The first thing anyone is told about betting NFL games is that the public loves favourites, that this pushes favourite prices too short, and that the value is therefore on the underdog. It is repeated in every beginner's guide and most podcasts.
Across 5,139 priced regular-season games from 2006-2026, a flat stake on every underdog moneyline returns -3.60% per bet. A flat stake on every favourite returns -3.07%. Both are approximately the house margin, and neither is far enough from it to argue about — the underdog figure carries a t of about -1.74, which is not a result.
So the aggregate claim is simply wrong: there is no free money on dogs as a class. That is the boring half. The interesting half is what happens when you sort them by size.
The error grows with the number
Every underdog price implies a win rate. Here is what the prices implied against what actually happened, in brackets by the size of the spread.
| Spread | Games | Price implied | Dog actually won | Gap (points) | Return per bet |
|---|---|---|---|---|---|
| under 3 | 1,097 | 47.1% | 46.2% | -0.9 | -1.69% |
| 3 to 6.5 | 2,486 | 36.5% | 36.8% | +0.3 | +0.94% |
| 7 to 9.5 | 883 | 25.1% | 23.1% | -2.0 | -7.13% |
| 10 to 13.5 | 500 | 18.1% | 15.6% | -2.5 | -13.74% |
| 14 or more | 173 | 11.7% | 8.1% | -3.6 | -33.68% |
The gap is the column that matters. Under three points the dog wins about what its price says, and from three to six and a half it is marginally better than its price. Then it turns and keeps turning: -2.0 points, -2.5, and -3.6 in the largest bracket.
In other words the pricing error runs the opposite way to the folk wisdom, and it is concentrated exactly where the folk wisdom is loudest. Nobody writes excited posts about a two-point dog.
Double-digit dogs
Taking every underdog of 10 points or more: 673 games, 92 wins, a return of -18.87% per bet at an average price of +554. The t is about -2.33.
That is a loss rate several times the house margin, on a bet that is sold as the value play.
| Seasons | Games | Dog wins | Return per bet |
|---|---|---|---|
| 2006-2009 | 157 | 22 | -9.07% |
| 2010-2014 | 157 | 19 | -30.26% |
| 2015-2019 | 157 | 25 | -10.81% |
| 2020-2024 | 164 | 24 | -14.03% |
| 2025-2026 | 38 | 2 | -66.45% |
Negative in all of them. The sample in any one bracket is small and the most recent is the smallest, so no single row carries the claim — what carries it is that the sign never changes.
The part that stops this being a betting tip
The same dogs, against the spread rather than the moneyline, go 339-315-19. That is 51.8%, against the 52.4% a bet at -110 needs to break even. Slightly short, and nothing like an edge in either direction.
So the mispricing lives in the moneyline and not in the spread. The two markets price the same game and only one of them is getting the tail wrong.
And the reverse trade is not available either. A loss of 18.87% on the dog is not a gain of 18.87% on the favourite, because both sides pay the margin — which is why backing every favourite returns the -3.07% quoted at the top and not something attractive. What we have measured is a pricing error, not a bet.
Why it probably exists
A long price is a lottery ticket, and lottery tickets are easy to sell and hard to value. At +554 the question "how often does this actually happen?" has an answer most people never check, and the answer is less often than the ticket says. The bias is in the direction you would expect once the market is the thing being described rather than the crowd.
Method
Every regular-season game in nerddawg.nfl_game_history with a final score, a closing spread and both moneylines: 5,139 games, 2006-2026. Pick'em games are excluded because there is no underdog in them. Returns are one unit staked on the dog's closing moneyline, profit on a win and the stake on a loss. Implied probability includes the vig, which is the point — it is the price a bettor actually pays. The t-statistics are approximate: a flat stake on a long price has a heavily skewed return, so they are reported for sign and rough magnitude, not as exact tail probabilities. Measured 2026-10-13 and recomputed whenever this page is rebuilt.
