Read this one differently from the other findings here. The rest grade prices against what really happened. This one grades the board against the simulator, which is a claim about what the model says over every price on offer rather than about realized results. It answers a narrower question: of everything a book puts up on a night, how much of it does an independent estimate call a bad price?
Across 252 MLB games and 18 slates, 55.5% of the board came back Negative EV. Another 20.1% graded roughly fair, so three quarters of every price on offer had no edge in it either way. The split held on every single slate: the negative share never left the band from 53.1% to 58.1%, so it is not one strange night doing the work.
Every selection on the board
| Verdict | Share | Note |
|---|---|---|
| Negative EV | 55.5% | the majority of what is on offer |
| Roughly fair | 20.1% | inside the app's own fair band |
| Positive EV | 24.3% | see the note below on what this is not |
| No edge at all | 75.7% | negative and fair together |
Counting each book's own quote separately instead of the best price anywhere makes it slightly worse, at 58.6% negative across 83,790 rows. That is the expected direction. Shopping for the best number strips out part of the margin, and most people do not shop.
By price, where it gets expensive
| Price band | Negative EV | No edge |
|---|---|---|
| Longshots, +400 or longer | 71.0% | 80.0% |
| Dogs, +200 to +399 | 69.7% | 80.6% |
| Modest dogs, +100 to +199 | 55.7% | 77.4% |
| Heavy favorites, -200 or shorter | 55.6% | 81.3% |
| Favorites, -199 to -100 | 48.1% | 70.6% |
Seven in ten long prices grade out Negative EV. That is the shape behind the two things people buy most of, the cheap anytime home run and the long parlay, and it lines up with the one market here that has been graded against real outcomes rather than simulations: 18,468 anytime home run props came in around -23%.
Split by market instead of by price, batter props were worst at 56.4% negative, then totals at 49.2%, pitcher props at 49.0%, spreads at 47.0% and moneylines at 46.5%. Props are worse than game lines, which is the direction to expect. The margin is thicker where the market is thinner.
Method
Every MLB slate with both published odds and simulations over the window was taken whole: every player prop and team line, both sides, at the most-quoted point, with the best price across books, giving 23,059 distinct selections. Each one was graded by the production scoring code against that game's 5,000 pitch-by-pitch simulations, not by a reimplementation, and priced at the raw posted number with the vig still in it, because that is what a bettor actually pays. The verdict bands are the app's own: roughly fair when the edge is within 2 points and the relative expected value within 5%, otherwise the sign decides. Selections the scorer declined, where a player was not in the simulations or the market is unsupported, were counted separately and excluded rather than defaulted to a verdict. Tool: board-ev-census.
Inside the noise floor
The 24.3% that graded positive is the simulator disagreeing with the market, and this measurement cannot say how much of that is real edge and how much is model error. Those are not separable without grading the same selections against outcomes, so nothing here treats that quarter of the board as profitable bets, and neither should anyone reading it. The negative-side figures do not depend on resolving it.
Top comments (0)