Docs / Game Predictions / 7 · Where the Engine Stands
07 — Where the Engine Stands
Docs 1–6 describe how books and quants price NFL games, and each ended with a list of changes for our engine. This page records what we built from those lists, what the honest measurements showed once it was built, and where the work is aimed now. Last updated October 2026.
What shipped
| Idea (doc) | What the engine does now |
|---|---|
| Opponent-adjusted rating spine (2, 5) | A league-wide ridge power rating — margin of victory capped at 28, recent games weighted more (8-week half-life), home edge solved to about 1.55 points — enters the score as one capped term beside the defense-weighted baseline |
| Probability layer + key numbers (2, 5) | Margins and totals become cover/over probabilities through a key-number-aware distribution (spikes on 3, 7, 6, 10…), with a margin SD that widens with the spread and a total SD of 13.2. The distribution is centred so a model that agrees with the line gets exactly 50% |
| QB starter vs. backup (2, 4) | Each team has a sticky baseline starter. The engine prices the gap to the quarterback actually starting, including third-stringers, and moves the total as well as the side. A game with a QB change is no bet |
| Market-aware blending (5) | The published margin is 70% model / 30% market; the total keeps 30% of the model's view; the slate's totals are re-centred so there is no systematic over/under lean |
| De-vig, flagging, sizing (1, 6) | A two-part gate: at least 4 points of edge on a spread or 3 on a total, and a model cover probability of 52.4% + 3 points. Quarter-Kelly stakes, capped |
| CLV as the scoreboard (1, 6, 8) | Every flagged pick is logged at the moment it is issued, with the line it was issued at, then graded against the closing line and the final score |
| Leak-free validation (5) | Backtests rebuild every profile, team average and home-field edge from games before the season being tested. Tuning must also win on a season the tuner never saw before any change is adopted |
| Real weather and injuries (3) | National Weather Service forecasts for outdoor games inside a week; zero effect otherwise (never sampled). The injury feeds are live, but only the quarterback moves a price today |
What the measurements showed
Building the measurement layer first was the point: it told us, honestly, how much the model actually knows.
- Against the spread, 2022–25 (leak-free): 48.8%, 50.0%, 45.7%, 44.0% — 183-205 pooled, 47.2%. Totals: 47.1% pooled. Every season is below the 52.4% break-even. Earlier, higher numbers came from backtests that leaked future information into past seasons; they were wrong and are retired.
- ΔR² beyond the closing line: 0.0006–0.0104 by season. The model carries essentially no information the market does not already have.
- Calibration: fitted on 2022–24 and scored on 2025, the best trust factor for our cover probabilities is zero. Picks the model called 75% likely covered 49% of the time. The probability printed on a ticket is the calibrated one, so tickets appear as one-unit, one-star leans.
- Live 2026 CLV: as of October 10, 2026, the 81 graded picks with a closing line average +0.35 points of closing line value, and 44% beat the close. That is not yet distinguishable from zero (the margin of error is about ±0.5 points). An earlier figure of roughly zero was wrong: until October 2026, a snapshot taken days before kickoff, or one taken after a game had started, could stand in for the close. The close is now the last line captured before each game's kickoff. We also removed two week-4 picks that a Tuesday refresh logged after their games had finished. A pick can no longer be logged once its game has kicked off. The live record is on the NFL Predictions page.
- Things we tested and switched off: EPA-magnitude scheme blending (worse on both holdout seasons), fantasy-production form and "conversion gap" signals (no out-of-sample value), a team's record against the closing line (does not predict its next game), and a pre-registered wind-under rule (54.1% on the holdout, below the bar we set in advance).
This matches thirty-five years of research: a team-strength model does not beat the NFL closing line on full-game sides and totals.
Where the work is aimed now
The research points to places where a small, honest operator can plausibly find value. Each one is judged by CLV, not by backtested win rate:
- Timing. Openers are softer than closes. We now snapshot the market every day, so the first snapshot of each week serves as the opener. Every pick carries an exact flag time, and the scorecard breaks CLV down by how many hours before kickoff the pick was issued. It also shows what the same picks would have earned at the opener. Nothing about when we issue picks changes until those numbers say it should.
- Line shopping. Taking −105 instead of −110 moves break-even from 52.4% to 51.2%. A multi-book odds feed shows the best available number for every ticket.
- Player props. Prop markets are thinner and less efficient than sides. The engine prices fair prop lines from its player projections and grades them against box scores. They are paper-traded only until they show positive CLV over a real sample.
- Quarterback and availability pricing, calibrated. Replace hand-set QB tiers with values fitted to how the market actually repriced past quarterback changes, and price other injuries in points with caps.
- Expert consensus, tested. We collected every pick from a published five-writer panel against the spread, 2019–2025: 8,178 picks on 1,642 games, each graded against the line the writer saw and against the close. The majority pick went 47.6% against the published line and 47.1% against the close. No writer's edge carried from one season to the next. A skill-weighted consensus, built only from earlier seasons, reached 51.2%. The line did drift toward the majority before kickoff, by +0.07 points on average. That clears the bar we set in advance, but it is worth far less than the bookmaker's margin. Consensus is a research input, not a pick source. One pattern the data surfaced, fading a 3–2 split, is now being tested on 2026 games only, because the seasons that suggested it cannot also prove it.
- Honest memory. A ledger records every change to our number and the market's, week by week. From it the engine learns when to trust itself and which terms to neutralize. It does not move prices on "hot team" narratives, which the data says have no predictive value.
Not planned: more machine learning on the same box-score features for sides, live betting, or teasers at −130. The evidence says these lose.
← Back to README.