The Kneeldown

About The Kneeldown

Thank you for visiting The Kneeldown. This project was originally created in 2022 when I was in high school. I used my experience as an R&D with PFF to create the site and build the original concept. Since then, life has gotten in the way and I haven't been able to maintain it as much until now where I've attempted to revive it, building it better than ever.

Currently, it runs off of open source data meaning that there is no asymmetric information that is offered. The goal is to offer the average consumer the ability to understand football through a data-driven lens. If you have any suggestions, feedback, or questions, please reach out to me at ryweis1@outlook.com or on Twitter/X at @ryanweisman12 .

I am funding the hosting of this website by myself and plan to keep it free for the average user. If you have found the site valuable and would like to contribute to its hosting, feel free to do so here: Buy Me a Coffee .

Data Sources

Source	Data	Provider
nflfastR	Play-by-play, EPA, win probability	nflverse
FTN Charting	Ball quality, play action, blitz, interception worthy	nflverse via FTN
Participation	Coverage scheme, route, personnel, formation	nflverse
Next Gen Stats	Time to throw, separation, air yards, RYOE	nflverse via NGS
OverTheCap	Contracts, APY, guaranteed money, cap hits	nflreadr

Data Status

Every dataset behind the site, with its source, coverage, and when it was last rebuilt. Manifest generated 2026-07-26 00:12:32 UTC.

Dataset	Source	Rows	Seasons	Updated	Cadence
Play-by-play (nflfastR + participation + FTN) ⓘ	nflreadr::load_pbp / load_participation / load_ftn_charting	197,362	2022-2025	2026-07-26 00:12 UTC	Manual refresh; nightly upstream in season
NGS passing (weekly) ⓘ	nflreadr::load_nextgen_stats	3,631	2020-2025	2026-07-26 00:12 UTC	Nightly in season
NGS receiving (weekly) ⓘ	nflreadr::load_nextgen_stats	8,871	2020-2025	2026-07-26 00:12 UTC	Nightly in season
NGS rushing (weekly) ⓘ	nflreadr::load_nextgen_stats	3,703	2020-2025	2026-07-26 00:12 UTC	Nightly in season
Fantasy opportunity (ffopportunity)	nflreadr::load_ff_opportunity	36,063	2020-2025	2026-07-26 00:12 UTC	Weekly in season
Contracts (OverTheCap) ⓘ	nflreadr::load_contracts	51,734		2026-07-26 00:12 UTC	Periodic
Schedules ⓘ	nflreadr::load_schedules	1,411	2022-2026	2026-07-26 00:12 UTC	Refresh + schedule releases
Rosters ⓘ	nflreadr::load_rosters	15,507	2022-2026	2026-07-26 00:12 UTC	Season rollover + offseason updates
Snap counts (PFR)	nflreadr::load_snap_counts	106,148	2022-2025	2026-07-26 00:12 UTC	Weekly in season
ESPN QBR (season) ⓘ	nflreadr::load_espn_qbr	321	2022-2025	2026-07-26 00:12 UTC	Weekly/season
ESPN QBR (weekly) ⓘ	nflreadr::load_espn_qbr(summary_type='week')	2,273	2022-2025	2026-07-26 00:12 UTC	Weekly in season
Defensive player stats	nflreadr::load_player_stats(summary_level='week')	45,553	2022-2025	2026-07-26 00:12 UTC	Weekly in season
Team colors & logos	nflreadr::load_teams	32		2026-07-26 00:12 UTC	Static
Player identity dimension ⓘ	nflreadr::load_players	25,035		2026-07-26 00:12 UTC	Periodic

Charting-era caveat: 2022 coverage/route/pressure labels come from the NGS participation release (~84% of pass plays charted, different route vocabulary); 2023+ is FTN-charted (~100%). Cross-era rates are computed on different instruments — era-sensitive views say so in place.

Model Card — Play Caller EPA Model

What it does: predicts per-play EPA for a proposed pre-snap situation and target route family, scored under each coverage the selected defense actually plays (usage-weighted). It models the targeted receiver's route — not a full 5-man play concept.
Type: XGBoost regression (EPA), 45 features.
Trained: 2026-07-18 on 140,184 plays (2022 to latest season).
Random-split test RMSE: 1.3150 (correlation 0.334)
Out-of-time validation: trained only on past seasons, scored on a full future season it never saw. These fold RMSEs are not comparable to the random-split RMSE above: each fold scores a different season's plays, so the numbers describe different test populations (a lower fold RMSE does not mean better generalization). The random split also mixes same-game plays into training, so its RMSE overstates how well the model transfers to future games. The honest evidence is the model-vs-baseline margin WITHIN each fold: train 2022-2023 -> test 2024: model RMSE 1.310 vs situational baseline 1.385 (beats baseline); train 2022-2024 -> test 2025: model RMSE 1.305 vs situational baseline 1.380 (beats baseline) .
Top features by gain: is_pass, half_secs, ay_none, field_pos, distance
Limitations: trained on public charting only — no player tracking, no blocking assignments, no receiver identity; recommendations are historical-pattern estimates with real uncertainty, not guarantees. Coverage labels change provenance at the 2022/2023 charting-source boundary.

