The MLB overs edge doesn't transfer: -0.5% ROI across 2073 bets

What we found

We started from a result that held up. In MLB, betting game-total OVERS returned +7.2%. The question was whether that edge would transfer. We applied the same idea to a wider sample and graded every bet.

It did not transfer. Across n=2073 graded bets on 106 events, the record was 989-1084 for an ROI of -0.5%. The hit rate explains it. The market's prices implied a 47.9% win rate on these overs, and they won 47.7% of the time. The raw p-value is 0.481, which is nowhere near significance.

Verdict: NOT_SUPPORTED. The hypothesis that overs beat their price in general fails at n=2073. Where overs cleared the number, the market had already priced that in almost exactly. The +7.2% lane looks like it came from something narrower than "bet the over." Finding that narrower thing is the next job.

The trends

The 2026-09-27 board had 8 games, and it shows why a blanket overs rule is too blunt.

Totals skew over. 6 of 8 flagged total edges point over. The biggest is BAL @ NYY, where our model's fair total is 8.08 against a line of 7.5 (+0.58 runs). This is exactly the pattern the failed hypothesis tells us not to trust by default. A lean toward overs across the whole board is not an edge. The market already charges for it, as 47.9% implied against 47.7% actual shows. The useful question is which of those 6 has a real gap between the model and the line.

A pitcher trending the wrong way. DJ Herz (WSH, NYM @ WSH) has a 9.0 ERA over his last 5 starts and 9.0 over his last 3, against a season mark of 9.00. That is not a recent slump. It is a consistent level. A starter at that level is the kind of specific, game-level input a generic overs rule ignores.

A pitcher trending the right way. River Ryan (DET, PIT @ DET) has a 0.0 ERA over his last 5 with 9.0 K/9. A blanket overs bet treats his game the same as the Herz game. The results say the market doesn't make that mistake, so we shouldn't either.

The money is concentrated and pregame. Tracked whale money on the slate came to $587,301 across 9 campaigns. The largest was $117,165 on 510AZSD SD ML at 98% one-sided. Next were $95,697 on 505LADSF LAD ML at 97% one-sided and $85,703 on 510LAASEA SEA ML at 86% one-sided. The biggest single ticket on the board was $8,587. All of these are pregame moneyline positions, not live prices. The heaviest money is going to sides, not totals.

Our proven signals are still holding out-of-sample. TRAILING_LATE is at 11.4% ROI over 338 graded bets. VELO_DROP is at 9.0% ROI over 197 graded bets. Both are conditional on something specific happening: a team trailing late, or a pitcher's velocity dropping. Neither is a blanket rule.

What it means

Put side by side, the evidence points one way. Our signals that work are conditional and specific: TRAILING_LATE at 11.4% over 338 bets, and VELO_DROP at 9.0% over 197. The idea that failed was broad: overs as a category, at -0.5% over 2073 bets. The market prices broad tendencies well. In this sample it missed on overs by the gap between 47.9% implied and 47.7% actual, which is close to nothing.

So we are not scrapping the overs lane. We are narrowing it. The next test keeps overs only where there is a game-specific reason to expect runs, and drops the rest. The reasons we are testing:

- Model gap. Our fair total clears the posted line, as in BAL @ NYY at 8.08 against 7.5. - Weak starter. The starter is struggling on recent form, as Herz is at 9.0 over his last 5 and last 3.

If the original +7.2% came from games like these, the filtered version should separate from the 47.9% implied rate. The unfiltered version could not.

This is a research note, not a pick. We publish probabilities and their track record. We make no claim about what any team or pitcher will do on a given night. 21+ where legal.

What would change our mind: if the filtered, game-specific overs also land at the market's implied win rate on a comparable sample, we would retire the overs lane rather than keep refining it.