J3P Setups — what's triggering & executing
The exact rules behind every trade in the backtest, straight from the engine (
j3p_backtest.py). 5-min bars, RTH 09:30–15:55 ET, ≥50 bars warmup. VWAP & ±σ bands are the VCET session VWAP. ·
see trades on the chart → ·
edge tables →
1 · VWAP Pull-In LONG · vwap_pull_in
Brian Shannon — "don't chase the gap, wait for VWAP." Price shakes out toward/below VWAP (sucks in shorts), then recaptures VWAP and runs.
| Trigger | close > VWAP AND prior close ≤ prior VWAP (a fresh recapture) AND VWAP slope ≥ 0 (VWAP not falling). |
| Tier | By how deep the shakeout went — lowest low over last 6 bars vs the bands:
band1 shallow (above −1σ) · band2 to −1σ · band3 deep (≤ −2σ). |
| Entry | Next bar open + slippage. |
| Stop | −2σ (LB3) − 0.1·ATR — the param-surface "LB3 override" (LB3 stops won historically). |
| Targets | T1 = +2σ, T2 = +4σ. Ladder: 80% off at T1, 20% runner to T2; stop → breakeven after T1. |
| A | band3 + RVOL≥2.0 + MA stack 5>20>50 + MA20 rising + band-spread≥1.5% + VWAP slope≥0 + gap>3% |
| B | band2 + RVOL≥1.5 + MA aligned + VWAP slope≥0 + gap>2% |
| C | band1 + RVOL≥1.5 + MA aligned |
| – | fires but meets none of the above (the bulk of pull-ins — and where most of the bleed is) |
2 · Trend Trade LONG · trend_trade
Trader Dale 1st-deviation pullback — in an uptrend, a shallow pullback to the first deviation band, then continue. Stop = reaction point.
| Trigger | Uptrend regime (close>VWAP, VWAP slope>0, VCET state ∈ TREND_UP / EXPANSION_CONFIRMED / EARLY_EXPANSION) AND target room (+2σ > close) AND a shallow pullback whose reaction low (lowest low last 3 bars) sits in [VWAP, +1σ] AND resumes back above +1σ (close>+1σ and close>prior close). |
| Entry | Next bar open + slippage. |
| Stop | reaction low − 0.1·ATR (just under the pullback low — the Dale "reaction point"). |
| Targets | T1 = +2σ, T2 = +4σ (same 80/20 ladder). |
| A | RVOL≥2.0 + VWAP slope>0 + late session (14:30–15:55) |
| B | RVOL≥1.5 + VWAP slope≥0 |
| C | everything else (the bulk — and the worst-performing cell) |
3 · Backside Scalp SHORT · backside_scalp
Fade an over-extended move back to VWAP. (Fires rarely — only 21 of 804 trades.)
| Trigger | Price extended ≥ 2·ATR above VWAP AND (5-MA rolling over OR a 5-MA break) AND a rejection bar (new local high, then close<open and close in the lower 60% of the bar's range) AND RVOL≥1.5. |
| Entry | Next bar open − slippage (short). |
| Stop | recent swing high + 0.25·ATR (above the failed high). |
| Target | VWAP — cover into VWAP. Single target (no +4σ runner). |
| A | ext≥3·ATR + 5MA rolling + 5MA break + RVOL≥2.0 + late session |
| B | ext≥2·ATR + RVOL≥1.5 |
| C | everything else |
How trades exit (the reason shown on every trade)
stop | Price hit the stop → exit at stop −slip (long) / +slip (short). 86% of C-grade trades die here. |
be_stop | (long) After T1 took 80%, the runner's stop was moved to breakeven and price came back → exit ≈ entry. |
t2 | (long) The runner reached +4σ. Happens ≤5% of the time — the +4σ target is effectively a phantom. |
vwap_cover | (short) Covered into the VWAP target. |
eod | Still open at 15:45 ET → flattened at the bar close. |
Execution model (how the sim fills & sizes)
Constant $100 risk per trade → shares = $100 ÷ per-share risk (so a $18 name and a $600 name are comparable). Fills at next-bar open. Conservative slippage (≥1¢ or 0.10×bar-range on entry; 0.05×ATR adverse on stops). 0.5¢/share commission each side. One long + one backside-short slot per name (no pyramiding). R = net P&L ÷ initial risk.
What the backtest actually found (804 trades, 52 liquid names)
- No play has reliable edge as encoded — Trend −0.282R, Pull-In −0.161R; both stay negative out-of-sample. Backside +0.018R but only n=21.
- Grade is the real signal — A-grade +0.174R / 69% win vs C-grade −0.336R / 41%.
- C-grade dies by stop-out (86%) — enters into chop, gets shaken.
- The +4σ (T2) target is a phantom (hit ≤5%) — the exit ladder needs realistic targets.
- The short side barely fires (21/804) and midday (12–13 ET) is worst.
Refinement direction: kill the C-grade firehose, fix the +4σ target, tighten entries / avoid midday — then re-run the loop.