Model Card — Coverage-Mix Model

What it does: predicts which coverage a defense will show in a pre-snap situation, as a probability distribution. Those probabilities WEIGHT the Play Caller's per-coverage EPA predictions; the per-coverage EPA-allowed and box-count inputs stay empirical.
Type: XGBoost multiclass (multi:softprob), 45 features (pre-snap situation, offensive RB/TE counts, defense identity). Unit: one charted defensive pass play.
Classes: 10 charted coverage schemes (Cover 0–4, 6, 9, 2-Man, Combo, Prevent). BLOWN (busted coverage) is a data-quality bucket, never a class.
Trained: 2026-07-23 on 70,994 plays ( 2022–2025 ); validation mlogloss 1.6336 / top-1 37.5% at iteration 240 .
Out-of-time validation: trained only on past seasons, scored on a full future season it never saw. Promotion required beating the smoothed-empirical baseline on multiclass log-loss in BOTH expanding-window folds; it also wins top-1 accuracy in both:

Fold	Test plays	Log-loss (smoother)	Log-loss (model)	Top-1 (smoother)	Top-1 (model)
train 2022-2023 -> test 2024	19,829	1.8472	1.7506	31.2%	35.7%
train 2022-2024 -> test 2025	19,602	1.8401	1.8297	27.5%	27.9%

The 2025 fold's log-loss edge is thin (1.8297 vs 1.8401): on the most recent season the model is only marginally better than the smoother.

Limitations: pre-snap information only — no personnel packages beyond RB/TE counts, no motion, no injuries or roster context. Coverage labels change provenance at the 2022/2023 charting-source boundary (NGS participation vs FTN). Coverage calls are inherently noisy: even out-of-time, the best top-1 is ~36%.
Fallback: if the model bundle is missing or a defense/situation can't be encoded, the Play Caller logs one line and reverts to the hierarchically smoothed empirical mix (situation → defense → league, k = 15) — the previous production method, still fully tested.

Model Card — Defense Archetypes

What it does: clusters defensive coverage identities (10 smoothed coverage shares as isometric log-ratios + logit man rate + logit blitz rate, within-season centered so archetypes describe deviation from that season's league norm — the 2022 NGS vs 2023+ FTN vocabulary break is removed by construction). The Coverage Hub Identity tab shows posterior probabilities by PROJECTION onto this frozen fit — profiles are never re-clustered.
Type: Gaussian mixture (mclust 6.1.3 ), model VEE , k = 2 components, selected by ICL (entropy-penalized BIC — favors separated clusters, the right criterion for archetypes; k was never forced upward).
Fitted: 2026-07-25 on 128 defense-seasons ( 2022–2025 ).
Stability gate: B = 100 bootstrap resamples, mean Jaccard per cluster; ≥ 0.75 → named archetype, 0.60 – 0.75 → unnamed pattern, below → dissolved into “Mixed”:

Cluster	n	Mean Jaccard	Gate result	Shown as
1	18	0.41	dissolved	Mixed
2	110	0.83	named	Mainstream

Naming rationale: names are descriptive, centroid-derived, and explicitly unordered. On the frozen fit the league separates into one large stable cluster sitting essentially AT the season league norm (“Mainstream”, Jaccard 0.83) plus a small unstable fringe cluster (COMBO-avoidant, single-high/quarters tilt) that failed even the 0.60 pattern gate and dissolved into “Mixed” — the honest read at n = 128 is one mainstream identity plus fringe variation, not a menu of named schemes.
Limitations: 128 defense-seasons is a small pool — expect coarse clusters; archetype membership describes style, never quality (distinctiveness is not success); coverage labels change provenance at the 2022/2023 charting boundary (handled by within-season centering); the bundle is frozen and must be refit when a new season completes (new seasons project against the latest stored season center until then).

Model Card — Expected Interceptions (xINT)

Status: SHADOW-ONLY — not promoted. The model beat the league-rate, air-yards-bucket, and depth+situation baselines on log-loss in both out-of-time folds, but failed the cross-fit calibration gate in 1 of 4 assessment cells (2025 fold, fit even → assess odd weeks: slope 1.159 outside the pre-stated 0.85–1.15 band — the platt alternative fails the same cell (slope 1.168), so no calibrator family passes). The promotion rule was fixed before evaluation: every gate passes or nothing user-facing ships. Accordingly, no xINT or INTs-Over-Expected figure appears anywhere on the site — including the QB Decision Making view this model was built to extend. The evaluation artifact (models/xint_oot_eval.json) and the per-throw out-of-fold predictions (models/xint_oof_predictions.parquet) are kept so a recalibrated candidate can be judged against the exact same gates.

