The Zeitgeist Passer Rating

Why a second number

The NCAA passer rating was written in 1979. It counts completions, yards, touchdowns and interceptions per attempt, and it ignores three things the play-by-play can now see: sacks (a sack is a dropback that never becomes an attempt), garbage time, and who the opponent was. ZPR keeps the passer rating on the page because everyone knows what it means, and adds a number built to fix those three blind spots.

The three ingredients

• Positive-play rate (weight 24%) — the share of dropbacks that gained expected points, by College Football Data’s play-value model. Consistency.
• Yards per attempt (weight 56%) — explosiveness. Attempts exclude sacks.
• Sack rate (weight 21%, a penalty) — sacks divided by dropbacks. The most stable quarterback trait in the data, and the one passer rating cannot see.

Each ingredient is measured over non-garbage dropbacks only (the site-wide definition: the pre-play margin above 28 at any point, or above 16 in the fourth quarter), and each is opponent-adjusted: we subtract how much the defenses this quarterback faced allow relative to the FBS average, weighted by how often he faced each one. A quarterback who posted the same raw numbers against a harder schedule scores higher. An FBS defense needs 100 dropbacks faced before its own rate is trusted; until then it counts as average. FCS and lower-division opponents never reach that floor, so they are pooled into one shared rate (they allow roughly 1.5 more yards per attempt than the FBS average, every season), and a September game against one is scored as what it is rather than as average competition. That pooling was added on 2026-09-01 after it proved to be the largest single improvement in early-season accuracy we tested.

How they combine

The weights were not chosen by opinion. They are the coefficients that best predict a quarterback’s opponent-adjusted expected points per dropback in the other half of his own season, fitted across 2022–2025 and locked. They are refit at most once per offseason; a metric whose definition moves weekly is not a metric.

Reading the scale. ZPR is on 0–100 with 50 as the FBS average, and it reads like a percentile: an 85 means “better than roughly 85% of qualified FBS starters, given how much we have seen.” The n / (n + 225) term pulls a thin sample toward 50: a quarterback who has been brilliant for 60 dropbacks is probably above average, not the best in the country yet. Below 125 dropbacks we show the number with a “thin sample” chip and do not rank it.

Early in a season, before about 40 quarterbacks reach 125 dropbacks, the current year cannot define its own average and spread, so ZPR borrows the previous full season’s scale. Week-3 numbers are therefore comparable to last year’s, and the card says which season’s scale was used.

What was tested, and what was left out

Before choosing the ingredients we measured how stable each candidate is inside a season: split every quarterback’s games into alternating halves and correlate the two halves (Spearman–Brown corrected to full-season length). A number that does not agree with itself across halves is measuring luck.

• Raw EPA per dropback is left out as a headline. It is the obvious “modern” replacement, and it is less stable than the 1979 formula it would replace while correlating 0.89 with it. The extra variance it adds is mostly big-play and touchdown luck.
• Interceptions are not weighted. Interception rate is essentially noise inside a season (three of four years near zero). Passer rating charges 200 points per pick; ZPR does not charge for picks at all. This is the second most common reason the two numbers disagree about a quarterback.
• Touchdown rate is left out for the same reason (year-to-year 0.07), and because play-by-play touchdown attribution is incomplete in 2025.
• Throw depth is not in the score. Average depth of target is a stable style trait with no relationship to efficiency (the same finding the NFL analytics community reached). It belongs on the page as a descriptor, not inside the number.

The bar it had to clear

A rating is only worth showing if it tells you something about what the quarterback will do next. The test: score each quarterback on one half of his games, then correlate that score with his opponent-adjusted expected points per dropback in the other half. Weights were fitted with the scored season held out. Computed with the shipped code on 2026-09-01:

ZPR beats both raw EPA and the passer rating in every held-out season. The honest ceiling matters too: about 0.35–0.45 half-season to half-season is what any per-dropback rate can manage at 150–200 dropbacks. ZPR is better than the alternatives, not clairvoyant — which is why the card carries a confidence chip tied to sample size.

Against ESPN’s Total QBR

The fair comparison is the same test with ESPN’s per-game Total QBR (averaged over the same half-season games) as the predictor. Same quarterbacks, same halves, same target; 2021 added because QBR coverage there is complete. Computed 2026-09-01:

QBR is a little better at this — it wins 2021 through 2023 by a small margin and loses 2024 and 2025 — and the two agree strongly about who is good (correlation about 0.73 on the same games). The pooled gap is about 0.02 with a bootstrap interval that includes zero, so the honest reading is a statistical tie with QBR slightly ahead. We tested the obvious additions (the quarterback’s own runs, raw EPA, completion rate, explosive-play rate) and none closed the gap. The horizon matters: through week 3, QBR predicts the rest of a quarterback’s season about as well as ZPR does; by week 7, ZPR is ahead. QBR appears to carry a stronger prior early, and ZPR’s opponent adjustment appears to be the better one once the sample exists.

We also tested the thing QBR does that ZPR does not: dividing credit between the passer and the receiver. Using CFBD’s charted air-yards and yards-after-catch data (2025 weeks 9–15, the only charted window that exists), we rebuilt the score with the receiver’s yards-after-catch-over-expected removed from the quarterback’s ledger, on identical plays with identical opponent adjustment. It predicted the same quarterbacks’ efficiency in weeks 1–8 slightly worse (0.42 vs 0.45), and the receiver share itself was positively predictive of the quarterback’s future efficiency. Within a season a quarterback keeps his receivers and his scheme, so removing what they add throws away signal. Game-by-game averaging, weighting plays by leverage, crediting pass-interference penalties, and stronger sample-size shrinkage were also tested and did not help. Credit division may still matter across seasons, when receivers change; that test needs a full charted 2026 season and is queued for 2027.

QBR is ESPN’s proprietary model: you can read the number but not recompute it. ZPR is recomputable from public play-by-play, with every ingredient stored, and that is the point of it. It is shown next to passer rating and QBR rather than presented as the last word.

What ZPR does not claim

• The play values underneath (“expected points”) are College Football Data’s model. ZPR is ours in construction and weighting, not in the play-value model.
• Per-play quarterback attribution begins in 2020. There is no ZPR for earlier seasons.
• About 4% of pass plays have no quarterback attribution in the source; per-quarterback dropback counts run slightly under the box score. This is uniform enough not to reorder anyone.
• It measures the passer’s dropbacks. Scrambles are rushes in the source data and are not part of ZPR; a dual-threat quarterback’s running value lives elsewhere on his page.

Every stored row keeps the raw and adjusted ingredients, the z-scores, the shrinkage factor and the scale season, so a skeptical reader can recompute the headline from the parts.