Completion percentage over expected still isn't all quarterback
Completion percentage over expected arrived as the smart correction to a dumb number. Raw completion percentage rewards checkdowns and punishes downfield passing, so analysts built a model that estimates the probability any given throw is completed — based on depth, distance to the sideline, receiver separation, pressure, and more — and then measures how much better or worse a quarterback does than that expectation. CPOE is a real advance, and it correlates with good quarterback play more tightly than the raw number ever did. But somewhere between the research papers and the broadcast graphics, a useful adjustment hardened into a verdict, and CPOE started getting quoted as a clean readout of quarterback accuracy. It is not clean. It is a shared statistic that still hands one player credit for a play that took eleven.
What CPOE actually measures
CPOE takes the outcome of a pass — completed or not — and subtracts the completion probability the model assigned before the ball arrived. Sum those residuals across a season, divide by attempts, and you get a number that says how much more often a quarterback completes passes than a league-average passer would in the same situations. A positive CPOE means beating the model; a negative one means falling short of it.
That framing fixes the biggest flaw in raw completion percentage, which is that it treats a five-yard checkdown and a twenty-yard throw into coverage as equivalent successes. By pricing in difficulty, CPOE stops punishing quarterbacks for throwing hard passes. What it does not do is isolate the quarterback from everyone else involved in whether those hard passes were caught.
The model is only as good as its inputs
CPOE is not a measurement. It is the output of a model, and a model can only adjust for the variables it contains. Public versions know depth of target, field position, and sometimes pressure, but many do not know receiver separation, coverage type, or whether the throw was a designed layup against a soft zone. Every difficulty factor the model omits gets silently absorbed into the quarterback’s residual as if he earned it.
That means two quarterbacks with identical CPOE can have faced completely different degrees of actual difficulty, and the number will not tell you which. A passer whose scheme manufactures easy throws the model rates as hard will post a flattering CPOE he did not earn. The metric is precise about the wrong thing: it is exact arithmetic performed on an estimate.
Different public and proprietary CPOE models also disagree with one another, because they include different variables and weight them differently. When two credible models rank the same quarterback several points apart, the disagreement is not a rounding error — it is a reminder that CPOE is a modeling choice, not a fact about the throw. The version quoted on a broadcast is one estimate among several, and the confidence with which it is cited rarely reflects that.
Receivers finish what the quarterback starts
A pass is completed by two players, and CPOE credits only one of them. A perfectly placed ball that clanks off a receiver’s hands is scored as a quarterback failure — the model expected a completion, the throw was accurate, and the residual goes negative anyway. A slightly errant ball that a great receiver reaches back and reels in is scored as a quarterback success. Drop rate and catch skill live entirely on the receiving side of the line, and CPOE cannot see the difference.
Over a full season the noise from receiver hands does not fully wash out, because quarterbacks do not have random receivers — they have the same targets every week. A passer throwing to reliable hands inherits a persistent CPOE tailwind, and one throwing to a drop-prone group carries a persistent drag. The number attributes both to the quarterback because he is the only player it is built to grade.
Scheme manufactures the easy throw
Modern passing offenses are designed to make completions cheap. Play-action, motion, and route concepts that flood a zone create wide-open receivers, and a wide-open receiver is an easy completion the model may still rate as moderately difficult based on depth alone. A quarterback in a scheme that engineers separation gets to bank the difference as personal accuracy.
Quick-game offenses do the same thing from the other direction, trading depth for a rhythm of high-probability throws that pad the completion side of the ledger. None of this is a knock on the quarterbacks — executing a good scheme is a skill — but it means CPOE is measuring the quarterback and the coordinator together. The graphic names only the quarterback.
Sample size and the noise floor
CPOE is a small difference between two large numbers, which makes it volatile in the samples people quote it in. Over a few games the metric swings wildly on a handful of contested catches and drops, and mid-season CPOE leaderboards routinely feature names that will not survive to January. It takes most of a season, sometimes more, for a quarterback’s CPOE to stabilize into something predictive.
This matters because CPOE gets cited weekly as if each new reading were a firm measurement. It is not. Early-season CPOE is mostly noise, and treating a four-game figure as a settled fact about a quarterback’s accuracy is exactly the sample-size mistake the analytics movement was supposed to cure.
The volatility also interacts with everything above it. Because a season of dropbacks is a modest sample for a stat this sensitive, a few weeks of good or bad receiver play, favorable or hostile scheme, and simple variance can push a quarterback’s CPOE well off his true level before it settles. The number needs both time and context before it earns the weight people put on it.
What to read instead
CPOE is most honest when read as one input among several rather than a standalone grade. Pair it with the model’s own uncertainty and a full season of attempts before trusting the number. Read it beside receiver drop rates and average separation to gauge how much help the quarterback got, and beside pressure and time-to-throw data to see how much the offensive line and scheme shaped the throws he faced.
Expected points added per dropback and success rate answer the question that ultimately matters more than accuracy — whether the completions actually moved the team — and they contextualize a quarterback who completes a lot of passes that go nowhere. CPOE tells you he hit the throw. The value stats tell you whether hitting it was worth anything.
The honest read
Completion percentage over expected is a genuine improvement on the raw number, and quarterbacks who post strong CPOE over full seasons are, more often than not, accurate passers. The metric deserves its place in the toolkit. What it does not deserve is the clean-readout status it has quietly acquired, because it still bundles receiver hands, scheme design, and modeling assumptions into a figure printed under one name.
Read CPOE as evidence, not proof — a signal that gets stronger with sample size and weaker the more you learn about who a quarterback throws to and what his coordinator asks of him. It is the best widely available accuracy proxy we have. It is still not the quarterback alone, and anyone quoting it as though it were is crediting one man for a completion that took a whole offense.