What it estimates: the probability that a single thrown ball is intercepted, from only what was true when it left the QB's hand — depth, location, coverage (busted coverages included as real context), man/zone, pressure, time to throw, contested-catch situation, down/distance/field/score/clock, box count, and rush count. It scores how dangerous the throw was: a ball thrown into danger that a defender drops is the same throw as one that gets picked. No QB or receiver identity — two QBs making the same throw into the same look get the same xINT.
Eligibility: charted pass attempts, 2022–25: pass play, ball actually thrown (no sacks — an un-thrown ball can't be intercepted), FTN-charted coverage, down 1–4; two-point tries excluded. 74,608 throws kept, 1,683 interceptions (2.26% league rate). Throwaways stay in: real thrown balls with near-zero INT risk (2,999 charted throwaways, 4 INTs 2022-25) — the model learns it.
Type: XGBoost binary classifier, no class reweighting (the true ~2.3% base rate is preserved so predicted probabilities can be read as probabilities), early stopping on the chronological tail of the training seasons — never on the test season.
Deliberately excluded — is_catchable_ball: P(INT | uncatchable) ≈ 0 near-definitionally, so the flag absorbs the target and quietly changes the question to "INT given catchable" — and it is charted with knowledge of how the play ended. An ablation WITH it scores better on paper ( 2024: log-loss 0.0813 vs 0.0966 primary; 2025: log-loss 0.0865 vs 0.1005 primary ) — exactly the signature of a feature leaking the outcome, not of a better model. It is reported here and used nowhere.
Never a feature — is_interception_worthy: FTN's INT-worthy flag IS a human expected-interception judgment — using it as an input would be pure leakage. It serves instead as an independent validation label (agreement below).
Kept with a caveat — is_contested_ball: charted at the catch point, so it is partly an outcome of the ball in flight rather than purely pre-outcome context.
Out-of-time validation (gate 1 — passed ): expanding-window season folds — train on past seasons only, score a full future season the model never saw. All three baselines are fit on the training rows only. The gate: beat every baseline on log-loss in BOTH folds.

Fold	Test throws	INTs	LL league rate	LL air-yards	LL depth+situation	LL model	AUC	Beats all 3
train 2022-2023 -> test 2024	18,438	403	0.1052	0.0985	0.0989	0.0966	0.743	yes
train 2022-2024 -> test 2025	18,218	405	0.1066	0.1028	0.1018	0.1005	0.707	yes

Cross-fit calibration (gate 2 — FAILED ): a beta calibrator is fit on one week-parity half of the test season and assessed on the other half — fit and assessment never share a block. Tolerances stated before evaluation: |Cox slope − 1| ≤ 0.15 and |mean predicted − observed rate| ≤ 0.005. Winning family: beta (mean assessed log-loss 0.0985 vs 0.0986 platt). All 4 cells must pass.

Test season	Cross-fit direction	Assessed throws	Cox slope	Intercept (prob)	Assessed log-loss	Gates
2024	fit odd → assess even weeks	8,990	1.098	+0.0015	0.0931	pass
2024	fit even → assess odd weeks	9,448	0.894	-0.0015	0.0996	pass
2025	fit odd → assess even weeks	8,834	0.863	-0.0028	0.1055	pass
2025	fit even → assess odd weeks	9,384	1.159	+0.0027	0.0959	FAIL

What the failed cell means: on the most-recent season, one assessment half needs its probabilities stretched by more than the tolerated amount (slope > 1 = the model's danger estimates run too flat there). Per-QB expected-INT totals are sums of these probabilities, so a breach this size would silently bias every total — which is why none are shown.
Agreement with human charting (out-of-fold): on test-season throws the scoring model never trained on, throws FTN charted as INT-worthy carry a mean xINT of 4.1% vs 2.1% on throws charted clean (n = 1,207 worthy / 35,449 not; AUC 0.727 predicting a label the model never saw). The model independently rates human-flagged throws about twice as dangerous.
Leaderboard discipline: per-QB "expected INTs" are only honest when every throw is scored by a model that never trained on it — the evaluation builds them strictly out-of-fold (2024 scored by the 2022–23 model, 2025 by the 2022–24 model, each half calibrated on the other half). Those predictions exist on disk, but publishing rankings from a model that failed its calibration gate would present biased sums as facts. This is a rebuild of previously published xINT research; the gates exist precisely because the original shipped random-split, in-sample, uncalibrated numbers.
Limitations: public charting only — no player tracking, no separation at the catch point; post-throw only (not a pre-snap risk model); coverage labels change provenance at the 2022/2023 charting boundary (2022 lacks COVER_9/COMBO/BLOWN — the season feature absorbs the break); contested-ball outcome-adjacency above; and the most-recent-season calibration drift that the failed gate caught.

