NFL QBs Visualized

Attribute profile: two scales, sources and limits

Dropbacks without a sack (%) occupies the desired pocket-presence position. Scramble contribution occupies the desired mobility position using the separate annual Mobility export: 559 finite accepted points among 560 qualifying player-seasons. Kyle Boller 2007 is withheld pending recorded-role review; 2011 is excluded. Sack avoidance reflects protection, receiver availability and decisions with other costs. Scramble contribution combines usage with whole-play shared value. Neither isolates the desired trait.

The short, medium and deep positions show completions above expectation on throws in each depth bin, in percentage points. These depth-specific completion residuals describe recorded outcomes relative to the supplied model; they do not isolate throwing accuracy or ball placement. Read the depth bins and formula.

Each percentile describes where a recorded outcome sits among its fixed same-year peers. For dropbacks without a sack and each of the three depth outcomes, peers are that year’s qualifying passers with a finite value for that outcome. Annual qualification requires at least 200 included throws across at least 10 included-throw games; depth bins add no minimum throw-count rule. Peer membership comes from the validated source records and cannot change with the selected opponent or search text. Each available outcome shows its year and peer count, N.

Annual Mobility percentiles use finite accepted scramble contributions from the qualifying player-seasons in that year. The 2007 pool has 26 values because the Boller point is withheld; 2023 has 29 and 2024 has 32. These are separate from the 20-QB pools used for the repeatability study. Each of the other four outcomes uses its own comparison pool. Compare each stat’s displayed year and finite peer count.

The midpoint-percentile formula is P = 100 × (B + 0.5 × T) / N. Here B counts peer values below the selected value, T counts exact ties including the selected record, and N counts all finite peer values including that record. Ties use unrounded native values. Higher recorded outcomes receive higher percentiles on all five measured spokes. A constant finite cohort gives every record 50; fewer than two peers or an unavailable native value gives no percentile. Missing evidence never becomes zero or 50, and another season never supplies a replacement.

The five outcome spokes use radius P / 100; 50 describes midpoint standing, not a zero native outcome or average ability. The two expert spokes instead use the ordinal drawing index (11 - rank) / 10: rank #1 is outer and #10 is at one-tenth radius. This index is not an NFL percentile or a calibrated ability grade. Numeric ring labels are omitted because the two scales have different meanings. Whole-number labels round half up using floor(P + 0.5); geometry retains the full percentile. Endpoints are not stretched to force a 1 or 100. With 20 untied peers, percentiles run from 2.5 to 97.5 in five-point steps. Whole-number labels do not imply one-point skill resolution.

Farther out means higher same-year standing on that outcome. Cross-year comparisons use different reference cohorts; they do not establish absolute ability or trait improvement. Equal percentile gaps need not mean equal differences in native outcomes or ability. The three depth outcomes share context and correlated evidence, so they do not count as three independent skill votes. Compare one spoke at a time: no area, aggregate score or winner combines them. Lines join adjacent available positions only. No line bridges a gap. The final edge closes only when all seven positions are available; the shape has no fill.

Optional full evidence retains raw values, counts and native-unit conditional game-based ranges. Those raw ranges are not percentile confidence intervals. They do not form a joint confidence region, predict performance or test a player/year difference. Individual sack-avoidance uncertainty was not estimated. The negative-depth diagnostic remains outside the radar. Read the range assumptions, the play populations, the sack formula, and the scramble formula.

Expert assessments: source, coverage and measurement limits

The data contain 84 published ESPN rank facts for same-year annual-eligible quarterbacks in the 2021–2025 preseason editions: 41 arm-strength ranks, 26 field-vision ranks and 17 decision-making ranks. Missing or unlisted entries remain unavailable; they do not mean last, low ability or an inferred rank. The source rank label, tie and marker are retained.

