Wednesday night, New England at Seattle. The market has the total at 44.5 and Seattle laying three. Our totals model looked at the same game and said 51.1 — a fat six-and-a-half-point disagreement, the kind of number that makes for a confident post and a bad week.
We published 44.8 instead. Here is why.
01What the closing line already knows
Before trusting any model, it is worth measuring the thing it is arguing with. Across 1,087 regular-season games from 2022 through 2025, the closing total missed the actual result by 10.19 points on average, with a residual standard deviation of 13.01 and a bias of just +0.72. Overs went 48.8%.
That standard deviation is the number to sit with. An NFL total is so noisy that a perfect forecast of the true average would still miss by about ten points a game. Most of what looks like model error is not error at all. It is the sport.
02What our model’s disagreements have been worth
So we tested it the only way that means anything: walk-forward, training on nothing but games that had already finished, then regressing what actually happened against what the model claimed would happen.
The slope came back at 0.043, with a standard error of 0.111 and a p-value of 0.70. In plain terms, about four cents of every claimed dollar has historically shown up in the result, and even that is statistically indistinguishable from zero. Graded at real prices across the 2025 season, the model went 49.2% on 181 bets for −5.8% ROI — and it got worse as its claimed edge grew. The six-point disagreements were its worst bucket, not its best.
That is not a broken model so much as an honest one meeting an efficient market. Our features are trailing efficiency numbers. Every sportsbook models those too. By the time a line closes, that information is priced.
Which is how 51.1 becomes 44.8. We shrink every projection toward the market by the amount its own track record justifies, and this model has not earned much. Against a 44.5 line, our published probability on the over is 50.9% where the market’s own de-vigged number is 49.2%. A gap of a point and a half, on a game where the honest answer is that we do not know much the market does not.
03Where the actual edge was
Now the part that matters. Six books were quoting the over at 44.5 on Wednesday evening, and they were not quoting it at the same price.
Same side. Same number. Same game. At −102 you need the over to hit 50.50% of the time to break even. At −108 you need 51.92%. That 1.43-point spread in breakeven is, by itself, comparable to the entire probability edge our model claims — and unlike the model’s edge, it is not an estimate. It is arithmetic on prices you can see.
Run the expected value and the point lands hard. Our 50.9% over, bet at lowvig’s −102, returns about +$0.76 per $100. The identical opinion, bet at −108, is negative. The bet is not good or bad. The price is good or bad, and the same wager flips sign depending on where you place it.
A winning $100 over ticket pays $98.04 at lowvig and $92.59 at −108. Nobody is going to build a season out of five dollars and change on one game. But that is exactly the point: this is the margin that compounds across a year, and it is available to anyone willing to check more than one screen. The book hold on this game was 4.7%, which is what you are paying for the convenience of not looking.
04What we are actually saying about this game
Seattle’s implied team total is 23.75, New England’s is 20.75. Those come straight from the market — 44.5 split by a three-point spread — and they are the most defensible numbers on the page, because they are derived from prices rather than predicted from history.
Our model likes the over. Our track record says not to lean on that. If you take the over anyway, take it at −102, because that is the difference between a bet with a sliver of value and a bet without one.
- Market total 44.5, Seattle −3. Implied team totals: SEA 23.75, NE 20.75.
- Model says 51.1 raw; 44.8 after shrinking to its measured record.
- Our over probability 50.9% against a market fair price of 49.2%.
- Best over −102 (lowvig); worst at the same number −108. Same bet, opposite sign.
The uncomfortable truth of totals betting is that the market is very good and your model probably is not. We would rather publish that, and the price that actually pays, than a six-point edge we cannot defend.