Methodology
How the model works, and where it fails
Every figure on this page comes from a walk-forward test on this project’s own data: ratings for a given week are built using only games played before it. None of it is theoretical.
The headline
The model does not beat the closing line
Closing line error
10.26 pts
mean absolute error on margin
This model
11.06 pts
worse by design, not by accident
Model side ATS
49.93%
52.38% needed at −110
Fitting w · model + (1 − w) · closing line over 5,199 games puts the best weight on our model at zero. Error rises with every unit of model weight. The closing line already contains what the model knows.
An earlier version of this product flagged “+EV value bets” from model-vs-market disagreement. It fired on 91.7% of available selections and returned −3.01% ROI across 11,488 simulated bets. A detector that fires on almost everything is measuring its own noise, so we removed it.
What is real
Key numbers
Football scoring is discrete and lumpy. Across 6,967 regular-season games since 1999, a final margin of 3 occurs in 15.0% of games and 7 in 9.1% — while 9, no further from the mean, occurs in just 1.6%. Margin 3 is over nine times more likely than margin 9.
So the model prices margins on that lattice rather than with a smooth curve: it re-weights a normal by empirically fitted key-number factors (w(3) = 2.79, w(7) = 1.85, w(9) = 0.50) and renormalises. Fitted on 1999–2019 and tested on 2020–2025, that improves log-likelihood by +0.108 nats per game.
The practical consequence: at a spread of exactly 3 — 15.9% of all closing spreads — the observed push rate is 8.7%. This model says 7.3%. A plain normal says 2.7%, underestimating by a factor of three on the single most common line in the sport.
Where it is weaker, stated plainly: at a spread of exactly 7 the model says 5.0% against an observed 2.0–2.6% in the test window. Those samples are small (n = 50 and 38), but the model over-weights 7 there.
Calibration
Fitted constants
walk-forward 1999-2025, 6967 games
EPA → points
15.12
per unit net EPA/play
Home field
2.22 pts
fitted for the 2026 era
Margin σ
14.23
QB scale
32
points per unit EPA/dropback
Home-field advantage is not a constant. Fitted by era it has decayed from 2.56 points (1999–2007) to 1.71 (2020–2025), so each era uses its own value rather than a single average that would misprice both ends of the archive.
Limits
What is not validated
The quarterback adjustment is not covered by the backtest above. Testing it walk-forward would need each quarterback’s form as of that week, and the cached form is a season aggregate — using it would leak results the model could not have known. Rather than report a flattering number built on look-ahead, we leave the term untested and say so.
It is at least bounded sensibly. Quarterback EPA is regressed toward the league mean by sample size, because passers under 100 dropbacks span −0.995 to +0.739 EPA per dropback while those over 300 span only −0.192 to +0.250 — a fourfold spread that is noise. Without that shrinkage a 20-dropback third-stringer pinned the adjustment against its ±10-point clamp and dominated whole predictions. With it, no game in 2025 reaches the clamp.
Staking
Quarter-Kelly, and why
Full Kelly is optimal only when the probability you feed it is correct. Ours is measurably less accurate than the market’s, so full Kelly is a route to ruin rather than to growth. Where a stake is shown at all, it is a quarter of full Kelly.
The honest framing throughout this product: a model disagreeing with a sharper market is usually evidence that the model is wrong. NFL Intelligence is built to help you understand football. It is deliberately not built to tell you it has found free money.
Data
Sources and coverage
- · Play-by-play, schedules, rosters and historical closing lines from nflverse.
- · EPA per play, success rate and explosive rate are computed from play-by-play, because the published team-season files contain no defensive EPA and no success rate at all.
- · Live scores and standings from the ESPN public endpoints.
- · Next Gen Stats are published only through 2024, so anything depending on them is hidden for later seasons rather than shown as zero.
Not affiliated with the NFL. Back to scores.