NFL Edge Research — July 2026 (offseason build window)

Date: 2026-07-02. All code in this directory; nothing in production was touched. All results below are graded against closing lines, flat stakes, pushes excluded, one bet per pick (dedup'd). Given the July 2026 track-record audit, every historical "NFL record" found on disk was re-verified before use — two of them turned out to be retrodictions (see 1.3).

1. Current NFL edge surface (what we exploit today: effectively nothing)

1.1 Live pipeline. [redacted] contains zero NFL alerts ever (store began 2026-02-01, after the NFL season). [redacted] (source tag nfl_model_v3) is wired to scan h2h + spread + total vs devigged book lines using the temporal model + residual sigmas (sigma_spread~12.67, sigma_total~13), but it has never produced a live, settled record.

1.2 Model quality (honest, walk-forwardwalk-forwardEvaluating week by week using only what was knowable before each week, mimicking how the model would actually have been used at the time.). From experiments/iter_03_FINAL_holdout_2025.json (2025 holdoutholdoutData deliberately set aside and never looked at while developing an idea, then used once at the end as a fair test. Peeking at it first would defeat the purpose., leak-free): - ML Brier 0.2116 / accuracy 65.8% — vs the closing market's 0.2109 / 66.2% (measured in 04_h2h_market_efficiency.py). The model is at market level, not above it. - ml_flat betting at close: -5.26% ROIreturn on investmentProfit as a percentage of the money wagered. +2% means $2 profit per $100 bet. (n=179). The +6.5% "kellyKelly criterionA formula for how much to stake given your edge. Full Kelly maximizes long-run growth but swings wildly; most people use a fraction of it." ROI in the same file is on 2.46 total units staked — noise, not evidence. - Totals model backtest (data/totals_backtest_predictions.csv, 861 games 2021-2024): ROI -2.0% to -3.2% at every edge threshold (|edge|>=0/2/4/6). No edge in the totals model.

1.3 Invalid "records" (do not cite, ever). - Root model_performance.db: 1,087 NFL predictions all created in a 32-min window on 2026-01-18 covering 2021-2024 — hindsight retrodiction (already flagged by the audit). - [redacted]::stat_prop_bets claims +18.9% ROI on 4,431 dedup'd prop picks (62.4% win rate). Verified: this is a retrodiction written by scripts/backfill_stat_props.py on 2025-12-27 covering 2024 weeks 1-16 and 2025 weeks 1-17, with mismatched alt-line odds (e.g. an OVER graded at the standard 69.5 line but paid at +800 alt pricing; 25% "push" rate on X.5 lines). The number is fabricated-by-construction. Any props work must start from zero live evidence.

Bottom line: the platform currently has no proven NFL edge and no live NFL record.

2. Data inventory (what's actually on disk / freely reachable)

3. Candidate edges tested (scripts 01_-05_, reproducible)

3.1 Wong teasers (6-pt, 2-team) — 01_wong_teasers.py — BEST CANDIDATE

Rule fixed a priori (Wong 2001): tease dogs +1.5..+2.5 -> +7.5..+8.5, faves -7.5..-8.5 -> -1.5..-2.5 (both cross 3 and 7). Per-leg breakevens: 72.4% at -110, 73.9% at -120, 75.8% at -135 (current DK/FD standard).

Honest read: dog legs only are consistent across two independent 5-6 season eras (77.7% / 77.6%) and clear the -120 breakeven by ~2 SE; they clear the -135 breakeven by only ~1 SE. Fav legs have decayed below breakeven — the market fixed that half of the strategy. The dogs/favs split is post-hoc here but matches the published literature (favorite Wong legs died first). Volume is small: ~50 dog legs/season ~= 25 two-team teasers/season. Verdict: real but modest, and entirely price-dependent — +EV at <=-125, roughly breakeven-to-thin at -135. Requires shopping teaser pricing.

3.2 Wind unders — 02_wind_totals.py — NOT CONFIRMED

Threshold fit on 1999-2015 (wind >=18 mph: 61.1% under, +16.7% ROI), evaluated once on 2016-2025: 54.8% under, n=73 (SE +/-5.8%), +4.6% ROI — statistically indistinguishable from breakeven. The market demonstrably already prices wind (closing totals drop ~2.3 pts from calm to 21+ mph). ~7 qualifying games/season, and live betting relies on wind forecasts, which makes real results strictly worse than this backtest. Verdict: no deployable edge.

