Why NFL Spread Models Land Near a Coin Flip

The bar is 52.38%, not 50%

At the standard price of −110, you risk $110 to win $100, so you need to be right 52.38% of the time just to break even. The gap between 50% and 52.38% is the sportsbook's cut. A method that is exactly as good as the line loses money; it has to be meaningfully better than the line, not just good at football.

What the line already knows

The things that look like obvious edges are the things the market prices first. Three examples from the 2015–2025 seasons, all from the same public game file:

• Home field. Home teams won by 1.92 points on average; the closing line made them 1.90-point favorites on average. They covered 48.9% (see what home field is worth).
• Rest. A team coming off a bye against an opponent on normal rest won by 1.51 points on average. The line had it favored by 1.54. Against the spread those teams went 160–158–8, or 50.3%.
• Being better. Favorites won 66.1% of their games outright but covered only 48.4%. Winning is easy to predict; winning by more than the market expects is not (see whether underdogs cover more).

A model built from public factors starts out knowing what the line knows. Whatever edge it finds has to come from weighting those factors differently than the market does, and the market has millions of dollars riding on getting the weights right.

In-sample vs out-of-sample

A back-test replays past games and counts how often a method would have been right. If the method was fitted on those same games, the back-test is in-sample, and it flatters the method: the fit has already seen the answers. An out-of-sample test fits on one stretch of seasons and scores a later stretch the fit never saw. It is the harder and more honest number.

The NFL Spread Dial shows its in-sample back-test and labels it as one. In its September 2026 fit, over 3,060 games since 2015, the default weights went 1505–1479–76 against the closing line, 50.4%. On games where its edge was 3 points or more, 51.6%. When it was fitted on 2015–2021 only and scored on 2022 onward, it was right about 48% of the time. None of those clear 52.38%, and the page says so.

Why a spread tool can still be useful

If a lean is not a pick, what is it for? Three things:

• Seeing why. Every game on the dial has a table that adds up to its lean: how many points came from team strength, recent form, rest, home field, recent cover rate and play efficiency. You can see when the numbers disagree with the line and which factor is doing it.
• Testing a belief. If you think rest matters twice as much as the fit says, turn that dial to 2× and watch the back-test. Usually it gets worse, which is itself an answer.
• Calibration. Knowing that a 3-point disagreement with the line has historically been worth a percentage point or two, not twenty, is the most useful thing a record can teach.

Be wary of any source that reports a long-run spread record far above 52.38% without saying whether it was measured in-sample. The most common way to get a great back-test is to fit and score on the same games.