The Marginbeta
Methodology · Phase B fitted

AUDIBLE — how the NFL team composite works

AUDIBLE is a descriptive team-strength composite built from eight position callsigns and Phase B Ridge-fitted weights. Two of three ship gates pass; the third — sign-consistency on offensive skill-position callsigns — fails. We publish anyway, because the composite is honest about how a team is good even when it has no edge against the closing line.

1. The strategy pivot

Phase 3 of the NFL player rebuild aimed to replace v1.2's team-feature backbone for spread prediction by aggregating per-position AUDIBLE ratings up to the team. End-to-end testing in 2026-05-15 showed no spread lift over the existing v1.2 model. The player aggregation is descriptively useful (we can say sharp things about how a team is good, position-by-position) but not forecasting-useful (we don't beat the line by enough to call it a betting signal).

Phase B is the formal write-up of that finding: we ran a 10K-bootstrap Ridge on game-level outcomes, found that 2 of 3 ship gates pass but the third (sign-consistency on all eight callsigns) does not, and decided to publish the result as a descriptive composite rather than as a forecaster. This page is the honest record.

2. How we got here

v1 (hand-weighted). The first cut used hand-picked proportions: POCKET 0.40, ROUTE 0.20, TRENCH 0.20, HINGE 0.10, ENGINE 0.10, with the defensive unit subtracted at full weight. Calibrated by face-validity, not by data — and overweighted QB relative to what the data actually says (POCKET's no-spread variance share is 9%, not 40%).

DS + football consults. Two memos (EXPERT_DS_AUDIBLE_REDESIGN_2026_05_15.md andEXPERT_FOOTBALL_AUDIBLE_REDESIGN_2026_05_15.md) predicted the outcome before the fit ran: the line absorbs skill positions efficiently; the “structural” units (OL, defensive sides, ST) are the rock-stable signals.

Phase B Ridge. We built a training matrix of 2,111 games with home-minus-away differences in the eight callsign ratings plus the closing spread. LOSO CV chose λ=19.3; 10K bootstrap on the full-sample fit gave per-coefficient 95% CIs. The fitted weights are what feeds this page.

3. The eight callsigns

4. The fitted weights

Phase B fit two Ridge models on the same eight callsigns: one with closing spread as a feature, one without. The page publishes the no-spread β as the headline weight (it captures what each callsign explains on its own) and reports the with-spread 95% CI for transparency.

CallsignPositionWith-spread β95% CISign% (8 LOSO folds)Var share (no-spread)
STORM-PASSPass defence+2.53[+1.94, +3.12]100%50.9%
TRENCHOL unit+1.29[+0.72, +1.89]100%24.6%
BOOMSpecial teams+1.11[+0.57, +1.67]100%7.1%
STORM-RUSHRush defence+1.01[+0.44, +1.58]100%5.4%
POCKETStarter QB+0.30[-0.29, +0.93]⚠83.5%9.4%
ENGINELead RB+0.22[-0.34, +0.76]⚠77.9%0.7%
ROUTEWR corps+0.19[-0.39, +0.76]⚠73.7%1.9%
HINGETE × tgt share-0.01[-0.59, +0.57]⚠51.9%0.0%

⚠ = bootstrap CI spans zero. β values are standardised (a +1σ improvement in that callsign moves expected home margin by β points, after controlling for the closing line where applicable).

5. Ship-gate verdict

GateThresholdResultVerdict
1. RMSE ex-best fold≤ 13.1012.81PASS ✓
2. ATS ex-best fold≥ 51.5%55.85%PASS ✓
3. Sign-consistencyAll 8 callsigns: CI excludes zero AND no LOSO sign-flips4 / 8 (TRENCH, STORM-PASS, STORM-RUSH, BOOM pass; POCKET, ROUTE, HINGE, ENGINE fail)FAIL ✗

Headline: RMSE and ATS pass; sign-consistency fails. We publish as a descriptive composite, not as a forecaster. See §7 below for what the failed gate means in football terms.

6. What the data told us — variance partition

In the no-spread fit, full R² (LOSO out-of-fold) = 0.153. The per-callsign share of that explained variance is the most useful read on which signals carry the composite:

STORM-PASS  50.9%  ████████████████████████████████
TRENCH  24.6%  █████████████████
POCKET   9.4%  ██████
BOOM   7.1%  █████
STORM-RUSH   5.4%  ████
ROUTE   1.9%  █
ENGINE   0.7%  ▌
HINGE   0.0% 

STORM-PASS alone explains half of what the eight callsigns can say about game outcomes. TRENCH is #2 at 25%. Together, pass defence + offensive-line is three-quarters of the composite's signal. This is the football consult's §1/§5 prediction verbatim — “defence is more stable than offence, and the pass-D / OL block is the core.”

HINGE (TE × target share) contributes essentially zero to explained variance. ENGINE contributes 0.7%. The original hand-weights gave ENGINE 10% of the composite — that's ~15× overweight. We surface HINGE and ENGINE on the page for descriptive completeness, but the data is clear: at the season-level team-rating grain, RB workload and TE quality add almost nothing on top of OL + defence + ST.

