The AlmanacGame Theory of Play-Calling · nflverse 2016–2025

The Predictability
Paradox

Every NFL offense telegraphs its run–pass tendencies to some degree, and some are far more readable than others. Conventional wisdom says the predictable ones pay for it. They don’t — and the number that actually separates good offenses from bad is a different one entirely.

0.51–0.78
range in offensive legibility (AUC)
+.06
predictability’s effect on EPA — the paradox
+.25
what aggressiveness (PROE) is worth
+.10 AUC
the tell hiding in your formation
ACT I

How predictable is measured

Take every early-down snap in a competitive game and ask what the situation alone — down, distance, field position, score, clock — says a team should do. A legible offense does exactly that: its pass rate climbs steeply with the situation. An unpredictable one stays flat — running when you’d pass, passing when you’d run.

Kliff Kingsbury’s 2022 Cardinals are the most legible offense of the last decade — a near-perfect diagonal. Kyle Shanahan’s 2016 Falcons, the best offense in football that year, are nearly flat. Same league, opposite legibility.

Actual pass rate by what the situation calls for · neutral early downs
0%25%50%75%100%≤44%44-48%48-52%52-65%>65%situation says pass →↑ actual pass rateLeagueATL 2016ARI 2022

Legibility is the steepness of this line, scored as the AUC of the league tendency model against a team’s calls. It is not the same as how often a team passes: ARI 2022 and ATL 2016 threw at similar rates overall — they differ in when.

Actual pass rate by situational expected-pass bin
Situation≤44%44-48%48-52%52-65%>65%
League — “what the situation says”34%45%49%56%75%
ATL 2016 — unpredictable (a top offense)55%44%57%58%76%
ARI 2022 — most legible22%36%51%75%87%
Most predictable · 2016–25
1LAC0.70
2TB0.70
3NE0.69
4ARI0.68
5WAS0.68
6LV0.68
Least predictable · 2016–25
32DET0.63
31PIT0.63
30PHI0.63
29CHI0.64
28BUF0.64
27DEN0.64
ACT II

The paradox: legible ≠ beaten

Here is every team-season on two axes of identity: how legible it is (left–right) and how aggressive it is — pass rate over expected, or PROE (bottom–top). Each dot is colored by how good the offense was. If predictability were the tax everyone assumes, the red would pool on the left. It doesn’t — what faint order the color carries runs top-to-bottom, with aggressiveness, not left-to-right.

Each dot is a team-season, 2016–2025 · color = offensive EPA per play (redder is better)
worse → better offense (EPA)
0.550.600.650.700.75−6.0−3.00.0+3.0+6.0← unpredictable · LEGIBILITY (AUC) · predictable →↑ passes more than expected (PROE) ↓ lessATL 2016ARI 2022
Legibility → offensive EPA+.06[−.03, +.16]no effect — the paradox
Aggressiveness (PROE) → EPA+.25[+.15, +.35]this is what matters

Legibility and success are statistically independent (r = +.06, its 95% interval straddles zero). PROE and success are not (r = +.25). And the two axes don’t move together (r = +.04) — they are genuinely separate dimensions of who an offense is. Single dots carry ~±.03 of sampling noise; read the field, not the point.

Most and least legible team-seasons with PROE and offensive EPA
Team-seasonLegibilityPROEOffensive EPA
ARI 20220.78+1.7−.067
TB 20190.77+0.6−.091
WAS 20240.76+2.3+.011
TB 20200.76+1.9+.087
TB 20160.76−2.3−.011
ARI 20210.75+5.3+.019
DET 20240.55−5.1+.118
WAS 20230.55+10.0−.069
ARI 20230.55−7.0+.011
CHI 20250.54−1.0+.078
CHI 20220.54−11.5−.032
SF 20160.51−6.1−.091
ACT III

What actually gives you away

If tendencies don’t cost you, what does a defense read? Not the things fans obsess over. Train a model on the pre-snap picture and only one family of cues moves the needle: alignment — are you in the gun, how many backs, pistol or under center. Motion and personnel grouping add essentially nothing.

Predicting pass vs run · holdout AUC, trained 2022–23 and tested 2024–25
xpass only0.66
+ shotgun0.71+.04
+ backfield, QB alignment0.76+.05
+ motion, personnel0.76+.00
QB alignment (shotgun/pistol/under center)+.039signalBackfield count (empty/loaded)+.017signalPersonnel grouping+.001noisePre-snap motion-.001noise

Pre-snap alignment adds +.10 AUC over the situation alone, and about +.05 of that is beyond simply being in the shotgun. Motion and personnel — the classic “tells” — test as noise once alignment is known.

Pass rate by backfield
0 (empty)93%
155%
233%
3+23%

An empty backfield is the loudest tell in football.

Pass rate by QB alignment
Shotgun68%
Under center34%
Pistol28%

The gun screams pass; under center and pistol both lean run — pistol most of all.

ACT IV

Every offense, decade by decade

Legibility is a coaching fingerprint — it moves with coordinators, not rosters. Pick a team to trace its predictability and aggressiveness across the last ten seasons, and see the one alignment that gives it away.

ARI · legibility & aggressiveness by season
0.500.600.700.80'16'17'18'19'20'21'22'23'24'25— legibility- - PROE
ARI legibility, PROE and EPA by season
SeasonLegibilityPROEEPA
20160.73−1.5+.041
20170.72−4.1−.073
20180.64−7.6−.180
20190.61+4.3+.027
20200.73−0.0+.002
20210.75+5.3+.019
20220.78+1.7−.067
20230.55−7.0+.011
20240.64+2.4+.109
20250.67+8.1−.027
ARI · career legibility 0.68

Biggest tell: 0 (empty) 95% pass.

0 (empty)95%
154%
242%

Pass rate by backfield count, 2022–2025.