Methodology — QB After Disruption

What it shows: how a QB's dropbacks after a disruption event (pressure, QB hit, sack, interception, strip-sack) compare to his own baseline, in four response windows: next dropback, rest of drive, next same-QB drive, rest of game.
The estimand — a within-QB difference-in-differences, never a raw split: the displayed delta is (situation-adjusted EPA residual on post-event dropbacks) minus (the SAME QB's residual on his non-post-event dropbacks, reweighted by situation bucket to the post-event situation mix). A pooled baseline would leave QB quality inside the delta; and because events select on transient bad process, post-event play regresses up by construction — so a matched placebo (“after a situation-matched bad play that wasn't the event”: non-event dropbacks below the QB's own median EPA, importance-weighted to the event's situation distribution) is displayed beside it. If event and placebo deltas match, the “response” is ordinary mean reversion, not something the event caused.
Sequencing guards: plays ordered by game_id + play_id (monotonicity asserted); nullified no-play snaps skipped, never counted as the response; kneels, spikes, and two-point tries excluded; “next drive” = the next drive with the same offense AND the same passer (never naively fixed_drive + 1, which is usually the opponent's drive); a turnover on the event play empties the rest-of-drive window; truncated windows are dropped from denominators, never scored 0; a benched QB's window resumes on his return without bridging the backup's plays.
Uncertainty: per-QB deltas are shrunk toward the league delta (empirical Bayes, k = 25 events) and carry 95% game-clustered bootstrap CIs (B = 500 — plays within a game are dependent, so games are resampled, not plays).
Reliability tiers (measured anchors, ≥300-dropback QB-seasons): median per QB-season is 148 pressures / 32 sacks / 10 INTs / 2 strip-sacks. So pressure, hit, and sack windows are per-QB viable with shrinkage; interception windows are pooled 2022–2025 (a narrower season pick auto-falls back, with a notification); strip-sack windows are league-level only, never per QB.
Reading discipline: observed situation-adjusted output after the event, relative to this QB's own baseline — reflects selection and mean reversion; descriptive, not a causal bounce-back, never toughness/psychology.

Methodology

Pizza Charts: Percentile rankings computed against the full position pool using ECDF.

Per-metric definitions (EPA/Play, Success Rate, CPOE, Matchup Grades, and every other displayed metric) moved to the Metric Dictionary card below — each entry states its exact denominator, sample gate, shrinkage policy, charting-era caveat, and data source.

Metric Dictionary

Every metric the site displays, from one shared registry (the same entries power the in-place ⓘ tooltips). Each entry states the plain-language definition, the exact denominator, the minimum sample shown, any shrinkage applied, whether the 2022 charting-source break affects it, and where the data comes from.

EPA/Play

Expected Points Added per play — the change in expected points caused by a play, from nflfastR's expected-points model. The site's universal currency: positive = good for the offense.

Denominator: all plays in the selected sample (QB coverage/route views: the QB's pass plays vs the scheme; team views: scrimmage plays, rps %in% c('pass','run'), excluding kneels/spikes/special teams)
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: epa_play

EPA/Target

The same nflfastR EPA, averaged over a receiver's targets — every play on which he was the targeted receiver.

Denominator: the receiver's targets (plays with him as targeted receiver)
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: epa_target

EPA/Dropback

The same nflfastR EPA, averaged over ALL dropbacks — including sacks and throwaways, not just throws. Used by the Play Caller QB-vs-Coverage value box and the Pressure sub-tab's EPA/Dropback Allowed.

Denominator: all dropbacks (rps == 'pass'), sacks and throwaways included
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: epa_dropback

Success Rate

Share of plays with EPA > 0.

Denominator: plays in the current sample with a non-missing EPA
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: success_rate

TO +/-

Turnover margin: takeaways minus giveaways, counted over ALL plays (special teams included) in the selected seasons. A turnover is an interception or a lost fumble (fumble_lost). Attribution follows possession: the offense (posteam) commits, the defense forces — EXCEPT on punts, where posteam is the PUNTING team while nearly every punt fumble lost is a receiving-team muff (97 of 98 in 2022-25). A punt fumble_lost with positive posteam EPA is therefore credited as a giveaway BY the receiving (returning) team and a takeaway FOR the punting team; a punt fumble_lost with EPA <= 0 (botched snap, failed fake) stays a punting-team giveaway. Kickoffs need no flip: posteam there is already the receiving team.

Denominator: counting stat, not a rate: all plays in the selected seasons, special teams included
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: to_margin

Stabilized EPA (EB shrinkage)

Empirical-Bayes stabilized EPA: eb_shrink(x, n, prior, k) = (n·x + k·prior) / (n + k) with prior weight k = 25 observations — small-sample means are shrunk toward a relevant prior so thin cells can't dominate rankings. Raw values and n always remain visible.

Denominator: the cell's own plays/targets (n), shrunk toward the view's stated prior: Route×Coverage heatmap — LEAGUE same-coverage mean computed from the full filtered pass frame (never the Team/QB-scoped frame); Play Caller QB-vs-Coverage concepts — league mean for that concept vs that coverage; receiver ranking — situation-pool mean; Read Progression — per-read league mean
Min n shown: none (all n shown)
Shrinkage: eb_shrink k=25 toward the view's stated prior
Era caveat: No
Source: derived
Registry id: stabilized_epa

CPOE

Completion Percentage Over Expected — actual completion minus nflfastR's expected completion probability given play difficulty.

Denominator: pass plays with a non-missing cpoe value (~80% fill on pass plays)
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: cpoe

ADOT

Average Depth of Target — mean air yards on throws. The canonical instrument is pbp air_yards (coalesced with the 2022-only NGS ngs_air_yards); Weekly Stats shows NGS avg_intended_air_yards.

Denominator: throws with a charted air-yards value
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: adot

Deep Throw %

Share of throws with air yards ≥ 20 (the deep_throw flag, derived from coalesce(air_yards, ngs_air_yards)).