7. Why we don't claim to predict spreads better than v1.2

Phase 3 (separate workstream, end-to-end test) and Phase B (this fit) both arrive at the same conclusion through different paths. Phase 3 ran the player-aggregate team features through v1.2's spread model and got no measurable lift over the production baseline. Phase B's with-spread Ridge fit finds the four skill-position callsigns have bootstrap CIs that span zero — Vegas has efficiently priced QB / WR / TE / RB quality, and the marginal information beyond the closing line is approximately zero on the offensive skill side. The team page publishes the composite as a descriptive number; it is not a betting signal.

8. Situational signals — tracked, not yet folded in

Alongside the composite, the team drawer surfaces six situational metrics for the season:

These are tracked separately because the football consults call them out as orthogonal to the EPA aggregates already captured in the eight callsigns. We'll watch them for a month and decide whether to fold them in. Until then they are explicitly descriptive: shown beside each team but not in the composite.

9. PASPA cliff disclosure

We refit the Ridge separately on pre-2019 (n=799) and post-2019 (n=1,312) games. The closing-spread coefficient is virtually unchanged (Δβ = 0.02). The callsign coefficients drift substantially on the offensive side (POCKET, ROUTE, ENGINE, STORM-RUSH all shift by 0.4-0.9 standardised units). The market has evolved how it prices the callsigns, but its overall calibration is rock-stable.

Practical implication: any production deployment should refit periodically with a rolling window. The bootstrap CIs reported above slightly under-state real uncertainty because the data is non-stationary. The current cadence target is a yearly refit on a rolling 8-year window.

10. Season-sim backtest — Sprint 4 (2026-05-15)

We Monte-Carlo simulate each season Y from the prior season's AUDIBLE ratings (Y−1, walk-forward — never any in-season data leakage) and check three forecasting gates on outcomes 2018-2024. Two bugs were caught and fixed during Sprint 4:

  1. Defense missing from 2026 view: team_aggregator was reading pass_rating / rush_rating from the def_unit_v1 components blob, but the 2026 schema writes def_pass_rating / def_rush_rating. 51% of model variance (STORM-PASS + STORM-RUSH, see §6) was silently zeroed for the live publication. Fixed: aggregator now reads both key names and prefers the new one.
  2. No regression-to-mean between seasons: a single prior-season elite team (BAL 2024 had def-contrib +11.13) was being projected forward at full strength into Y+1. Added a per-callsign blend toward the prior-season league mean with weight w = 0.30, picked from a 5-value grid sweep (0.00, 0.15, 0.20, 0.25, 0.30) by minimum 7-season win MAE.
Forecasting gateThresholdPre-Sprint-4 (mr=0.00)Post-Sprint-4 (mr=0.30)
Win-totals MAE (7-season mean)< 2.502.565 → 2.463PASS ✓
Playoff hit rate (per-conf top-7 logit)> 60%53.6% → 71.6%PASS ✓
SB champ in top-8≥ 50%4/7 → 4/7PASS ✓

Headline (post-v3, 2026-05-15): all 3 gates pass. The breakthrough was decoupling the playoff decision from the margin-RMSE-optimized win-total scalar. The v2 Ridge still drives per-game margin and win totals; a new per-conference top-7 logistic on makes_playoffs — fit on the 8 callsigns plus 12 situational features (3rd-down conv, RZ-TD%, turnover/drive — levels AND year-over-year deltas) plus coaching-shock features (new HC binary + tenure) — replaces league-wide top-14 ranking. Hybrid 7-season hit rate jumps 53.6% → 71.6%. See docs/AUDIBLE_PLAYOFF_LIFT_2026_05_15.md for the experiment trail (DS1 + DS2 expert memos, 4 candidate feature stacks, ablations).

Sweep results (10K sims each, final winner re-run at 50K):

mr=0.00  MAE 2.565  playoff 53.6%  SB 4/7  → 1/3 gates
mr=0.15  MAE 2.518  playoff 53.6%  SB 4/7  → 1/3
mr=0.20  MAE 2.506  playoff 53.6%  SB 4/7  → 1/3
mr=0.25  MAE 2.497  playoff 53.6%  SB 4/7  → 2/3
mr=0.30  MAE 2.486  playoff 52.6%  SB 4/7  → 2/3 (winner)

11. Reproducibility

The full Phase B report lives at docs/AUDIBLE_V2_PHASE_B_2026_05_15.md with the training matrix at derived/nfl_players/audible_v2_training_matrix.parquet on the external data drive. The fitted weights this page publishes are at derived/audible_v2_fitted_weights.json + derived/audible_v2_supplemental.json. The team aggregator reads them at import time and rebuilds the composite via polyedge.modules.nfl.players.team_aggregator.

One-line summary

A descriptive composite that's honest about which signals it believes in and which it doesn't. The pass defence and OL unit are the load-bearing 75%; everything else is published with a 95% confidence interval that lets the reader judge whether the ⚠ flag applies.