Whet for the MLB.
The sport we have the deepest record in — and the one where we can show you exactly which bets have made money and which have lost it.
Baseball is where the sample lives.
Whet scans MLB every cycle, all season. It is the only sport with enough graded picks for a segment to clear the sample floor the forward test set in advance, so it is the only sport where we are willing to quote a rate at all.
Below is the MLB slice of the same public ledger /proof serves. It updates on its own; nothing here is typed in by hand. Losing segments are in red and are not removed.
28,747 graded picks across 8 segments, on at least 60 game dates. 7 of them have reached the sample floor.
swipe the table for ROI and the 95% CI →
| Segment | Graded | Win rate | ROI | 95% CI (win · ROI) |
|---|---|---|---|---|
| prop · over | 14,166 | 46.5% | −1.32% | 45.4–47.8% · −4–1.6% |
| prop · under | 8,666 | 60.4% | +10.9% | 59–61.7% · 8.3–13.4% |
| spread · under | 1,670 | 55.6% | −2.25% | 52.2–59% · −8.2–3.9% |
| spread · over | 1,507 | 59.9% | +5.66% | 56.3–63.3% · −0.9–12.1% |
| total · under | 1,313 | 52.6% | −0.62% | 49.2–56% · −7.3–5.8% |
| total · over | 1,094 | 54.4% | +10.77% | 50.6–58.2% · 3–18.8% |
| moneyline · over | 219 | 57.3% | −2.51% | 50.2–64.1% · −14.3–8.9% |
| moneyline · under | 112 | Too early — 112 of 200 graded picks. No rate is published below the floor. | ||
Pre-registered basis: every pick the frozen engine issued after 2026-07-14 and graded against the real outcome. Floors — 200 graded picks and 14 independent game dates — were written down before the results existed. Wins and losses both, every segment, no subset.
The unders carry it. The overs lose.
The most useful thing our own record says is unflattering. MLB prop overs are the single largest segment in the ledger, they are past the sample floor, and their return is negative. Prop unders, on the same engine over the same games, are the strongest segment we have.
You can read both numbers in the table above rather than take our word for it. The gap is not small and it has held across the whole frozen window.
So the engine penalises its own overs: an MLB prop over loses 35 points of Whet Score before it is ranked against anything else. That is a deliberate handicap applied to our own output because the graded record earned it — not a filter we invented and never ran.
The gates are market-quality gates.
Whet does not price a baseball game from box scores. It prices the market against itself, and rejects prices that fail these tests on every cycle:
- Two-way consensus — the fair line is built only from books quoting both sides of the market. A market nobody prices on both sides is skipped.
- Three independent books, minimum — the best price has to be one at least three independent books agree on, within two points of implied probability. One book quoting an outlier is an error, not an edge. Fitted on 21,782 graded MLB picks.
- Stale-price rejection — a quote more than 15 cents off the consensus is dropped rather than scored.
- Minimum edge — 3% EV on player props, 2% on spreads, totals and moneylines. Below that the vig eats it.
- Injury gate — props for players ruled out are removed before your board renders.
- The over penalty — the 35 points above, applied to our own picks.
No park factors, no BvP, no lineup gate.
This page used to describe a park-factor model, a batter-versus-pitcher filter with a 20-plate-appearance floor, a bullpen-fatigue adjustment and a gate that held props back until the lineup card posted. Those were written as product copy for code that never received the data it needed, so none of them has ever changed a pick you were shown. They are gone from this page rather than quietly left up.
What is left is what the ledger can back. If we build the lineup gate for real, it will appear here with a graded record attached to it, the same as everything else.
See today's MLB board?
Sharp ($39.99/mo) is the full live board, every cycle, no delay. The free board shows a slice of the same picks 10 minutes later.