with a position filter; every value beside its interval and n.
  3. 03 The full table, every season the metric covers, under a disclosure, and the file it was read from.
  4. 04 For the “what holds” families: week to week, between players and naive as tabs on one page, so you see how much the answer depends on the measure.
Open air-yard bins, passer

Studies and method

18 registered hypotheses, each with its verdict: 13 retired, 2 published nulls, 2 not testable, 1 open.

  • The register every hypothesis tested, with the committed script, the interval and the verdict, the nulls included.
  • Method how every figure on these pages is computed, and what the calibration record says about it.