Denominator: throws with charted air yards (~93% of pass plays each season); deep = air yards ≥ 20
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: deep_throw_rate

Time to Throw (TTT)

Seconds from snap to release, from Next Gen Stats. QB-tab buckets: <2.3 / 2.3–2.8 / 2.8–3.3 / 3.3+ seconds.

Denominator: dropbacks with an NGS time_to_throw measurement
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: NGS weekly
Registry id: ttt

Pressure Rate

Share of opposing dropbacks with an FTN-charted pressure on the QB (was_pressure).

Denominator: charted pressures / opposing dropbacks (rps == 'pass')
Min n shown: 100
Shrinkage: none
Era caveat: No
Source: FTN charting
Registry id: pressure_rate

Blitz (5+ rushers)

Share of dropbacks with five or more total pass rushers (n_pass_rushers ≥ 5). One of two blitz definitions shown side by side, never blended.

Denominator: 5+-rusher dropbacks / opposing dropbacks with a charted n_pass_rushers
Min n shown: 100
Shrinkage: none
Era caveat: No
Source: FTN charting
Registry id: blitz_rate_rushers

Blitzers (2nd-level) ≥ 1

Share of dropbacks where FTN's n_blitzers ≥ 1. n_blitzers counts only SECOND-LEVEL rushers, so this is a different construction from the 5+-rushers definition — the two are shown side by side, never blended.

Denominator: n_blitzers ≥ 1 dropbacks / opposing dropbacks with a charted n_blitzers
Min n shown: 100
Shrinkage: none
Era caveat: No
Source: FTN charting
Registry id: blitz_rate_blitzers

Pressure w/o Blitz

Pressure generated without sending extra rushers: the pressure rate restricted to dropbacks with four or fewer pass rushers.

Denominator: pressures on ≤4-rusher dropbacks / ≤4-rusher dropbacks (charted n_pass_rushers)
Min n shown: 100
Shrinkage: none
Era caveat: No
Source: FTN charting
Registry id: pressure_no_blitz

QB-Fault Sack Share

Share of sacks in the sample that FTN charted as the QB's fault (is_qb_fault_sack).

Denominator: FTN QB-fault sacks / sacks in the sample
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: FTN charting
Registry id: qb_fault_sack_share

Coverage Usage

How often a defense lines up in each charted coverage scheme, over dropbacks only (rps == 'pass' — QB scrambles with charted coverage are excluded, matching the QB coverage views). BLOWN (busted coverage — a data-quality bucket, not a callable scheme) is excluded from numerator AND denominator, so shown usage sums to 100%.

Denominator: team dropbacks in the scheme / team charted-coverage dropbacks (rps == 'pass'); league baseline = mean of per-team usage shares over ALL teams in the same filtered sample, with zero shares for teams that never show the scheme
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: FTN charting
Registry id: coverage_usage

Coverage EPA Allowed

Mean EPA a defense allows on dropbacks in each charted coverage scheme. Dropbacks only (rps == 'pass'): QB scrambles carry charted coverage too but belong to the run population, so they are excluded — the same denominator as the QB-vs-coverage views. League baseline is the play-weighted mean over all dropbacks in the scheme under the same filters.

Denominator: dropbacks (rps == 'pass') with the scheme charted vs the selected defense; league baseline over all such dropbacks league-wide
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: FTN charting
Registry id: coverage_epa_allowed

Target Share

Player targets divided by team targets — computed PER TEAM STINT (player × team over that stint's weeks) so traded players are not blended across teams; the most recent stint is reported.

Denominator: player targets with the team / that team's total targets over the stint's weeks
Min n shown: 30
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: target_share

Catch %

Completions on the receiver's targets (mean of complete_pass over targets) — of ALL targets, unlike Drop %, which uses catchable targets only.

Denominator: receptions / all targets (catchable or not)
Min n shown: 30
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: catch_rate

Drop % (of catchable)

Drops divided by CATCHABLE targets (FTN-charted catchable balls) — never of all targets (league ≈5.0%). Lower is better; inverted in pizza percentiles.

Denominator: FTN-charted drops on catchable targets / catchable targets
Min n shown: 25
Shrinkage: eb_shrink k=25 toward the play-weighted league drop rate (computed from everyone before the display filter)
Era caveat: No
Source: FTN charting
Registry id: drop_rate_catchable

Contested Catch %

Contested-catch conversion: completed catches on FTN-charted contested targets.

Denominator: contested catches (contested & complete) / contested targets
Min n shown: 25
Shrinkage: eb_shrink k=25 toward the play-weighted league contested conversion (computed from everyone before the display filter)
Era caveat: No
Source: FTN charting
Registry id: contested_conv

INT-Worthy Rate

Share of dropbacks with an FTN-charted interception-worthy throw. Uncharted flags count as not-flagged (≥99.6% of plays are charted).

Denominator: INT-worthy throws / QB dropbacks (rps == 'pass')
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: FTN charting
Registry id: int_worthy_rate

INT-Worthy Throws Intercepted

Share of a QB's interception-worthy throws that were actually intercepted — computed as the intersection sum(int & int_worthy) / sum(int_worthy), because tipped-ball INTs aren't always charted worthy. League ≈48.5% pooled 2022–25. Descriptive; never framed as “luck”.

