MLB totals: both sides close to fair, neither measurable

Both sides of a totals bet pay close to even money and look symmetrical. Graded against real final scores over 1,067 games, the over came back about -4.4% per bet and the under about -0.7%, both inside their noise floor. The direction survives; the size does not.

Written by Jesse, NegativeEV. Last updated 2 August 2026.

• Sample: 1,067 MLB games, pushes excluded
• Window: 2026-04-19 to 2026-07-31
• Priced at: best price across books on each side, at the most-quoted total
• Measured: 2026-08-02

Realized EV per bet	Figure
Every over	-4.4% (± 3.0, z = -1.5)
Every under	-0.7% (± 3.0, z = -0.2)
Mean per-book two-sided hold	4.7%

An earlier July-only pass over 342 games put the over at -10.6%, narrowly outside its noise floor, and that framing is retired: the full archive window pulls both sides back inside their floors. Overs still grade worse than unders, and that is all this census supports.

The per-book hold of 4.7% is the margin in the quotes. It is not the realized cost of either side and the two should not be read as the same quantity - the sides came out at -4.4% and -0.7% around it.

Method

For each game the most-quoted total point is taken, then the best over price and best under price across books at that point, and both sides are graded against the real combined run total from the mapped play-by-play corpus. Pushes are excluded. Games missing from the corpus are excluded rather than defaulted.

Inside the noise floor

Neither side clears its own noise floor. On 1,067 games the over is -4.4% ± 3.0 (z = -1.5) and the under is -0.7% ± 3.0 (z = -0.2), and both intervals cover zero. Smaller windows of this same market have read differently: a 279-game pass put the over at -7.8% ± 5.8 and a 342-game pass at -10.6% ± 5.2, the one figure that did clear its floor and the one the full archive retires. The decimals are estimates rather than measured costs. The large-sample prop and ATP censuses remain the sharper measurements.

Written up for readers: https://negativeev.com/about/totals All findings: https://negativeev.com/research