The Marginbeta
NFL · Research

The PolyEdge QB Rating

From the MarginRead the story behind this data We rebuilt the quarterback rating from scratch. Accuracy won.

Measuring the quarterback alone — stripped of his team, his coach, his line, his receivers, and his schedule — is the hard problem in football. Everything you see on Sunday is the QB times his circumstances. This rating is our best attempt to divide that circumstance back out and answer one question, honestly and stably: how well is this man actually playing quarterback?

It's a skill rating, not a box-score or a win total. We let ten years of play-by-play define what “good” means — points per drive, which a quarterback controls roughly twice as much as he controls his team's record — then kept only the four inputs that survived an honest, out-of-sample stress test: efficiency (EPA), accuracy (CPOE), avoiding sacks, and the value a QB adds with his legs. The result out-predicts QBR, passer rating, and CPOE, and ranks every qualified season and career 2016–2025 below.

What matters
  • Accuracy (CPOE) — the engine. 2–4× the weight of any other skill, and it survives controlling for the supporting cast.
  • Avoiding sacks — the second-biggest lever; part QB processing, part protection.
  • Rushing value — the legs are real. Crediting them made the rating both more accurate and more predictive (and finally rates Lamar and Allen correctly).
  • Efficiency, not volume — we grade quality per dropback, so a checkdown machine padding 45 throws doesn't outrank a surgeon on 28. Skill, not box score.
What doesn't
  • “Clutch” — fourth-quarter and red-zone performance barely repeats year to year. It's noise. Fade the narrative.
  • Raw arm strength — a big arm without accuracy produces bad offense (see Anthony Richardson). It matters ~4× less than completion accuracy.
  • Aggressiveness — throwing into tight windows correlates with worse outcomes; it's a symptom of receivers who can't get open, not a virtue.
  • ESPN QBR — the least stable number in football (0.24 year to year). If your take is anchored on QBR, it's anchored on sand.
What this is

One question: how well is he playing quarterback? A stable, MVP-aligned, narrative-proof measure of skill.

  • • Rank QBs against each other and history — the board below
  • • Find the cheap efficient ones before the cap catches up
  • • Fade takes built on clutch, arm talent, and QBR
What it is not
  • Not a production forecaster. Yards and attempts are volume — pace and game script (season yards R²≈0.36).
  • Not a wins predictor. Winning is ~⅔ defense and luck; the QB owns about a third (win% R²≈0.29).
  • Not a betting tool. Tested against real closing lines, it loses the vig — the market prices skill, and props are mostly volume the rating ignores.
The leaderboard — the best quarterbacks, skill isolated from circumstance

Every qualified season and career, 2016–2025, ranked against each other and against history. Showing the top 30 by default — filter by season or team, raise the minimum dropbacks, sort any column, or bump the limit to see deeper. Or flip to MVP factors to see who the voters actually crown.

30 of 355
#PlayerSeasonTmRatingPts/DrvEPACPOERushV/gSack%PsrRtWin%Dbk
1L.Jackson2019BAL135.63.390.3313.83.665.4107.150423
2A.Rodgers2020GB133.73.320.3257.20.183.7117.250543
3D.Maye2025NE132.52.910.29610.80.668.7103.667539
4P.Mahomes2018KC132.33.470.354.50.954.3108.980606
5L.Jackson2024BAL130.73.030.3414.60.694.7113.750494
6D.Brees2018NO130.23.480.2726.50.513.3111.480508
7J.Allen2024BUF129.23.270.2620.83.562.898.467497
8B.Purdy2025SF129.12.860.1947.21.83.795.967298
9J.Allen2020BUF127.53.080.2666.60.014.3102.470599
10B.Purdy2023SF127.53.040.2975.40.716106.4100469
11D.Brees2019NO1273.080.2416.10.23.1112.560386
12M.Ryan2016ATL126.53.110.2966.9-0.386.4108.925575
13P.Mahomes2022KC126.52.860.2583.61.433.8100.789677
14P.Mahomes2020KC126.12.860.2752.81.173.6103.990612
15J.Love2025GB1262.660.245.50.824.696.230461
16T.Brady2016NE125.42.890.3053.7-0.483.4108100447
17J.Goff2024DET1253.250.2765.7-0.185.5105.660568
18J.Allen2023BUF1242.520.12253.2488.157604
19P.Mahomes2019KC123.52.890.2522.50.873.4101.667502
20Aa.Rodgers2021GB123.42.960.2295.80.365.4105.875559
21R.Tannehill2020TEN123.33.130.2362.71.724.7101.267507
22D.Prescott2016DAL122.52.740.2184.21.135.299.260484
23R.Tannehill2019TEN120.72.50.1797.71.529.8105.90315
24D.Watson2020HOU120.12.510.1936.50.98.2102.60595
25A.Rodgers2016GB119.12.650.193.51.135.498.325645
26J.Allen2021BUF119.12.770.1122.13.273.988.345675
27J.Burrow2024CIN1192.710.1436.80.946.810146701
28K.Cousins2016WAS118.72.40.1675-0.183.793.519628
29J.Hurts2024PHI118.52.710.17.52.729.593.880399
30P.Mahomes2021KC118.42.890.1852.70.844.194.364687

Rating = 0.5·z(EPA/play) + 0.3·z(CPOE) − 0.15·z(sack rate) + 0.15·z(rushing EPA/game), scaled to mean 100. Qualified seasons (≥200 dropbacks), 2016–2025. Careers are dropback-weighted, ≥3 qualified seasons. Filter by season or team, click any column to sort, and the “#” column is rank within the current filter and sort.

The “MVP factors” toggle — best quarterback vs MVP

The skill rating answers “who's playing best?” The MVP factorstoggle answers a different question — who will win the award? — and the two diverge more than you'd think. The MVP model adds what voters actually reward (team record and seeding, touchdowns, rushing) on top of the skill rating, fit to the actual AP MVP vote shares 2016–2025 and leave-one-season-out validated. It's tight: the eventual MVP lands at an average rank of 1.6, in the model's top three 90% of the time — because, unlike the Heisman, the NFL MVP is a remarkably predictable, team-success-driven award.

The clearest case is 2025. On pure efficiency the rating had Matthew Stafford 6th— but he quarterbacked a 12–5 Rams team into the top seed, and the MVP factors lift him to #1 at a 66% win probability. He won. Watch the arrows in MVP mode: they show exactly how far each quarterback travels between “best” and “most valuable,” and that gap is the team success and narrative the skill rating is built to leave out.

The full story — how we defined success, the overfitting trap we walked into, the dual-threat flaw we fixed, and the day we tried to bet it and couldn't — is in the model journal: We rebuilt the quarterback rating from scratch. Accuracy won.

We took the same idea to college football — where there's no public EPA, the schedules are wildly unequal, and the Heisman turns out to be a power-conference award more than a quarterback one: the PolyEdge College QB Rating.

Observational research on nflverse play-by-play, 2016–2025. A skill rating, not betting advice — backtested against real closing lines, it shows no edge (it measures the quarterback; props measure volume the market already prices). It also fades narrative MVPs on purpose: Stafford 2025 ranks 6th, Lamar 2023 sits outside its top 5.