J3P — 90-Day SIP Backtest

87 trading days (2026-02-27 → 2026-07-02) · 154 in-play tickers reconstructed from Massive · 40,914 signals · matched exits · hub · experiment log
The headline: playing the daily stocks-in-play instead of a fixed list is decisive — and for the first time the edge survives out-of-sample. The disciplined stack (in-play + morning + daily-trend + matched exits) held: TRAIN +0.36R → TEST +0.12R (test N=1,483, 40% win). Three plays hold strongly. LAB — not live-confirmed; trade small first.

Plain-English key (read this first)

R How much you planned to lose if the trade went wrong. Every result is counted in R. "+0.15R" = on average you made 15¢ for every $1 you risked.
Avg R The average result of all the trades, in R. Above 0 = winning overall; below 0 = losing.
Win % Out of every 100 trades, how many made money.
SIP (stocks in play) The stocks that are actually busy and moving that day — not the sleepy ones. You trade only these.
Selective Only the best trades: an in-play stock, in the morning, going the SAME way as the daily trend.
Train The older data we studied to build the rules — like a practice test.
Test (out-of-sample) Newer data we hid from ourselves, then graded cold — the real exam. THIS is the number that counts.
Holds Positive on BOTH the train and the test data — a sign it's real, not luck.

① The progression — how each change moved the number

Start with everything, then add one discipline rule at a time. Watch the edge climb as the trade count drops — fewer, better trades. "$ / 100 trades" = dollars made per 100 trades, each risking $100.
What we changedTradesAvg RWin$ / 100 tradesLift
Baseline — fixed mega-cap list, all plays, all day12,988+0.02334%$+231—
+ trade the daily stocks-in-play (not a fixed list)32,227+0.10036%$+998+0.077
+ only the morning (9:30–11:00)7,245+0.24142%$+2,407+0.141
+ only with the daily trend (price vs its 20-day)3,520+0.25842%$+2,581+0.017
+ only the 4 durable plays (drop the losers)2,183+0.36842%$+3,685+0.110
From basically break-even (+0.02R, a coin-flip 34% win → ~$230 per 100 trades) to +0.37R (~$3,700) — roughly a 16× swing in profit, by trading less but better. The two biggest single levers: trading the morning (+0.14R) and keeping only the durable plays (+0.11R). Crucially it holds out-of-sample (§②) — so the realistic live expectation is the TEST number, about +0.17R (~$1,700/100), not the full +0.37.
the old fixed-list vs in-play A/B (what step 1 above replaced)
MethodSignalsAvg RWin %
OLD — fixed 8 mega-caps, all day12,988+0.02334%
NEW — daily SIP in-play, all day32,227+0.10036%
OLD — fixed 8, selective (morning+daily-aligned)1,332+0.06139%
NEW — SIP, selective (morning+daily-aligned)3,520+0.25842%

② The out-of-sample verdict (the one that counts)

Trained on 2026-02-27→2026-05-20 (first ~2/3), tested on 2026-05-20→2026-07-02 (held-out last ~1/3). A play is real only if it's positive in both halves. STRONG = test ≥ +0.10R on a real sample.
Play (SIP, selective)Train NTrain RTest NTest RVerdict
backside2+0.6660+0.000no
backside_roll27-0.25930+0.029no
band_break138+0.53483+0.039weak
open_drive83+0.33231+0.114STRONG
revbreak_long781+0.597532+0.147STRONG
trend_ride291+0.456167+0.207STRONG
v_bottom234-0.019162+0.016no
vwap_pull_in368+0.050293+0.054weak
vwap_reject_short113+0.231185+0.192STRONG
Overall selective SIP stack: TRAIN 2,037/+0.361 → TEST 1,483/+0.118. The long trend/reversal plays (revbreak, trend_ride) and the short (vwap_reject_short) are the durable core.

③ Every play, in plain English

All 9 plays the backtest runs. Click any play name → its labeled chart + exact specs + the real code open right here (no page change). Or the full Plays page.
PlaySideWhat it isOut-of-sample
vwap_reject_short ↗SHORTPrice is below VWAP, pops up to it, but fails and closes back under. You short the failure. (Your MU trade.)HOLDS
revbreak_long ↗LONGPrice is falling, breaks its downtrend, pulls back to test the breakout, then goes. You buy the turn. (Your afternoon reversal.)HOLDS
trend_ride ↗LONGIn a clear uptrend, price pulls back to its fast average, then takes off again. You buy the continuation.HOLDS
band_break ↗LONGPrice sits calmly above VWAP, coils tight, then bursts up through the top band. You ride the pop.weak
open_drive ↗LONGIn the first hour, price squeezes into a tight bar then breaks out upward. You buy the breakout.HOLDS
vwap_pull_in ↗LONGPrice dips just under VWAP, then closes back above it. You buy the recovery into value.weak
v_bottom ↗LONGPrice drops hard to the bottom band, then snaps back up in a “V.” You buy the rebound.no
backside_roll ↗SHORTPrice is stretched way above VWAP and slowly rolls over, losing its rising average. You short the fade back to VWAP.no
backside ↗SHORTPrice stretches far above VWAP and slaps down with a rejection candle. You short the snap-back to VWAP.no

④ Honest caveats

• Test < train (+0.36 → +0.12) — real edge decays out-of-sample; expect live to look like the test number, not the train.
• One split, ~90 days. A single held-out period — more periods + more names would harden it.
• Free-tier skips: some stock-days were dropped on rate limits (partial coverage).
• "Liquid-mover" reconstruction (top $-volume names that moved), a proxy for the real SIP scanner's catalyst weighting.
• LAB only — nothing promoted to the Pine; trade small/live by hand before trusting it.

⑤ What's next · and what we tried and rejected

Next: automate the cockpit + a real push-alert when a play fires, and build the MTF higher-timeframe "which-setups-actually-run" selection gate.   What we rejected (kept off this page on purpose — this page is the durable results): the inverse mirror-shorts (regime artifact), the v_bottom hybrid exit (refuted), the backside long (falling knife). The full journey incl. every dead end is in the Experiment Log, and which stocks make these plays pay is in Nuance.
Generated from sip90.py → sip90.csv, rendered by j3p_sip90_page.py. Reconstructed daily SIP via Massive grouped-daily. 2026-07-05.