We tested every way you'd think to profit from capper picks — tail the career winners, fade the losers, chase hot streaks, fade cold ones, bet the consensus — walk-forward, so no strategy ever saw the future. Zero survived. Most lost money outright.
| Strategy | Variants | Best result | Verdict |
|---|---|---|---|
| Tail proven winners | 9 | all negative | DEAD |
| Chase hot streaks (the 15-0 guys) | 6 | ≈ zero, none significant | DEAD |
| Fade the losers | 5 | −5.5% ROI (vig eats the fade) | DEAD |
| Consensus (from old DBs) | 4 | untestable — 0.6% coverage | → REPLAY |
| Bet everything | 6 | −1.2% ROI | DEAD |
The one mechanical "survivor" (NBA subset, +14%) failed the fragility screen: 82% of its profit came from ONE month (May playoffs), gone by June. Flagged artifact by its own finder AND the independent verifier. And the red-team caught the real prize: the consensus strategy had a look-ahead leak — it was peeking at same-week results. Fixed → its edge dropped to zero. A lesser system ships that leak as a "finding." Ours caught it and killed it.
Why everyone loses, in one line of arithmetic: the pool needs 52.62% to beat its own juice and hits 50.95%. Cappers aren't dumb — they're average, and average pays the book −2.35% forever.
Your idea, executed: pull what cappers actually posted on past days, build the card with zero knowledge of outcomes (results firewalled from the card-builder), grade against real scores. The 5-day pilot went 4–1. So we scaled it to 24 days — and watched the streak revert, exactly like honest sampling does.
| Sample | Record | Units | ROI | 95% CI | Read |
|---|---|---|---|---|---|
| Pilot (5 days) | 4W–1L | +3.10u | +62% | — | 1-in-6 luck |
| Full (24 days) | 9W–9L | +0.14u | +0.8% | −45% … +47% | coin flip |
This is the most honest thing on the page. The exciting pilot number regressed to dead-even the moment we added data — which is what a lucky streak is supposed to do. Verdict: the consensus card is currently indistinguishable from a coin flip that pays the juice. Not proven dead — proven unproven. Needs ~250 graded tickets (bare minimum) for real significance; the forward archive now grows it automatically every day, for free.
The realistic frame: betting income from picks alone = not supported by the data. Discipline (band, sizing, no-chase) + CLV-verified line shopping = where a real but small edge could live. The system now measures exactly that, daily, for free.
BOARD_REFRESH_OFFThe upgrade script is staged. Paste one line into Claude any time:
ultracode — run UPGRADE-PROMPT.md from bets-data, full power
What it fires: replay scale-up to 30+ days · CLV verdict analysis · fade-gate scoring · new-source scouting (Discord/Reddit signal lanes) · board UX wave · fresh adversarial audit of everything since the last one. Details in UPGRADE-PROMPT.md.