A p-value of 0.0085 on 1,872 bets, and we still threw it out

Every backtest shop eventually meets a number like this one: p = 0.0085. Run one test, see that figure, and you would publish. We ran 84.

The hypothesis was clean and old as the sport: *September offenses are behind September defenses.* If true, early-season college football totals should be set too high, and blindly betting unders should print. So we graded it — every qualifying early-season under, no cherry-picking.

What we found

Across n = 1,872 bets the record came back 929-943-13, a 49.6% win rate against a breakeven of 52.4%, for a -5.3% ROI. The direction of the deviation was strong enough to produce a raw p = 0.0085 — but this test ran inside a batch of 84 hypotheses, and under our Benjamini-Hochberg procedure (threshold 0.01026) the batch produced 0 survivors. Nothing in the sweep, this test included, gets promoted to the live book. A tout would headline the p-value. The discipline is the batch: when you run 84 tests, small p-values show up for free, and a result that can't clear the multiple-comparisons bar is a coincidence with good marketing.

That's the near-miss worth writing about. At 49.6% over 1,872 bets, the market's early-season totals are not systematically inflated in a way you can bet. The idea fails at scale, and saying so is the finding.

The trends

The same discipline shapes how we read today's MLB board (8 games, slate of 2026-09-02), where the directional signals are loud.

Concentrated money is landing pregame: the top campaign is $154,941 on NYY/LAA Over 6 at 62% one-sided, with $153,499 on STL/LAD Over 9 at 72% one-sided and $148,371 on the Marlins moneyline at a striking 96% one-sided. In total we're tracking $1,146,314 across 12 campaigns on today's games, with the biggest single ticket at $22,957.

The model's totals skew the same way the money does: 7 of 8 flagged total edges point over, led by CWS @ HOU, where the model's fair total of 9.29 sits +0.79 runs above the 8.5 line. Part of that read is a pitcher trending hard in the wrong direction — Davis Martin's ERA has deteriorated from a 4.32 season mark to 9.95 over his last 5 starts and 14.62 over his last 3. On the other side of the ledger, Reid Detmers is the hottest arm on the board at a 1.5 ERA over his last 5 with 8.7 K/9.

The model also leans underdog in 6 spots today — MIL ML at +251 (model 35.2%, EV +23.7%), MIA at +590, MIN at +160 — which is worth reading against where the concentrated money sits, since the whale flow above is mostly chasing overs and the Miami favorite side.

What it means

None of those directional reads is a finding yet, and that's the point of pairing them with the CFB result. The signals that *have* cleared the bar keep working out of sample: VELO_DROP is at 15.2% ROI over 147 graded bets, and TRAILING_LATE at 13.9% over 227. Both survived the same correction machinery that just killed a p = 0.0085 result — which is exactly why we trust them.

So the early-season unders hypothesis is dead in its blanket form. What we test next is the conditional version: whether the effect concentrates in specific slices of the early slate rather than across all 1,872 spots, run inside a fresh, pre-registered batch so the correction is honest. If a slice survives BH there, it gets a forward-tested paper period before a dollar touches it.

We'd change our mind if a pre-registered re-run of the hypothesis cleared its batch's correction threshold out of sample — until then, 0 survivors means 0 bets. 21+ where legal.