New to these reports? Start here
- Dotted-underlined words have a plain-English definition — hover or tap them. Every term is also on the glossary page.
- "Null" means we found nothing, not that something broke. Most reports here are negative results, on purpose — knowing an idea doesn't work is the point.
- Two questions get asked separately. First, is the effect real? Second, is it already priced into the betting odds? An effect can be completely real and still useless to bet on.
- A "calibration" row is a self-check. It runs the same method on something already known to be true. If that fails, the whole report is unreliable — so it's reported alongside the findings.
- If a confidence interval includes zero, the real effect might be nothing at all, so no claim gets made.
NFL consensus-stacking evaluation — REJECT (2026-07-04)
Question: does the NHL v3 "top-down" recipe (tiny logistic on devigged closing consensus + small orthogonal features) beat the NFL closing moneyline?
Prior art and the gap filled
backtests/nfl_edges/FINDINGS.md §3.4 already stacked a rival MOV-Elo onto the
close (no gain; EloElo ratingA rating system, originally from chess, that moves a team up or down based on results and the strength of the opponent. coef decayed to ~0.01-0.04 by 2024-25; sims negative). This
study fills the untested gap: the NHL-style structural feature family
(rest, short-week, bye, divisional, neutral site).
Data
nflverse schedules 2006-2025: 5,281 non-tie games with closing ML (the single
closing ML is the consensus). odds_history.db line history is Jan-2026-only
(unusable); unified_odds.db has no NFL.
Results (walk-forward train 2006..T-1, test 2012-2025)
| pool | n | consensus LL | stacked LL | ΔLL (95% CI95% confidence intervalThe range the true value is plausibly in. If this range includes zero, we cannot rule out that the real effect is nothing at all.) | cons acc | stack acc |
|---|---|---|---|---|---|---|
| REG+playoff | 3,815 | 0.6092 | 0.6125 | +0.0033 [+0.0007, +0.0058] | 0.6647 | 0.6658 |
| REG only | 3,649 | 0.6094 | 0.6127 | +0.0033 [+0.0010, +0.0055] | 0.6657 | 0.6654 |
Stacking is significantly worse on log losslog lossA score for probability forecasts that punishes confident wrong answers harshly. Lower is better. in both pools. Smoking gun: mean |structural coefficient| decays monotonically 0.103 (2012) → 0.045 (2025) while the consensus coef holds ~0.82-0.86 — the market progressively absorbed the schedule structure, same pattern §3.4 found for Elo.
Verdict
REJECT with high confidence. Neither a rival model (§3.4) nor structural features beat the NFL close. The only documented NFL soft spot remains the Wong-teaser dog-leg structure (price-gated scanner, separate work). Revisit only if a genuinely private pre-close signal (e.g. real-time inactives) becomes available.
Reproduce: python3 backtests/stacking_eval/nfl/run_nfl_stacking.py