NFL prop overs do not carry the MLB leak: +1.3% ROI at n=470 fails the test

The hypothesis was simple and, in our view, the most likely of the cross-sport transfers to hold: books misprice player prop OVERS in MLB, so they should misprice them in the NFL too. Same product, same behavioural bias, different sport. We graded it. It does not hold.

What we found

The test covered 470 NFL yardage and touchdown prop OVERS across 41 events. The record was 332-138, good for a 1.3% ROI. That sounds like a win until you put it next to the price. The market-implied win rate on those tickets was 69.7%. The actual win rate was 70.6%. The raw p-value on that gap is 0.4502, which is a coin flip. The verdict is NOT_SUPPORTED.

This is the cleanest kind of negative. It is not that the overs lost. They hit almost exactly as often as the price said they would, which is what an efficient market looks like. The MLB leak, whatever its cause, is not a general property of prop OVERS. It is specific to baseball pricing, and the NFL book has priced this particular product correctly across 470 tickets.

One caveat on the sample. With n=470 across 41 events, a small true edge could still be hiding under the noise. But a 1.3% ROI at p=0.4502 is not something we can bet, and we will not describe it as an edge.

The trends

The reason a null result matters is that the working signals keep working, and the contrast tells you where to point the next test.

The proven-signal book is holding out-of-sample. VELO_DROP is at 11.2% ROI over 176 graded bets. TRAILING_LATE is at 10.6% ROI over 300 graded bets. Both are baseball-native reads, one on pitcher fatigue and one on in-game state. Neither is a behavioural bias about how bettors like overs. That is the pattern: the edges we can verify are structural to baseball, not generic to the betting public.

Today's board is a good illustration of where the structural edges live. Concentrated money is landing pregame, and the top campaign is $319,026 on 315SFSTL SF ML at 99% one-sided. The same game carries $267,403 on Over 9 at 63% one-sided. In 310CWSCLE there is $127,072 on CLE ML at 99% one-sided. The biggest single ticket on the board is $47,000, and total tracked whale money on today's games is $1,459,958 across 12 campaigns.

The model leans underdog in 8 spots today, among them DET ML +239, WSH ML +182 and KC ML +153. The top moneyline read is DET @ TOR: DET ML +239, model 38.8%, EV +31.5%. That is worth reading against where the money sits, because the whale campaigns above are concentrated on favourites at 99% one-sided while the model is finding its value on the other side.

Totals skew over. 8 of 8 flagged total edges point over, led by MIA @ AZ, where the model fair is 9.68 against a line of 8.5, a gap of +1.18 runs. Note the irony against the NFL result: in baseball the model itself is generating over leans from the simulation, not from a bias about the public. That is the distinction the NFL test just drew for us.

Pitcher form is trending in opposite directions in LAD @ CIN. Andrew Abbott is trending the wrong way, with an 8.96 ERA over his last 5 and 15.0 over his last 3 against a season mark of 4.78. Blake Snell is the hottest arm on the board at a 0.96 ERA over his last 5 with 11.5 K/9. Reads like VELO_DROP exist precisely because this kind of divergence is measurable before the market fully prices it.

What it means

The NFL prop-OVERS transfer is dead as a standalone signal at n=470. We are not going to keep betting it hoping the sample turns. The useful output is the shape of the failure: a 70.6% actual against a 69.7% implied says the NFL prop book is well calibrated on this product, so any NFL edge will have to come from a structural input the book is slow on, not from a public-bias story.

What we are running next: the same grading framework, but on NFL props conditioned on a measurable state variable, the way VELO_DROP conditions on pitcher velocity and TRAILING_LATE conditions on game state. Unconditional overs failed. Conditioned overs are the next test, and we will publish that result whether it passes or not.

21+ where legal. We publish probabilities and their track record, not outcomes.

What would change our mind: the same hypothesis holding at a materially larger sample with a p-value well below 0.4502, or a conditioned version clearing the bar unconditional overs could not.