Processing uses field vision in 2021–2023 and decision-making with the football in 2024–2025. Each is a different partial editorial assessment, not measured thinking speed or a complete decision-quality grade. The selected values show the actual category, preseason edition, source date and panel or individual author. Different editions, categories and authorship do not form a comparable skill time series. A rank difference does not measure improvement. The expert axes use ordinal drawing indices, not outcome percentiles; equal rank steps do not establish equal skill gaps. Source uncertainty remains unquantified.

• ESPN 2021 preseason edition · 2021-08-23: NFL experts; Panel; individual and joint contributor slots. Points combined from personal top-ten ballots; weights not specified.
• ESPN 2022 preseason edition · 2022-08-22: NFL Experts; Panel; individual and joint contributor slots. Points combined from personal top-ten ballots; weights not specified.
• ESPN 2023 preseason edition · 2023-08-14: NFL experts; Nine named analysts. Points combined from personal top-ten ballots; weights not specified.
• ESPN 2024 preseason edition · 2024-08-12: Dan Orlovsky; Individual judgment. Current performance and upcoming-season expectations; article uses tape observations.
• ESPN 2025 preseason edition · 2025-08-11: Dan Orlovsky; Individual judgment. Current performance and upcoming-season expectations; tape-based; starters only.

Comparable isolated arm strength would need calibrated ball-speed observations and throw geometry. Processing would need validated observations of available options and choices. Those measurements remain unavailable. The sourced judgments supply a distinct kind of evidence; they do not establish those missing measurements. No expert confidence interval or composite score is invented.

How the models work

The data provider supplies completion chances and expected-points values. This project summarizes recorded outcomes using those values. The fixed-cohort studies examine associations across game groups and seasons; the wider annual and depth data provide descriptive comparisons.

The supplied completion model uses recorded throw/context information. The inspected upstream implementation includes depth/location, down/distance and a QB-hit field; these are not all independent of QB decisions. The project neither fits nor retrains this model. The supplied data do not identify which model or training sample generated these values, their out-of-sample status or their calibration in each era. Completion model implementation.

The upstream expected-points model uses game-state inputs such as field position, down/distance and time left. It estimates scoring-event probabilities and converts them to expected points. EPA then values the resulting play. The project uses supplied raw EPA, not receiver-fumble-adjusted QB EPA, and fits no expected-points model. The generating expected-points model and training vintage remain unknown for these snapshots. Modeled game-state adjustment does not isolate protection, receivers or coaching. Comparing seasons with different football conditions and unknown model calibration does not produce an era-neutral ability comparison. Official expected-points explanation · EPA implementation.

What the game-based ranges mean

An approximate range based on game-to-game variation, under the study’s assumptions. Here, the bars describe approximate uncertainty around each summary, not the spread of game performances. These conditional ranges are not probabilities for true talent, predictions for next season or bounds on every source of uncertainty.

The annual EPA range removes one complete player dropback game at a time and recalculates the pooled EPA/dropback ratio, holding qualification fixed. It uses all games containing that player’s recorded dropbacks. The run-contribution ranges for the 20-passer study use all combined-opportunity games, including games with zero events of the named run role.

Depth ranges use all of the player’s included-throw games in that season, including games with zero throws in the named depth bin. Let G be that full game count, n_bg the bin throws in game g, and s_bg their sum of supplied CPOE values. The full estimate is theta_b = sum_g(s_bg) / sum_g(n_bg). The leave-one-game-out estimate is theta_b[-g] = (sum_h(s_bh) − s_bg) / (N_b − n_bg). A zero-bin game remains a deletion with no removed bin throws; the positive-bin-game count does not replace G.

Every deletion must retain a positive denominator and finite estimate. The analysis neither drops failed replicates nor requalifies the player after deletion. An empty bin leaves the estimate unavailable; an invalid deletion leaves the range unavailable, with an explicit status. For valid deletions, m = sum_g(theta[-g]) / G, V = (G−1)/G × sum_g((theta[-g]−m)^2), and the displayed range is theta ± 1.96 × sqrt(V). This is an approximate marginal 95% conditional game-deletion sensitivity, not the minimum-to-maximum deletion range.