3.3 Situational ATS angles — 03_situational_ats.py — DEAD

12 classic angles x 2 eras vs closing spread. In 2015-2025 nothing clears breakeven meaningfully: post-bye -7.3%, divisional home dogs -4.5%, rest-edge

=4 days -3.0%, primetime dogs -3.2%. The nominal winner (freezing outdoor home teams, 56.4% n=140) is exactly the ~1 false positive expected from 24 tests. Verdict: fully priced; do not build.

3.4 h2h market efficiency / NHL-style stacking — 04_h2h_market_efficiency.py — NFL IS NBA-LIKE, NOT NHL-LIKE

Walk-forward logistic stack of [logit(devigdevigRemoving the bookmaker's cut from odds to recover the market's actual implied probability. closing ML), logit(MOV-Elo)], train 2007..T-1, test T for 2018-2025: - Stack Brier vs market Brier: mean delta +0.00001 (no improvement, any season). - EloElo ratingA rating system, originally from chess, that moves a team up or down based on results and the strength of the opponent. coefficient shrinks toward zero over time (+0.10 in 2018 -> +0.01-0.04 by 2024-25): the closing lineclosing lineThe final odds right before a game starts. It reflects everything the betting market knows, which makes it the hardest benchmark to beat. has absorbed everything public-data models know. - Betting sim (>=2% edge vs devig, paid at actual closing price): -9.0% ROI, n=207. - Corroborated by the platform's own stronger model: -5.3% flat at close (1.2). Verdict: do NOT attempt an NFL h2h/spread/total model-vs-close edge. The NHL v3 playbook does not transfer; NFL close is the sharpest line we track.

3.5 Off-consensus line shopping — 05_multibook_line_shopping.py — OVERLAY, NOT AN EDGE

2020-2024, 1,408 games, near-close multi-book snapshots. Bet any book >=0.5 pt better than the side's median consensus (dedup per game-side): - n=1,522: +0.7% ROI at the outlier line/price vs -1.8% for the same sides at consensus -110 -> line shopping is worth ~2.5% ROI. - But >=1.0 pt outliers lose -4.2% (n=337): big deviations are informed books being right (or limits/staleness you can't actually bet), not free money. Verdict: shopping recovers most of the vigvigThe bookmaker's built-in cut. It is why a coin-flip bet at -110 needs you to win about 52.4% of the time just to break even. but is ~breakeven standalone. Use as an execution overlay on any real edge (e.g., teaser pricing); never chase large outliers.

3.6 Player props / early-week CLV capture — INSUFFICIENT DATA

• Genuine prop capture covers 20 games (late Dec -> SB). No honest backtest possible; the only "record" (+18.9%) is fabricated (1.3).
• Zero early-week->close NFL line history exists on disk, so "capture CLVclosing line valueWhether you got a better price than the market settled at. Widely used as a faster signal of skill than profit, which takes ages to measure. as the edge" (bet Tue/Wed numbers that beat the close) is untestable today. Verdict: cannot be evaluated until we capture a season of data — see 4.

4. Ranked recommendation for the 2026 season

1. Ship a Wong-teaser dog-leg scanner (only proven +EV item). Rule: current spread +1.5..+2.5, 6-pt tease, pair legs across games, only alert when 2-team pricing <= -125 (pass at -135 unless a book discounts). Honest expectation: low-single-digit % ROI on ~25 teasers/season — a small, real edge, not a headline. Track it live from week 1 with true timestamps so a real record exists by January.
2. Start capture in August (this is the actual high-value move): (a) NFL spreads/totals/ML snapshots from Tuesday open through close (extend the existing unified_odds poller — it already does this for other sports); (b) full-season prop lines with open+close (extend the player_props poller beyond the 20-game trial); (c) teaser/alt-line pricing per book. This makes early-week-CLV, props, and derivatives honestly testable by mid-season 2026 / offseason 2027.
3. Line-shop everything (~+2.5% execution overlay), but never treat an off-market line as signal — >=1 pt outliers lose.
4. Explicitly do NOT build: NFL h2h/spread/total model-vs-close edges (market is NBA-efficient; our own model loses 5% at close), situational angles, wind unders. And never resurface stat_prop_bets or root model_performance.db numbers — both are retrodictions.

Reproduce

python3 01_wong_teasers.py          # teaser legs, eras, prices
python3 02_wind_totals.py           # train/test wind threshold
python3 03_situational_ats.py      # 12 angles x 2 eras
python3 04_h2h_market_efficiency.py # stacking test + betting sim
python3 05_multibook_line_shopping.py

Data snapshot: schedules_1999_2025.feather (nflverse, downloaded 2026-07-02).