THE LAB · bets-data

H1 ACCUMULATING 49d to verdict

DOG BUCKET

Do underdog ML picks (+120 to +200) produce positive ROI in a forward paper sample?

IN-SAMPLE FINDING

+11.8% ROI · n=300 · CI lower bound barely clears break-even; 5 buckets tested

FORWARD STANDING

n = 2 / 200 · ROI +136.0% CI [+108.6%, +163.4%]

H2 ACCUMULATING 49d to verdict

SPREAD OVER ML

Do spread picks outperform ML picks in a forward paper sample?

IN-SAMPLE FINDING

Spread +2.9% vs ML -3.9% — gap +6.8pp · CIs overlap; selection bias unresolved

FORWARD STANDING

Spread n = 0 / 150 (accumulating)
ML n = 23 / 150 (-23.9%)

H3 DEFERRED 49d to verdict

LATE POST

Do picks posted within 2h of game start produce higher ROI than picks posted 6h+ early?

IN-SAMPLE FINDING

Late +4.1% vs Early -2.7% — gap +6.8pp · time-to-game is a proxy (±1-2h error)

FORWARD STANDING

n = 30 · game-start timing unknown (proxy failure)

H4 PAUSED ROLLING

NBA POOL

Do NBA ML capper picks produce positive ROI across the 2026-27 season?

IN-SAMPLE FINDING

+15.0% ROI · n=658 · one playoff window; self-selection likely; multiple-comparison risk

FORWARD STANDING

PAUSED — tracking begins Oct 2026 (2026-27 NBA season)

H5 OFF-TRACK 49d to verdict

CONVICTION

Do high-conviction (3u+) plays outperform low-conviction (1u-) plays forward?

IN-SAMPLE FINDING

High-conv +9.5% (n=636) vs Low-conv -3.1% — gap +12.6pp · survivorship risk; units unreliably stored

FORWARD STANDING

High-conv n = 6 · ROI -29.3%
Low-conv n = 18 / 200 · ROI -13.4%

H6 ACCUMULATING ROLLING

CONSENSUS CARD

Does the daily consensus card produce positive ROI over 250+ graded tickets?

IN-SAMPLE FINDING

Pilot 9W-9L · ROI +0.8% · n=18 · CI spans 92pp — far below significance threshold

FORWARD STANDING

Accumulating — 0 / 250 graded tickets

H7 DELEGATED 26d to verdict

FADE THE INFLATED

Can you produce positive ROI by fading cappers who are proven losers in-sample?

IN-SAMPLE FINDING

In-sample losers identified via EDGE-REPORT.md §3 · gate fully delegated to fade_gate.py

FORWARD STANDING

Tracking delegated to fade_gate.py — see Fade Gate section below

H7 · Fade Gate — per-capper tracking 26d to verdict

Gate start: 2026-07-12 · Verdict: 2026-08-09 · Automated via fade_gate.py

ACTIVE MrBigBets

Accumulating — no graded picks yet

p_value=None / 1.0 means insufficient data (n<5 non-push outcomes)

ACTIVE ThisGirlBetz

Accumulating — no graded picks yet

p_value=None / 1.0 means insufficient data (n<5 non-push outcomes)

ACTIVE Underdog Sniper

Accumulating — no graded picks yet

p_value=None / 1.0 means insufficient data (n<5 non-push outcomes)

How pre-registration works

01 The gate — metric, threshold, minimum n, and verdict date — is locked to hypotheses.json before any forward picks are observed. Changing these after seeing data is goalpost-moving and invalidates the test.
02 Forward n increments only when a graded result is logged with a timestamped receipt. Unknowns remain "—"; they are never fabricated or imputed.
03 Seven hypotheses run simultaneously. Expected false positives by chance at alpha=0.05: ~0.35. Clearing p<0.01 without clearing Bonferroni-adjusted p<0.0071 is noted as suggestive — not confirmed at family-wise alpha.
04 All tests are paper-only. In-sample findings describe the historical dataset — they are descriptive, not predictive claims. The forward gate must clear before any edge is asserted.

WHAT'S ON TRIAL