The QB stat that actually predicts winning
The NFL has spent fifty years trying and failing to find a single number that summarizes quarterback play. Passer rating, the league's official answer, is a formula written in 1973 that nobody can explain on the spot and that gives equal weight to four inputs that emphatically don't deserve equal weight. Wins as a starter, the broadcast's favorite shortcut, credits a quarterback for the work of his defense and offensive line. Touchdowns and yards reward volume more than quality. Completion percentage is meaningfully affected by how aggressively a passing game throws downfield, which is the exact decision the stat should be neutral on.
None of those are useless. All of them, used in isolation, will mislead you. The metric that gets closest to what actually correlates with winning is called expected points added per dropback, or EPA per dropback, and it has been the gold standard inside front offices for at least a decade. The reason it hasn't taken over the broadcast is mostly that it can't be expressed in twenty words.
What EPA per dropback measures
EPA, in any context, is the change in the expected value of a possession before and after a single play, measured in points. A first-and-ten at midfield has an expected point value — say, two points, given how often offenses score from there. If the next play is a 25-yard completion, the new field position has a higher expected point value, maybe three and a half, and the play is credited with one and a half EPA. If the next play is a sack, the expected value drops, and the play is credited with negative EPA. Every play in the NFL has an EPA value attached to it, computed against a baseline of league-average outcomes from that situation.
EPA per dropback aggregates that across every passing play a quarterback was involved in, divides by the number of dropbacks, and produces a number that says, on average, how many points a single passing decision was worth. League average is roughly zero, because the baseline is league average. The best quarterbacks in any given season land between 0.20 and 0.30. The worst land negative. The spread between the two is much larger than the spread in passer rating across the same players, which is part of why EPA per dropback is more useful — it has more resolution.
Why it correlates with winning
Two reasons. The first is that EPA is built on the right currency. Football is a points game; statistics that are not denominated in points always require an extra translation step before you can use them to predict points. Passer rating's mix of completions, touchdowns, interceptions, and yards is correlated with EPA, but each component is correlated with different strength. EPA collapses all of them into the only output that matters.
The second is that EPA implicitly handles down and distance. A seven-yard completion on third-and-six is more valuable than a seven-yard completion on second-and-fifteen, even though both produce the same line in the box score. The first kept the chains moving; the second extended a punt-likely possession by half a first down. Passer rating doesn't see this. EPA does, because the expected point value at the start of each play already encoded the down and distance.
What it doesn't capture
Offensive line, mostly. A quarterback under pressure on 35% of dropbacks will have lower EPA per dropback than the same quarterback under pressure on 25%, all else equal — but most of the difference is the line. Public EPA models can't fully disentangle that. There are pressure-adjusted versions in front offices that try to, but the public versions effectively credit the quarterback with the line's contribution. This is the same criticism that applies to most quarterback stats, including the bad ones, so it's not unique to EPA, but it is a real caveat.
Receiver play matters too. A wide-open receiver dropped pass is a negative EPA event credited mostly to the quarterback. A contested-catch touchdown is a positive EPA event also mostly credited to the quarterback. The advanced versions of the model try to assign some of that to the catcher, and the next generation of public stats — completion percentage over expected, for example — is doing pieces of the same disaggregation. None of them is fully clean.
And rushing matters. EPA per dropback is a passing stat. A scrambling quarterback who gains forty yards on a play that started as a passing concept gets credit for it in some versions of the stat and not in others. The cleaner versions count scrambles, designed runs, and passes separately. The simpler versions don't. This matters more for some quarterbacks than for others — Lamar Jackson and Josh Allen have a meaningfully different rank order depending on which version you read.
What about wins?
Wins-as-a-starter is the most common quarterback shortcut on broadcast graphics, and it should be on the same list as touchdowns- thrown-against-a-soft-defense for "stats that are easy to find but should not be used as a verdict." A quarterback's win record depends on the quarterback, sure, but it also depends on the defense, the run game, the kicker, the schedule, the weather, and the random variation that comes with playing only seventeen games a year. The correlation between win percentage and EPA per dropback is real but loose. Many of the highest-EPA quarterbacks of the last decade played for teams that lost a lot of games. Their passing was the only thing keeping those teams competitive.
The clean way to use wins is at the team level, not the quarterback level. A team's win expectancy is closely tied to its net offensive EPA, which is closely tied to its quarterback's EPA per dropback. The quarterback is the largest single input. He is not, however, the only input, and pretending he is by attaching his name to the team's record produces persistently bad evaluation.
The five-second read
When someone tells you a quarterback "won twelve games this season," ask them what his EPA per dropback was. They probably don't know. That's fine — most people don't, including most broadcasters. But the answer, if you can find it, will tell you much more about how the quarterback actually played than the win column does. Above 0.20 EPA per dropback is genuinely excellent. Between 0.10 and 0.20 is solid starting-level play. Between zero and 0.10 is mid-tier. Below zero is bad, regardless of what the team's record was. This holds across every era the public play-by- play data covers.
A second useful number, which the public sites also publish, is completion percentage over expected. It tells you whether a quarterback completed more passes than a league-average passer would have given the same throws — same depth, same defenders, same situation. It's a tighter measurement of the quarterback's accuracy specifically, separate from receiver quality. Used together, EPA per dropback and CPOE produce a pair of numbers that diagnose the quarterback's effectiveness and decision-making better than any traditional stat does individually.
Both numbers are publicly available. None of them require a subscription or a team's tracking-data feed. They just don't fit on the lower-third graphic that runs during a broadcast. That's the actual reason passer rating and wins-as-a-starter still get the airtime. The math has been done. The communication is the part still catching up.