We rebuilt the quarterback rating from scratch. Accuracy won.

We set out to answer a question that usually gets answered with highlight reels and narrative: what actually makes a quarterback good? Not who feels great, but which measurable skills predict that a passer will keep producing next year. The rule we gave ourselves was simple and a little unforgiving — no guessing. Let the data name the factors, score them, and then check the answer against the people history already deemed great.

We pulled every play from the 2016–2025 seasons (real nflverse data), built one row per qualified quarterback season — 344 of them, 100 different quarterbacks — and went to work. What follows is the honest version: what we found, the metric we built, the flaw we caught, and the bug we caught in our own data the day before publishing.

The full sortable leaderboard — every qualified season and career, 2016–2025 — lives at the PolyEdge QB Rating page.

First, we let the data define "success"

Everyone's instinct is that a great QB wins games. The data disagrees with how much of that is him. We modeled two targets from QB skill: how many points his drives produce, and his team's win percentage.

• QB process explains about 69% of points-per-drive but only 34% of win percentage. Roughly two-thirds of winning is defense, special teams, schedule, and luck — things the quarterback never touched.
• Win percentage is also the least repeatable thing a QB does year to year, and it barely travels when he changes teams.

So we anchored everything on points per drive. It folds in the things people actually mean by "good" — it correlates 0.97 with touchdown-drive rate and 0.96 with scoring-drive rate — without crediting the QB for his defense. Wins stay on the page as context, never as the thing we optimize.

What predicts it: accuracy is the engine

We ran the full field of passing skills against points-per-drive, controlling for the supporting cast (receiver separation, yards-after-catch, and the strength of the defenses each QB faced). The standardized effects:

Two findings worth saying out loud. Accuracy dwarfs arm strength — a cannon attached to a low completion-over-expected number (see Anthony Richardson) produces bad offense. And aggressiveness is a negative: the quarterbacks who throw into tight windows most often are the ones whose receivers can't get open. The gunslinger ideal does not survive the data.

We also went hunting for more. We tested eleven additional factors — red-zone finishing, third-down play, deep-ball value, clutch performance (fourth quarter, one-score games), play-action, splitting the QB's throw value from his receivers' yards-after-catch. None of them added any predictive signal beyond accuracy, efficiency, and sack avoidance. "Clutch" in particular barely repeats from one year to the next. It is, statistically, noise. Fade the narrative.

The rating — and the confession

Here is where we have to be honest, because the first version was a trap.

We fit a fancy, forward-weighted model. In-sample it looked dominant. Then we ran a real time-split — train the weights on 2016–2021, test on seasons the model had never seen — and its ability to predict the future collapsed. It had memorized the past. We threw it out.

The thing that actually works is embarrassingly simple, which is the point — a simple blend can't overfit:

PolyEdge QB Rating = 0.5·z(EPA/play) + 0.3·z(accuracy) − 0.15·z(sack rate) + 0.15·z(rushing value)

Scaled to look like passer rating (100 is average). We bake it off against the establishment on out-of-sample data — does this year's number predict next year's production, and how stable is it season to season?

The rating is the most stable and the best forward predictor in the field. And a finding we did not go looking for: QBR is the least stable number in football — it swings so much year to year that last season's QBR tells you remarkably little about next season's. If your QB take is anchored on QBR, it's anchored on sand.

The flaw we caught: it was blind to running

The first rating had a real problem, and it was exactly the problem you'd worry about in 2026: it was built on passing and quietly punished quarterbacks who run. A buried data quirk meant scramble value was being dropped while sacks were still counted — a double penalty on the most exciting players in the league. It is why an early version had Jalen Hurts' best season outside the top ten.

So we rebuilt it to credit the legs — scrambles and designed runs, the EPA a quarterback adds with his mobility. The result is the 0.15·rushing value term above, and it did three things at once:

• It made the rating more predictive and more stable, not less.
• It made the rating agree with history more — the average MVP's rank improved from 2.6 to 2.0.
• It lifted the right players without inflating the wrong ones: Hurts 2022 went from #12 to #4, Jackson 2023 from #10 to #7, Daniels 2024 from #14 to #8, Allen 2024 from #3 to #2 — while Justin Fields' run-heavy, throw-poor 2022 stayed near the bottom, because his dropbacks were genuinely inefficient. The rating credits rushing value, not rushing volume.

The next generation of quarterbacks is dual-threat. A rating that can't see Lamar Jackson's and Josh Allen's legs is measuring the wrong sport.

Does it survive history? Yes.

