CPOE (Completion Percentage Over Expected)

What it measures

Raw completion percentage rewards quarterbacks who throw short. CPOE removes that excuse. The nflverse model estimates, for every attempt, how likely an average passer is to complete that exact throw (air yards, field position, down and distance, and so on), and CPOE is the passer's actual result minus that expectation. A checkdown artist completing 70 percent of layups can post a negative CPOE; a downfield thrower completing 66 percent of hard throws can post a strongly positive one.

By construction the league average is close to zero. Sustained seasons above +3 are excellent; below -3 is a real accuracy problem.

How it is computed here

The mean of the play-by-play cpoe value across a passer's attempts that carry a CPOE estimate (not every attempt does; spikes and batted throwaways may lack one). We do not re-model anything: the expected completion probability is the nflverse model's output, and our column is a straight average of the per-attempt differences. The league_baselines cpoe column stays near zero and serves as a sanity anchor.

How fast it stabilizes

Slowly. Accuracy differences between NFL passers are small relative to throw-to-throw randomness, so CPOE needs a big pile of attempts before it means much. Under roughly 150 attempts, treat it as descriptive. One week (30 attempts) is mostly noise, and situation-bucket CPOE within a single season is thinner still.

How it gets misused

• Small-sample worship. A hot three weeks of CPOE is one of the most quoted and least meaningful stats on football social media.
• Treating it as a complete passing grade. CPOE measures accuracy against expectation, not decision quality, sack avoidance, or downfield aggression. A passer can lead the league in CPOE while producing middling EPA. Read them together: EPA per dropback for value, CPOE for accuracy.
• Comparing CPOE across eras or models. It is model output, and different models (or different seasons of data) shift the baseline.