Denominator: intercepted INT-worthy throws / INT-worthy throws
Min n shown: 100
Shrinkage: none
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: FTN charting
Registry id: int_worthy_conversion

Throwaway Rate

Share of dropbacks FTN charted as throwaways (is_throw_away).

Denominator: throwaways / QB dropbacks (rps == 'pass')
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: FTN charting
Registry id: throwaway_rate

Separation

NGS avg_separation — yards between the receiver and the nearest defender at catch/arrival; the receiver's weekly NGS average joined per play.

Denominator: targets with an NGS separation value (weekly player average joined per play)
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: NGS weekly
Registry id: separation

RYOE

Rush Yards Over Expected — the NGS model's actual minus expected rush yards given blocking/box, as a per-attempt weekly average joined per play.

Denominator: carries with an NGS RYOE value (per-attempt weekly average)
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: NGS weekly
Registry id: ryoe

Tendency Flag Usage

Team usage of the FTN play flags (motion, no-huddle, RPO, shotgun, play action, screens) vs the play-weighted league under the same filters. Association, not play-calling causation — the flags are not randomly assigned.

Denominator: Motion / No-Huddle / RPO / Shotgun per scrimmage play (rps %in% c('pass','run')); Play Action / Screen per dropback (rps == 'pass')
Min n shown: 100
Shrinkage: none
Era caveat: No
Source: FTN charting
Registry id: tendency_usage_rate

Matchup Grade

Score = EPA the defense allows on the player's most-TARGETED route families vs league average, weighted by the player's TARGET share on each family (0.5 = league-neutral). Targets are the charted data — routes run are not tracked (ff_opportunity route counts are weekly aggregates and are not wired here). Letter grades curve the Score across the graded (100+ target) players: A = top 20%, B = next 20%, C = middle 20%, D = next 20%, F = bottom 20%.

Denominator: the receiver's charted-route TARGETS overlaid on the opponent's coverage-allowed EPA; graded pool = players with ≥100 targets
Min n shown: 100
Shrinkage: reliability curve: sub-100-target scores pulled toward the DISPLAYED population's mean raw score (recomputed from the current filtered field before curving — not a constant 0.5) by min(targets/100, 1); letter grades only for ≥100-target players
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: derived
Registry id: matchup_grade

Targets/Gm

Targets per game in the pizza-chart metric tables (WR/TE and RB). The denominator is games the player APPEARED in — any offensive involvement on a play: targeted (receiver), carried (rusher), or dropped back (passer) — counted once per season-week. It is NOT games-with-a-target (which doubled the rate for players targeted in half their games) and not official games played: pbp has no participation log, so a zero-touch appearance cannot be counted.

Denominator: distinct season-weeks with any offensive involvement (target OR carry OR dropback) within the filtered sample (compute_player_games_appeared, R/07-utils-pizza.R)
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: nflfastR
Registry id: targets_per_game

Pizza Percentile

Percentile of a player's metric vs the qualifying position pool, via the pool's empirical CDF: P(pool ≤ value), so ties share credit. Lower-is-better metrics (INT-Worthy%, TTT, Fumble%, Drop%) use the ECDF of NEGATED values — P(pool ≥ value) — which is tie-symmetric: players tied at the best value all get the top percentile (the old 1 − ECDF form was the strict P(pool > value) and charged every tie).

Denominator: position pool at the qualifying volume (QB ≥100 plays, WR/TE ≥30 targets, RB ≥40 carries), roster positions as pool authority
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: derived
Registry id: pizza_percentile

APY Commitments / Headroom (GM ledger)

The GM simulator's ledger: the sum of APY over the team's active contracts, plus dead money from simulated cuts. This is an APY-BASED ESTIMATE, not official cap space — real cap hits differ from APY because signing bonuses prorate over up to five years and offseason accounting counts only the top-51 charges, so summing APY exceeds the $301.2M cap for most cap-compliant teams (headroom = cap − commitments is then negative, and the share-of-cap gauge exceeds 100%). Cut/sign/trade moves stay internally consistent in APY units; cap-hit modeling (prorations, restructures) is deliberately deferred.

Denominator: counting stat, not a rate: sum of OTC APY ($M) over the selected team's active contracts + simulated dead money; share of cap divides by the $301.2M 2026 league cap
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: OTC
Registry id: apy_commitments

FA Estimated Market Value

Estimated APY for a simulated free-agent signing. Anchored on the player's own most recent contract APY (the market's last real observation of him); players with no usable prior APY fall back to the position's median APY among recent (2023+) signings. Where the player has recent play-by-play EPA, his EPA percentile within the pool's position group scales the anchor by a bounded 0.6x–1.8x multiplier. An estimate — age, injury, scheme fit, and bidding wars are not modeled, and the anchor lags when the last deal is stale.

