NFL · Research · Methodology
The NFL spread and total backtest
A research model of home margin and total points, tested on 855 games it had never seen. Nothing on this page is a pick or a record.
The result in one paragraph
The model's predicted margin is well calibrated (slope 1.003 on the holdout, where 1 is perfect), and it is less accurate than nflverse's published spread: its average miss on the home margin is 10.239 points against 9.786 for the published spread, and 10.328 against 10.12 for the total. Once the published number is known, the model adds nothing measurable to it. Betting its disagreements would have lost money at standard prices on every rule tested in advance. We publish it because a model shown with its error is more useful than one shown without, not because it beats anything.
What it is tested against
nflverse, the open football data project, publishes one spread and one total for each game. That is the only historical number we have, and this page calls it nflverse's published spread and nflverse's published total, because that is all it is known to be.
For 1999 to 2024, nflverse does not document which sportsbook the numbers come from or at what time they were taken, and we could not establish either. For 2025 onward, the spread matches DraftKings, as shown on ESPN to within a point in 99.1% of 225 games we compared (exactly in 81.3%). Because the test seasons span both periods, the description for 1999 to 2024 applies throughout. There is one number per game, so nothing here measures whether a bettor could have taken a better number at any particular time.
Does the number behave like a betting number?
Before modelling anything we checked that the data reproduces what football betting markets are known to look like, over 7,276 games from 1999 to 2025:
| Check | Result | n |
|---|---|---|
| Home team covers nflverse's published spread | 49.05% (95% CI 47.88 to 50.21) | 7,081 |
| Favorite covers | 48.72% (47.56 to 49.89) | 7,050 |
| Game goes over nflverse's published total | 49.44% (48.29 to 50.6) | 7,170 |
| Push rate when the spread is exactly 3 | 9.02% | 1,153 |
| Push rate when the spread is exactly 7 | 6.52% | 491 |
| Pushes on half-point spreads (must be zero) | 0 | 3,229 |
| Favorites' straight-up win rate vs the moneyline's implied chance, 2010 to 2025 | 66.33% vs 66.33% | 4,362 |
It does: covers and overs sit just under 50%, pushes cluster on 3 and 7, and favorites win exactly as often as their price implies. We also tested 60 simple angles (home underdogs, divisional games, rest, wind, temperature, surface and so on) on 1999 to 2022. 12 looked interesting at the usual 5% level, about 3 would by chance alone, none survived a correction for testing so many, and none of the 4 strongest replicated on 2023 to 2025.
The model
Two ridge regressions, one for the home team's margin and one for total points. Inputs are team efficiency from play-by-play (EPA per play and success rate on offense and defense, overall and split by pass and run, over rolling windows), pace, rest and travel, a quarterback-change flag, and roof, surface, temperature and wind. Every input for a game uses only games played before it. The model never sees the published number.
Each season is predicted by a model fitted only on earlier seasons. On the development seasons, 2006 to 2022 (4,576 games), its average miss on the margin was 10.61 points against 10.32 for the published spread. The rules to evaluate, the thresholds and the warning signs were then written down and committed, and only after that were the 2023 to 2025 seasons opened, once.
Holdout, 2023 to 2025 (855 games)
| Target | Model avg miss | Published avg miss | Calibration slope |
|---|---|---|---|
| Home margin | 10.24 | 9.79 | 1.003 ± 0.083 |
| Total points | 10.33 | 10.12 | 0.999 ± 0.128 |
Against nflverse's published spread and total, with win rates excluding pushes, 95% intervals, and return at a flat −110 price (break-even 52.38%):
| Rule | W-L-P | Win rate | 95% interval | Return |
|---|---|---|---|---|
| Spread, model disagrees by 2+ points (primary) | 227-216-10 | 51.24% | 46.6 to 55.9 | -2.18% |
| Total, model disagrees by 2+ points (primary) | 208-220-3 | 48.60% | 43.9 to 53.3 | -7.22% |
| Spread, every game | 409-427-19 | 48.92% | 45.5 to 52.3 | -6.60% |
| Total, every game | 433-415-7 | 51.06% | 47.7 to 54.4 | -2.52% |
| Total, disagrees by 4+ (best development row, 55.85% on 2006 to 2022) | 81-75-1 | 51.92% | 44.1 to 59.6 | -0.87% |
By season, the primary spread rule went 67 of 138 (48.55%) in 2023, 76 of 148 (51.35%) in 2024, 84 of 157 (53.5%) in 2025: no more variation than chance predicts. For win probability, the moneyline with its margin removed was the better forecaster (Brier score 0.2100 against the model's 0.2224; lower is better).
How much can 855 games tell you?
Less than it looks. On every game, a win rate has a standard error of about 1.7 percentage points; on the roughly 450 games the primary rules select, about 2.4. A rule that truly wins 54% of the time could not be told apart from a coin flip on this sample, and the primary spread rule's interval runs from 46.6% to 55.9%. What the sample can show is accuracy: across 855 graded predictions on each target, the published number missed by less.
Limits, stated plainly
- One number per game, and for 1999 to 2024 an undocumented source. A bettor may not have been able to get it.
- Some inputs know a little more than a forecaster would have: the weather columns are what was observed at the game, the quarterback flag uses the actual starter, and the play-by-play model that produces EPA was fitted later. All of these favor the model.
- No injuries beyond the quarterback, no adjustment for opponent strength, and no market information. A stronger model is possible; this one was fixed in advance so the holdout result could be trusted.
- The model leaned toward home teams and overs on the holdout, so its disagreements are not a list of selections.
- NFL samples are small: 272 regular-season games a year. Nothing NFL on this site is official, and nothing from this model enters any record.
Research tier. The data is nflverse's schedules and play-by-play, 1999 to 2025. Related: team efficiency, weather.