Does Week 1 carryover exist in CFB? We asked 6 ways. Zero survived at n=1,243

We don't test one angle and publish the winner. We test the family. This week the family was `native_week1` in college football: the idea that what a team was last season — turnover-lucky, turnover-cursed, efficient on a yards-per-play basis — carries tradable information into the first games of the next one. We ran 6 tests across 1,243 bets. Zero survivors. Weighted ROI across the family: -3.74%.

What we found

The headline cut was the classic regression trade: fade teams whose last season was turnover-fuelled, on the theory the market still prices the luck as skill. In Week 1 it lost — -6.0% ROI over 128 bets, p_raw 0.2369, which is to say indistinguishable from noise. The mirror image — buy teams whose last season was turnover-cursed — showed +7.4% ROI over 144 bets, but at p_raw 0.1763 that is a coin that came up slightly heads, not a signal. The prior-season yards-per-play edge never even produced a usable result at n=45 — the sample was too thin to grade.

The one cut that deserves a second look is the widened window: fading last season's turnover-fuelled teams across weeks 1-4 rather than Week 1 alone. That ran -9.5% ROI over 521 bets at p_raw 0.0115. Read carefully: that is a *negative* return on the fade — the fade itself lost money at a p-value that clears casual thresholds. But 0.0115 raw, inside a family of 6 tests, is exactly the kind of number multiple-comparison correction exists to humble. We are not trading it. We are re-running it on a holdout before we believe the market actually *overprices* the regression story.

So the bias hypothesis fails at n=1,243. That's the finding, and it's worth exactly as much as a winner: it tells us where the edge isn't, which prunes the search tree. Next up for this family: re-testing the early-window fade out-of-sample, and rebuilding the yards-per-play cut with a sample large enough to grade.

The trends

While the CFB corpus grinds, the MLB board shows why we keep testing at all — the market is not a static object. Concentrated money is landing pregame today: the top campaign is $126,244 on 940ATLCHC Over 10 at 72% one-sided, with $119,361 on 840CWSCLE Over 7 at 51% one-sided and $112,311 on 907DETTOR Under 8 at 71% one-sided behind it. Total tracked whale money on today's games: $1,013,294 across 12 campaigns, with the biggest single ticket at $23,405.

Lines are drifting with that money: 040SDCOL-SD moved 57.0c to 64.5c (+7.5) in the last 20 hours, 840LADCIN-LAD went 65.0c to 68.5c (+3.5), and 940NYYMIN-NYY went 59.5c to 62.5c (+3.0) — while 040SDCOL-COL drifted away, 42.0c to 36.5c (-5.5). Totals skew over: 8 of 8 flagged total edges point over, led by NYY @ MIN, where the model's fair total is 10.15 against a line of 8.5 (+1.65 runs). The model also leans underdog in 4 spots today (MIN ML +174, COL ML +852, MIA ML +125 among them) — worth reading against where the money sits. And Tomoyuki Sugano (COL) is trending the wrong way: a 9.38 ERA over his last 5 starts, 7.88 over his last 3, against a season mark of 5.30 — the context behind the model's COL ML +852 read at 13.7% and EV +30.1%.

What it means

The contrast is the point. Six plausible CFB stories, tested honestly, produced nothing tradable. Meanwhile the signals that *did* survive this process keep paying out-of-sample: VELO_DROP is at 10.9% ROI over 171 graded bets, and TRAILING_LATE is at 10.1% ROI over 290 graded bets. That's what a survivor looks like — and why we're comfortable publishing a family with 0 of them. Most ideas die. The discipline is showing you the graveyard, not just the graduates.

What would change our mind: the early-window turnover fade holding its sign and significance on a fresh out-of-sample run.

*Probabilities, not promises. 21+ where legal.*