๐ฌ J3P Trust Report โ verify before you build
Every load-bearing claim adversarially re-checked against the source data (6 independent verifiers, each hunting the traps we've been burned by). Only build on โ
. ยท
hub ยท
log
Bottom line (updated โ the verification went deeper than the first pass): the headline claim does NOT survive. The backtest only ever traded names it knew โ with hindsight โ had already moved that day; rebuild the watchlist on honest pre-market data and the out-of-sample edge falls from +0.17R to ~0. So the +0.17 was largely a look-ahead artifact, and the live SIP "diverging" from it was actually the live being MORE honest, not broken. What still stands: riding the 20-EMA beats capping the band (a relative exit finding, independent of name selection). What's now the ONLY honest measure of the real edge: the live tracker (Phase 2), capturing pre-market-selected results forward from today. Until that live scorecard proves an edge over a real sample, treat J3P as UNPROVEN โ trade tiny or on paper.
Does the edge survive removing look-ahead? (the biggest finding)โ REFUTED
NO. The backtest picks its 'in-play' names using THAT DAY'S completed move (close-vs-open โฅ3%, or range โฅ4%) โ which you cannot know until the day is over. It only ever traded names that, with hindsight, had already moved.
- Rebuild the watchlist on pre-market-knowable data (the PRIOR day's movers โ the best honest proxy) and the out-of-sample edge collapses from +0.17R to +0.004R โ essentially ZERO (median โ0.49, win 37%). Every play is at or below zero out-of-sample, including the short (โ0.04).
- So the +0.17 'edge' was almost entirely an artifact of look-ahead name selection โ about 39% of the selective trades are names you could only know were in play after the close.
- This does NOT kill the idea โ it means the demonstrated edge is UNPROVEN pre-market. The only honest test now is FORWARD: the live tracker (Phase 2) is capturing real, pre-market-selected results starting today. Trade tiny / on paper until that live scorecard shows a real edge over enough signals.
Does the walk-forward test still pass?โ WEAKENED
The literal claim is true โ 5 of 6 time-blocks positive, +0.20R going forward. But it's fragile.
- Same tail problem: forward median โ0.38R, wins 38%, a handful of trades carry it (top-5 = 33% of the profit).
- The most-recent block is NEGATIVE and the edge is FADING month-over-month: May +0.50 โ June +0.17.
- Treat the recent softening as a real live warning, not noise โ watch it.
Do all 4 plays work on new data?โ WEAKENED
Only ONE is statistically durable. The other three are positive but shaky, and one is basically unproven.
- vwap_reject_short (the short) = the one solid edge โ significant on every test, survives dropping its best trades, works in up AND down markets (its 'only works in a falling market' reputation is FALSE).
- open_drive = UNVALIDATED โ only 31 test trades, and it turns NEGATIVE without its 3 best. Don't trust it yet.
- revbreak_long & trend_ride = positive but NOT statistically significant, top-heavy, and revbreak's July is sharply negative (โ1.50). All four win under 45% of the time.
Is the discipline 'recipe' right?โ WEAKENED
The end result (+0.37R with all rules on) checks out, but the recipe was oversold.
- The starting point wasn't break-even โ it was already +$660 per 100 trades, not +$200. So the real lift is ~5ร, not the '16ร' we claimed.
- The order is wrong: trading the right (SIP) stocks matters MORE than the 4-plays filter โ we had it backwards.
- The 'with-the-trend' rule adds ~3ร what we said (~+0.05, not +0.02). Treat the per-rule numbers as ranges, not exact.
Does riding the 20-EMA beat capping the band?โ
CONFIRMED
YES โ ~7ร more on new data, and durable (holds every month, survives dropping the top 20 winners). Keep this exit.
- But it's a MEAN effect only: most rides LOSE (median โ0.25R, win 42% โ actually LOWER than capping's 44%); a few rare runners carry it.
- One stat was wrong: riding beats capping on ~48% of trades, not 37% (the 37-42% is the win rate). Fix the copy.
- Any teaching page MUST show the full distribution โ never sell it with a single +7R monster trade.
Does the LIVE scanner match the backtest?โ WEAKENED
The play LOGIC matches (the scanner uses the backtest's exact definitions โ good). But the INPUTS diverge โ this is the most important finding.
- ๐ด The live SIP watchlist overlaps the backtest's PROVEN universe by only 1โ4 of 15 names. The live system is watching DIFFERENT stocks than the ones the +0.17 was proven on.
- The daily-trend gate mixes two price-adjustment bases and goes STALE silently after the last cached day.
- Live uses a different data vendor (kti) + cache than the backtest (Massive). โ Until the live SIP list is aligned to the backtest's construction, live results WON'T match the lab, no matter how well the plan is followed.
Nothing here is auto-executed. LAB only. Verifiers recomputed from sip90.csv / exit_variants.csv with medians, t-stats, top-N concentration, and clean-boundary split re-cuts. "Tail-driven" = the average is positive but the median trade loses โ show the distribution, never a single monster trade. Generated 2026-07-07.