A collapsed range means no observed deletion variation, not certainty. Whole-season CPOE, completion rate and individual sack avoidance have no supplied individual range here. The three depth ranges do not provide simultaneous 95% coverage. Marginal ranges neither test differences between depth bins, players or years nor establish rank certainty; overlap and non-overlap do not supply those missing tests.

The calculation assumes approximately independent game shocks, stable opportunity sampling and a useful normal approximation, conditional on fixed qualification and observed support. It omits uncertainty from selection, serial injuries, receivers, protection, coverage, scheme, source labels and the supplied models’ training/calibration. Unknown model vintage and era changes preclude talent-confidence, causal, era-neutral or forecasting interpretations.

CPOE = completion percentage − expected completion percentage.

EPA = value of the result − expected points before the play.

Without a sack (%) = 100 × (dropbacks − sacks) ÷ dropbacks.

Signed scramble contribution is C_S = sum(raw supplied EPA on recorded scrambles) / N_O, in expected points per included recorded quarterback play. It combines scramble usage with shared whole-play value. It does not measure speed, all rushing ability or isolated mobility. A fully observed zero-scramble role with positive O gives a genuine zero contribution. There are four such rows; zero does not mean zero mobility talent. Positive and negative EPA can also cancel.

Formulas: C_S = sum_S(EPA)/N_O; C_R = sum_R(EPA)/N_O. Shares are 100 × N_S/N_O and 100 × N_R/N_O, percent. EPA per run divides each role’s EPA by that role’s count and remains a diagnostic with no supplied interval or repeatability test. A fully observed absent role has zero share and contribution, with unavailable per-run EPA. Zero contribution can also result from positive and negative EPA cancelling; inspect the counts. Missing evidence never becomes zero. Adding these contributions to EPA per dropback would mix denominators and double-count scrambles.

Correlation is r = sum_i[(x_i−mean(x)) × (y_i−mean(y))] / sqrt(sum_i(x_i−mean(x))² × sum_i(y_i−mean(y))²) across equally weighted passers. Their pooled summaries supply x/y. Near +1 means strong positive linear alignment, not identical values. Zero means no linear association, not zero talent. Constant values make correlation undefined.

Research question

How do quarterbacks compare when we isolate them from their team and situation?

What repeatability measures

The findings below belong to the fixed 20-passer 2023/2024 cohort and its ten study records. The 560-player-season annual data and depth summaries do not establish repeatability for that broader sample.

Repeatability asks whether players with higher values in one group also tended to have higher values in the other. Across-passer association does not measure a selected player’s personal consistency, absolute agreement, retained-skill percentage or forecast accuracy.

The sack study’s 2023 halves alternate each player’s date/game-ID-ordered dropback games; run halves alternate combined-opportunity games. Whole player-games stay intact. Annual comparisons use full 2023/2024 summaries. These shared/overlapping studies are not independent replications and do not directly test future performance.

Annual CPOE association was 0.507, versus 0.452 for completion percentage on the same throws. Their difference was 0.055, with a conditional interval from −0.274 to 0.384. Omitting one paired player reverses that difference. The evidence does not establish CPOE superiority. The paired correlation difference is dimensionless, not a percentage-point completion gap.

Recorded sack avoidance had a primary interleaved association of 0.265 [−0.289,0.685]. Chronological and annual results answer separate questions; all three remain in the study table.

Primary scramble-contribution association was 0.313 [−0.190,0.686]; other-run association was 0.697 [0.004,0.938]. Annual associations were 0.727 [0.226,0.924] and 0.429 [−0.080,0.761]. All four endpoints remain in the study table.

Study ranges remove a whole contributing NFL game across both teams and selected players, keeping membership/maps fixed. Sack/run correlations use Fisher transforms and each season’s own deletion mean. Annual CPOE retains its accepted full-estimate centering; the paired difference uses the same game deletions for both metrics. These approximate conditional ranges assume approximately independent game shocks and suitable smooth-statistic behavior. The 223 + 208 annual game units do not form 431 independent quarterback pairs or repeated validation studies.