What is a quarterback's Clean-Core Rating?

The two buckets

Every dropback a quarterback takes is sorted by the situation it happened in. A dropback is in the clean core when all five of these are true: he was not pressured, the defense sent four or fewer rushers, he did not scramble, it was first or second down, and the ball was outside the twenty. The volatile bucket is the dropbacks where he was pressured, or was blitzed by five or more, or it was third or fourth down. Kneels are excluded. Both figures are expected points added per dropback; the stability gap is the first minus the second.

Who qualifies

Each team's primary starter — the quarterback with the most dropbacks — for 32 quarterbacks in all. They range from 253 to 674 dropbacks on the season, and the clean core is between 26.5 and 40 percent of each quarterback's dropbacks. No individual situational split is shown below n=20. A quarterback under 12 games or 350 dropbacks is marked as a part-season on the tables above.

Does clean-core production repeat? We checked. It does not.

CCR was built on a borrowed premise: that performance in clean conditions is the part of a quarterback's season that carries forward, while volatile production does not. We had never tested that, so before naming the metric we extended the same split back to 2016 and measured it.

It does not hold. Across 2016-2025, for quarterbacks clearing 200 dropbacks in consecutive seasons (n=238 pairs), clean-core expected points added per dropback predicts the following season at r=+0.185. Overall expected points added per dropback, over every dropback, predicts it at r=+0.401 — better, not worse.

Most of that difference is sample size rather than signal. The clean core is about a third of a quarterback's dropbacks, and any measurement taken on a third of the data is noisier. To separate the two we drew a random subset of each quarterback's own dropbacks the exact size of his clean core and ran the identical test: it repeats at r=+0.235 (standard deviation 0.047 across 25 draws). The clean core sits within that spread. Choosing clean dropbacks does not concentrate the repeatable part of a season any better than choosing dropbacks at random — it just leaves you with fewer of them.

So CCR is published here as a description of a season that has already happened, and nothing on this page projects a quarterback forward. We went looking for the repeatability claim, did not find it, and would rather say so than let a name imply it. The test is reproducible from scripts/qb_clean_core_stickiness.py.

One source note that matters

Pressure, blitz and time-to-throw come from participation charting (FTN Data via nflverse); everything else is nflverse play-by-play. Completion percentage over expected inside this dataset's situational splits is the nflfastR per-play figure, because Next Gen Stats publishes only season aggregates and has no per-situation rows. Next Gen Stats remains the canonical source for completion percentage over expected everywhere else on this site. A reader who assumes one source throughout will be off by as much as three points.

“In pocket” is a proxy — no public tracking field marks a collapsed pocket, so a dropback counts as in-pocket when the quarterback did not scramble. There is no play-action split: play action is in neither the play-by-play nor the participation file, and we would rather omit the bucket than fake it.