Edge Lab

Hypothesis: Familiarity compresses margins and favourites underperform.

Result: Flat. Divisional favourites win at their implied rate.

2006-2025 · 5,295 games · z < 1.17 · tested 2026-08-01

Hypothesis: Home dogs are systematically undervalued.

Result: Flat. No exploitable deviation from market probability.

2006-2025 · 5,295 games · z < 1.17 · tested 2026-08-01

Hypothesis: Some coaches own particular opponents.

Result: Flat. Head-to-head history adds nothing beyond the line.

2006-2025 · 5,295 games · z < 1.17 · tested 2026-08-01

Hypothesis: Turf/grass and Denver altitude create priced-out edges.

Result: Flat. Venue effects are fully absorbed by the market.

2006-2025 · 5,295 games · z < 1.17 · tested 2026-08-01

Hypothesis: Favourites sandwiched between big games underperform.

Result: Flat. No measurable letdown effect after de-vigging.

2021-2025 · Full odds backfill · n/s · tested 2026-08-02

Hypothesis: Market/ELO agreement (v1), then bookmaker dispersion and book count (v2), should predict which favourites are safest.

Result: Failed twice. v1 inverted - low-confidence games won 76.9% of the time because the market prices roster news ELO cannot see. v2 came back non-monotonic. Not shipped.

2021-2025 · Full odds backfill · inverted / non-monotonic · tested 2026-08-02

Hypothesis: Head-referee assignment shifts game outcomes against the closing line.

Result: Mixed. Every market-facing angle came back flat - favourite win rate, home win rate, totals and over rate all failed correction for multiple comparisons. But penalty-rate heterogeneity across crews is genuinely real and significant. Real effect, not exploitable.

2006-2025 · 3,024 games, 27 head referees · chi-sq = 45.1, p = 0.0014 (penalty rate); market angles n/s · tested 2026-08-05

Hypothesis: Closing lines are sharper than midweek lines in games involving teams with nothing to play for.

Result: Validated on the elimination arm: closing lines were significantly more accurate than midweek lines against a week 5/10 control. Waiting for late-week news is a real edge in those spots.

2021-2025, weeks 15-18 · 183 games (144 elimination-proxy, 39 clinch-proxy) · z = 3.21 · tested 2026-08-03

Hypothesis: A team whose pass rush and pass protection are better than its opponent's wins more often than its de-vigged market price implies.

Result: Season-to-date net sack differential is strongly collinear with the market's own pricing (r = 0.509), so a raw test would look positive for the wrong reason. After regressing de-vigged win probability out of the differential, the residual has no predictive value for outcomes (r = -0.006, z = -0.27). Quintiles of the residual all land within 1.2 points of their implied win rate (|z| <= 0.52, non-monotonic). The PFR pressure-rate version behaves the same way (collinearity r = 0.537, residual r = 0.013, z = 0.59) on ~27.5 player rows per team-week.

2021-2025, weeks 5-18 (REG) · 2,052 team-weeks (4-game minimum rolling window) · collinearity r = 0.509 with de-vigged win prob; residual r = -0.006, z = -0.27 · tested 2026-08-25

Hypothesis: Favourites underperform their de-vigged market price in primetime standalone windows.

Result: Flat. Primetime is not priced differently to begin with — the mean de-vigged favourite win probability is 0.6673 in primetime and 0.6681 outside it, so there is no collinearity to strip out. Favourites then finished +0.44 sigma above their implied win rate in primetime and -1.11 sigma below it elsewhere, a difference of z = +0.92 — not significant, and directionally the opposite of the premise: primetime favourites slightly overperform. Splitting by win-probability bucket produces no monotonic pattern.

2006-2025 · 1,230 primetime games / 4,051 non-primetime games · collinearity check: 0.6673 vs 0.6681 mean favourite win prob; primetime z = +0.44, non-primetime z = -1.11, difference z = +0.92 · tested 2026-08-25

Hypothesis: Home favorites of a given spread size win at a different rate than the de-vigged moneyline implies, with the gap varying by how big the spread is.

Result: No bucket clears the bar once the number of buckets tested is accounted for. Two buckets reach |z| ~ 2, which is roughly what is expected by chance across 23 tests, and they point in opposite directions from their neighbours. The 8.5 bucket cited by the original claim is -3.8pp with z = -1.33. Buckets above 11 points carry fewer than 70 games each and are too thin to trust. Pooled across every home favorite the gap is -1.0pp with z = -1.24.

2006-2025 · 3,212 home-favorite games across 23 one-point spread buckets · largest gaps: -5.1pp at 6 (z = -2.31) and -4.6pp at 3 (z = -1.94); all home favorites pooled -1.0pp, z = -1.24 · tested 2026-09-15

Hypothesis: Favorites in weeks 1-4 win roughly 3 percentage points less often than their price implies, because the market has not yet corrected for roster and form changes.

Result: The direction and the size of the claimed effect both replicate: early-season favorites underperform their de-vigged moneyline by about 3 points while weeks 5+ are exactly on the number. The evidence is real but weak. No individual early week clears |z| = 2 on its own (week 3 is the strongest at -2.10), 12 of 19 seasons show a negative gap, and both halves of the window (2006-2015 z = -1.26, 2016-2025 z = -1.74) fall short separately. The effect appears in both home and away favorites, which argues against it being a home-field artifact. Treat as a small directional prior, not a standalone edge.

2006-2025 · 1,133 favorite games in weeks 1-4 vs 3,928 in weeks 5+ · weeks 1-4: -2.97pp, z = -2.14 (95% CI +/- 2.72pp); weeks 5+: +0.28pp, z = +0.38; difference-in-edges z = -2.07 · tested 2026-09-15