Denominator: the free-agent pool: each player's latest inactive contract (year_signed ≥ 2022) for players holding NO active contract; EPA percentile over the pool's position group (whole pool when the group has < 10 EPA-attached players; EPA gates: 50 pass / 20 target / 30 rush plays)
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: derived
Registry id: fa_market_value

Draft Pick Trade Value (Jimmy Johnson chart)

Points from the actual Jimmy Johnson trade value chart, embedded for all 224 picks (3000 for pick 1, 590 for 32, 100 for 100, 2 for 224); picks 225+ continue the round-7 step (0.3/pick) floored at 1. Replaced (2026-07) an exponential approximation that ran 1.7x–53x high after round 1 and flattened the chart's early-pick premium. Player-for-pick verdicts map the player's APY to a pick equivalent and compare chart points; the JJ chart reflects 1990s trade norms, not modern surplus-value research.

Denominator: not a rate: chart points per pick (1–262); player trade value uses apy_to_pick_equivalent(APY) → pick → chart points
Min n shown: none (all n shown)
Shrinkage: none
Era caveat: No
Source: derived
Registry id: jj_pick_value

Defensive Distinctiveness

How far a defense's coverage identity sits from its season's league norm — the leave-one-out Mahalanobis distance of the identity vector (CLR of 10 smoothed coverage shares + logit man rate + logit blitz 5+ rate) from the season's play-weighted league centroid computed WITHOUT the team, under an OAS-shrunk covariance fit on within-season-centered defense-seasons 2022+ and refit without the scored team. Reported as rank/percentile among the season's 32 defenses only — the reference distribution supports no p-values. Distinctiveness measures difference from the league, NOT quality — a distinctive defense is not thereby a good one.

Denominator: not a rate: distance over the defense's eligible charted non-BLOWN dropbacks (n shown); percentile among the season's 32 defenses
Min n shown: none (all n shown)
Shrinkage: coverage shares via kd_smooth_coverage_mix k=15 (week -> defense-season -> league); man/blitz rates via eb_shrink k=15 toward the league rate
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: derived
Registry id: distinctiveness

Weekly Identity Change (JSD)

Jensen-Shannon divergence (log base 2, bounded 0-1) between week w's smoothed coverage mix and the exponentially decayed trailing mix of the team's PRIOR weeks (half-life 4 weeks, week w excluded; the first two observed weeks are not scored). Because both arms shrink toward the same priors, raw JSD is deflated in low-volume weeks — every week is therefore shown against a multinomial parametric-bootstrap null band (90th percentile of 500 no-change resamples at the true ns), and a week only reads as changed when it clears the band. Within-season only; descriptive — no persistence alerts.

Denominator: week-w charted non-BLOWN dropbacks vs the effective n of the EW-decayed trailing mix (both ns shown)
Min n shown: none (all n shown)
Shrinkage: both arms smoothed via kd_smooth_coverage_mix k=15 toward the defense-season then league mix
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: derived
Registry id: identity_change_jsd

Archetype Probability

Posterior probability from the frozen defense-archetype Gaussian mixture (mclust on ilr + logit identity vectors of all pooled defense-seasons, within-season centered; model/k selected by ICL; bootstrap-Jaccard stability gates: >= 0.75 named, 0.60-0.75 unnamed pattern, below dissolved to Mixed). Profiles are PROJECTED onto the frozen model — never re-clustered — so posteriors are deterministic. Archetypes describe style relative to the season's league norm, not quality; names are descriptive and unordered.

Denominator: not a rate: posterior probabilities over the frozen mixture components (sum to 1); fit pool = defense-seasons 2022-25 (n = 128)
Min n shown: none (all n shown)
Shrinkage: inputs are the smoothed identity features (kd_smooth_coverage_mix k=15; eb_shrink k=15 for rates)
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: derived
Registry id: archetype_prob

Similar-Defense EPA (LOO)

Mean EPA allowed per dropback of the k=5 nearest defense-seasons by the same Mahalanobis identity metric — same charting era only, min-n gated, and always EXCLUDING the team itself (leave-one-out). A descriptive baseline: 'vs stylistically similar defenses', not opponent-adjusted, and similarity of style says nothing about quality — distinctiveness is not success.

Denominator: all dropbacks (rps == 'pass') faced by the k nearest same-era defense-seasons with >= 200 eligible dropbacks, self excluded
Min n shown: 200
Shrinkage: none (plain mean over the neighbor defense-seasons)
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: derived
Registry id: similar_def_epa_loo

Play-Caller Era Mix Shift (JSD)

Jensen-Shannon divergence (log base 2, bounded 0-1) between the hierarchically smoothed coverage mixes immediately before and after a documented defensive play-caller change. The observed divergence is compared with the 90th percentile of 500 multinomial no-change resamples at the two eras' true sample sizes. A shift is descriptive, not a causal effect of the staffing change.

Denominator: charted non-BLOWN defensive dropbacks in the adjacent play-caller segments; both segments require at least 100 eligible dropbacks
Min n shown: 100
Shrinkage: both era mixes use kd_smooth_coverage_mix k=15 toward that defense-season and the season league mix
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: derived
Registry id: caller_era_shift_jsd

After-Event EPA (within-QB DiD)

Observed situation-adjusted output after a disruption event (pressure, hit, sack, INT, strip-sack), relative to this QB's OWN baseline: mean situation-bucket EPA residual on post-event dropbacks minus the same QB's residual on his non-post-event dropbacks reweighted to the post-event situation distribution (difference-in-differences). Reflects selection and mean reversion; descriptive, not a causal bounce-back, never toughness/psychology. Reliability tiers per measured event counts: pressure/hit/sack are per-QB-season viable, INT per-QB only pooled 2022-25, strip-sack league-level only (never per-QB — median 2 per QB-season).

