Better team getting points: p=0.0002, and still not a finding

What we found

We tested a hypothesis every bettor has had: when our own ratings say a team is better and the market still makes it a dog, the market is wrong. Sample: 347 NBA games. Result: 149-198-3, a 42.9% win rate against a 52.4% breakeven, for -18.0% ROI.

The raw p-value is 0.0002. Under a Benjamini-Hochberg correction across the 108 tests in that batch, the threshold was 0.00143 — so this one clears it. It is, statistically, one of the least likely results in the batch to be noise.

And it is not a finding.

The direction is inverted. A hypothesis that predicts outperformance and delivers 42.9% against a 52.4% bar has not located an edge; it has located the vig plus a real bias pointing the other way. The honest read is that the market's pricing of our "underrated dog" is better than our rating is, at n=347. Out of 108 tests in that batch, survivors: 0. We are not rolling this into a live signal, and the inverted version needs its own out-of-sample sample before it means anything either.

That is the whole lesson of the corpus. Significance is a filter, not a conclusion. Run 108 tests and the arithmetic of multiple comparisons will hand you something that looks like a headline anywhere else on the internet.

The trends

What a real signal looks like, by contrast, is the graded book. TRAILING_LATE is holding out-of-sample at 11.0% ROI over 301 graded bets. VELO_DROP sits at 10.6% ROI over 177 graded bets. Those are records, not p-values — sample accumulated forward, after the rule was fixed.

On today's 7-game MLB board, the same discipline applies to reading flow. Concentrated money is landing pregame: the largest campaign is $97,628 on 915PHINYM Under 9 at 54% one-sided — big dollars, but barely a lean in ticket terms. Compare 915KCHOU, where $85,471 sits on Over 9 at 68% one-sided. Same order of magnitude in money, very different conviction. A third, 940DETCWS, shows $82,267 on Under 9 at 58% one-sided. Across the day we are tracking $698,802 over 12 campaigns, with the biggest single ticket on the board at $10,050.

Lines are moving in both directions and the moves disagree with each other. Drifting toward: 138MINLAA-MIN 51.5c to 53.5c (+2.0) and 915PHINYM-NYM 49.5c to 51.5c (+2.0) over the last 20 hours. Drifting away: 138MINLAA-LAA 50.5c to 46.5c (-4.0) and 915PHINYM-PHI 51.0c to 48.5c (-2.5). Note that PHI is drifting away while the largest campaign on the board is on that game's Under — money and line are telling adjacent, not identical, stories.

The model's own totals read skews one way: 7 of 7 flagged total edges point over, led by KC @ HOU at a fair 9.69 against a line of 8.5, a +1.19-run gap. A one-directional sweep across every flagged edge is exactly the shape that should make you suspicious of a shared input rather than seven independent discoveries — the same instinct the 108-test batch demands.

Pitching trends run to both extremes. Wilber Dotel (PIT) is trending the wrong way: 13.5 ERA over his last 5, 13.5 over his last 3, against a season mark of 4.84 — the recent form is worse than the season, not noise around it. At the other end, Michael King (SD) is at a 1.4 ERA over his last 5 with 6.7 K/9. The top model moneyline read, SD ML -4900 at a model 98.6%, prices out to EV +0.6% — a rounding error of an edge on a price where almost nothing is in doubt.

What it means

A number that passes a significance test is the beginning of the work, not the end of it. We publish the tests that fail because the failures are what make the survivors credible. Next in the queue: re-running the underdog test split by rating-gap size and by rest, to see whether the inversion is uniform or concentrated in a subset our ratings systematically misprice.

What would change our mind: a forward, out-of-sample sample of the inverted version that clears the 52.4% breakeven. Until then this is a rejected hypothesis, not a fade. 21+ where legal.