The test that matters: does our number agree with the quarterbacks the game already reveres? Using MVP voting as the proxy (2016–2025):

• MVP winners average a rank of 2.0 in our rating; 88% land in the league's top three. Mahomes' 2018 and 2022, Rodgers' 2020 — all #1.
• The career leaderboard reads like a credible greats list: Brees, Mahomes, Allen, Lamar, Brady, Hurts near the top (with Brock Purdy floating up on a small, elite three-year sample — a caveat, not a verdict).
• The disagreements are the interesting part, and they rhyme. Our rating preferred Brock Purdy's efficiency over Lamar Jackson's 2023 MVP, and it ranks Matthew Stafford's 2025 MVP season sixth, behind the efficiency leaders (Drake Maye, Purdy, Jordan Love). Both were team-success and narrative MVPs more than dropback-efficiency MVPs — exactly the seasons where the story outran the numbers. That's not the rating failing; it's the rating doing its job and telling you why it sees the race differently. We treat a disagreement like that as a question to investigate, not an answer.

How it matches the money

We merged the rating against actual contracts (cap percentage). It correlates 0.51 with pay — the market roughly pays for production, but with enough slack that the misses are exploitable:

• Most overvalued: Deshaun Watson — bottom-tier production on a top-of-market contract. The rating independently flagged the most notorious deal in the league without being told to.
• Most undervalued: rookie-deal quarterbacks — Drake Maye, Bo Nix — producing at a fraction of the cap cost, before the extension reprices them.

That gap, between what a quarterback produces and what he's paid, is the single sharpest way we use this: find the cheap quarterback whose production has already arrived, before reputation and the salary cap catch up.

So can you bet it? We checked. No — and the reason matters.

This is the question we get most, so we ran the real test: bet the rating against actual closing lines (real nflverse line data, real −110 juice, no peeking — train on 2016–21, grade only on 2022–25). The result is unambiguous:

Every market loses the vig. Not "small edge" — zero edge, a coin flip that bleeds the juice. And the reason is the most important thing in this whole piece: the rating measures the quarterback; a prop measures the box score, and those are different sports.

A passing-yards prop is mostly volume — attempts, driven by pace and game script, not by how well the guy plays. We measured it directly: at the game level, the correlation between a QB's efficiency and how many times he throws is +0.005. Zero. Passing yards breaks down as attempts × yards-per-attempt, and attempts alone explain ~48% of a single game's yards while the rating's efficiency signal explains ~30% — and whatever's left, the market has already priced. The rating is built to ignore volume, because volume is circumstance, not skill. So it can be a great skill rating and a useless betting tool at the same time. It is both.

The one thing it does see early is who is playing the position best. Rebuilt week by week, the rating's top three contains the eventual MVP 87.5% of the time by Week 4 and stays there all year — it identifies the contender pool almost immediately, even though the single No. 1 doesn't lock until Weeks 12–15 (hot non-MVPs like a Garoppolo keep crowding the top). That's a statement about recognizing quality fast, not about beating a market. Don't confuse the two — we did, once, and the backtest set us straight.

The bug we caught before we published

One more, because a methodology piece that hides its mistakes is marketing. While doing a final data check, we found our leaderboard had duplicate rows — a join against an ID crosswalk had quietly multiplied veteran quarterbacks (Tom Brady appeared fourteen times). It inflated our sample and over-weighted long careers. We traced it, fixed it, and recomputed every number in this piece on the clean 344-season set. It also corrected a stat we'd been quoting internally — QBR's stability was even being mis-measured by the duplicates; the real number, 0.24, is the one above.

Nothing in the conclusions changed. But we'd rather tell you we caught it than have you find it.

One number, one question — and three it deliberately won't answer

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: how well is this man playing quarterback? It is a skill rating, and we've learned the discipline of saying what it is not, because conflating these is where every bad QB take comes from:

• It is not a production forecaster. Yards and attempts are volume — pace and game script. Even over a full season the rating tracks them only loosely (yards R²≈0.36).
• It is not a wins predictor. Winning is ~⅔ defense, special teams, and luck; the QB controls about a third of it. The rating tracks team win % at R²≈0.29 — by design.
• It is not a betting tool. We tested it against real lines and it loses the vig (above). Skill is real, but the market already prices it, and props are mostly volume the rating ignores.

What it is: the most stable, most MVP-aligned, least narrative-driven answer to "who's actually good." Use it to rank quarterbacks against each other and against history, to find the cheap efficient ones before the cap catches up, and to fade the takes built on clutch reputation, big-arm hype, and QBR. Just don't ask it who'll win on Sunday — that's a different sport, and it'll tell you so.

We didn't set out to have an opinion about quarterbacks. We set out to let ten years of data have one, to check it against the people who vote on MVPs and the executives who write the checks, and to break it ourselves — including the day we tried to bet it and couldn't. Accuracy is the engine. The legs are real. The rating knows who's good. It just doesn't pretend to know anything else.

Next, we took the same method to college — where there's no public EPA, the schedules are a free-for-all, and the rating taught us the Heisman is a power-conference award more than a quarterback one: We rated every college quarterback. Then we learned the Heisman isn't a quarterback award.