Denominator: genuine dropbacks (qb_dropback == 1; no_play, kneels, spikes, and two-point tries excluded) inside the selected response window — next dropback, rest of drive, next same-QB drive, or rest of game — same game and same passer only; truncated (empty) windows dropped from denominators, never scored 0
Min n shown: 20
Shrinkage: eb_shrink k=25 toward the league DiD; 95% CI from a game-clustered bootstrap (B=500) on the unshrunk DiD
Era caveat: No
Source: derived
Registry id: after_event_epa_did

After-Event Placebo (matched bad play)

The same within-QB DiD computed after a situation-matched bad play that was NOT the event: non-event dropbacks with EPA below the QB's own median, importance-weighted by situation bucket to match the real event's situation distribution. Shown beside the event DiD so the event-specific increment separates from generic bad-moment reversion — if the event and placebo deltas match, the 'response' is ordinary mean reversion, not something the event caused. Reflects selection and mean reversion; descriptive, not a causal bounce-back, never toughness/psychology.

Denominator: same response-window dropbacks as the event DiD, with placebo events = situation-matched non-event dropbacks below the QB's median EPA; same guards (truncated windows dropped, turnover plays empty the rest-of-drive window)
Min n shown: 20
Shrinkage: eb_shrink k=25 toward the league placebo DiD; game-clustered bootstrap CI (B=500)
Era caveat: No
Source: derived
Registry id: after_event_placebo

Event Response Window

The set of dropbacks scored as the response to a disruption event, under seven sequencing guards: plays ordered by game_id + play_id (monotonicity asserted); no_play/penalty-nullified snaps are skipped, never counted as the response; kneels, spikes, and two-point tries excluded; 'next drive' means the next drive with the SAME offense AND the SAME passer (never naively fixed_drive + 1, which is usually the opponent); a turnover on the event play empties the rest-of-drive window; truncated windows are dropped from denominators, not scored 0; and a benched QB's window resumes on his return without bridging the backup's plays. Post-event numbers reflect selection and mean reversion; descriptive, not a causal bounce-back, never toughness/psychology.

Denominator: events with at least one same-game, same-passer response dropback in the window (next dropback / rest of drive / next same-QB drive / rest of game); dropped-window counts are reported
Min n shown: none (all n shown)
Shrinkage: none (window definition, not an estimate)
Era caveat: No
Source: derived
Registry id: event_response_window

Expected INTs (xINT)

Model-estimated probability that a thrown ball is intercepted, summed over a QB's throws. Post-throw XGBoost model (air yards, pass location, coverage scheme incl. busted coverages, man/zone, pressure, time to throw, contested-ball, down/distance/field/score/clock, box, rushers) trained on charted attempts 2022-25 with cross-fit Platt-or-beta calibration, so probabilities sum honestly against the ~2.3% league INT rate. Deliberately EXCLUDES catchable-ball and INT-worthy charting (outcome-adjacent / the human label itself) and any QB or receiver identity — it scores the throw's danger, not the player. Promotion gates: beat league-rate, air-yards-bucket, and depth+situation baselines on log-loss in both expanding out-of-time folds, plus cross-fit calibration slope/intercept within tolerance (models/xint_oot_eval.json).

Denominator: charted pass attempts: rps == 'pass', pass_attempt == 1, sack == 0, FTN-charted coverage, down 1-4 — throwaways included (real throws with near-zero INT risk), two-point tries excluded (no down)
Min n shown: none (all n shown)
Shrinkage: none (calibrated model probability)
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: derived
Registry id: xint

INTs Over Expected

Actual interceptions minus expected interceptions (xINT) over the same throws. Positive = more picks than the danger of the throws implies; negative = fewer. Descriptive — reflects ball placement, defender plays, and variance; never framed as 'luck'. Leaderboards are built ONLY from out-of-fold predictions (each season scored by a model trained on prior seasons, calibrated cross-fit), never in-sample fits.

Denominator: the QB's xINT-eligible charted attempts in the selected window (out-of-fold seasons only for leaderboards)
Min n shown: 200
Shrinkage: none
Era caveat: Yes — affected by the 2022 (NGS participation) vs 2023+ (FTN) charting break; era notes appear in place.
Source: derived
Registry id: int_over_expected

Connect

Twitter/X Support the Site Email

Explore

Coverage by Down & Distance

Coverage Mix Over Time

Weekly Identity Change

Archetype Probabilities

Similar Defenses

Advanced Play Caller

Fantasy Matchup Tool

Be Your Own GM

Roster Diff

Who's Overpaid? Who's a Bargain?

Contract Roster View

Cap Allocation by Position

Contract Database

Run Direction vs Box Count

WR/TE Route Tree

Route × Coverage Heatmap

Underlying Table

Receiver Hands — league-wide