J3P Exit Ladder — Out-of-Sample Verdict (James's 60/20/20 Shannon/Dale ladder, graded honestly)
Broad sample:
28 names × 8 sessions (locked SIP allowlist + 7/1 in-play). Exit = your ladder: 60% at +2σ (T1), 20% at +4σ (T2), 20% trailed on the higher-TF higher-low. Swept trail×T1×T2,
optimized on the first 5 sessions, reported on the held-out last 3.
← J3P hub ·
replays →
🧪 LAB — this is a disciplined out-of-sample grade of your strategy. It is honest, not flattering.
🕹️ COCKPIT — the Pine. Nothing here is promoted — nothing held out-of-sample.
⛔ Verdict: no edge survives the honest test. The thin 3-name sample (that showed +0.25R) was over-fit. On 28 names:
the pull-in has no positive config even in-sample (best −0.129R); the backside-roll's in-sample sweet-spot (+0.157R) flips to −0.121R out-of-sample (textbook overfit). And with 76 parameter combos over ~150–220 in-sample trades, you have only ~2–3 trades per parameter — far too thin to optimize. This is a data/sample-size problem, not an exit-tuning problem. The fix is the deep-data adapter, not more knobs.
Broad-sample coverage (N per play)
| Play | N (broad) | N (old thin) |
| vwap_pull_in | 338 | 42 |
| backside_roll | 376 | 41 |
| backside | 49 | 2 |
| open_drive | 11 | 1 |
| band_break ⭐ (new, trend-day) | 380 | — |
| v_bottom ⭐ (new, reversal) | 236 | — |
| trend_ride ⭐ (new, continuation) | 358 | — |
Out-of-sample grade (in-sample-best config → held-out result)
| Play | N | IS-best (trail·T1/T2/run) | IN-SAMPLE R | OUT-OF-SAMPLE R | IS trades/param | Read |
| VWAP Pull-In (long) | 338 |
drop · 40/30/30 |
-0.129 |
+0.079 |
2.91 | no positive config even IN-SAMPLE — no edge |
| Backside Rollover (short) | 376 |
swing · 40/0/60 |
+0.157 |
-0.121 |
1.89 | IS sweet-spot does NOT hold out-of-sample (overfit) |
| Band-Break Breakout (long, trend) ⭐ | 380 |
bar5 · 80/0/20 |
-0.096 |
-0.063 |
2.11 | no positive config even IN-SAMPLE — no edge |
| V-Bottom Reclaim (long, reversal) ⭐ | 236 |
bar5 · 40/30/30 |
-0.078 |
+0.293 |
2.16 | no positive config even IN-SAMPLE — no edge |
| Trend-Ride Continuation (long) ⭐ | 358 |
bar5 · 40/30/30 |
-0.102 |
-0.129 |
1.68 | no positive config even IN-SAMPLE — no edge |
band_break (your AMZN 7/1 trend-day play) catches trend breakouts (AMZN 7/1 10:28 = +0.79R) but on the broad sample most breakouts fail in chop → net negative, ~36% win. Same story: catches the good days, no confirmed edge on ~2 weeks.
Green cell background = OOS-positive; the point is that the IS-optimal cell does not stay green out-of-sample.
OOS avg-R surface — Backside Rollover
| trail↓ / T1/T2→ | 40/0 | 40/10 | 40/20 | 40/30 | 50/0 | 50/10 | 50/20 | 50/30 | 60/0 | 60/10 | 60/20 | 60/30 | 70/0 | 70/10 | 70/20 | 70/30 | 80/0 | 80/10 | 80/20 |
|---|
| bar5 | -0.03 | -0.03 | -0.03 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.05 | -0.05 | -0.05 | -0.05 | -0.05 | -0.06 | -0.06 | -0.06 | -0.06 |
|---|
| bar10 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.05 | -0.04 | -0.04 | -0.05 | -0.05 | -0.04 | -0.05 | -0.05 |
|---|
| swing | -0.12 | -0.12 | -0.12 | -0.13 | -0.12 | -0.12 | -0.13 | -0.13 | -0.12 | -0.13 | -0.13 | -0.13 | -0.13 | -0.13 | -0.13 | -0.13 | -0.13 | -0.13 | -0.13 |
|---|
| drop | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.04 | -0.05 | -0.04 | -0.04 | -0.05 | -0.05 | -0.04 | -0.05 | -0.05 |
|---|
Rows = trail method · cols = T1%/T2% split. Cell = out-of-sample avg R. Note there's no robust green plateau — the surface is noise around zero.
OOS avg-R surface — VWAP Pull-In
| trail↓ / T1/T2→ | 40/0 | 40/10 | 40/20 | 40/30 | 50/0 | 50/10 | 50/20 | 50/30 | 60/0 | 60/10 | 60/20 | 60/30 | 70/0 | 70/10 | 70/20 | 70/30 | 80/0 | 80/10 | 80/20 |
|---|
| bar5 | +0.05 | +0.05 | +0.05 | +0.05 | +0.05 | +0.05 | +0.05 | +0.05 | +0.05 | +0.05 | +0.04 | +0.04 | +0.04 | +0.04 | +0.04 | +0.04 | +0.04 | +0.04 | +0.04 |
|---|
| bar10 | +0.11 | +0.10 | +0.10 | +0.10 | +0.10 | +0.10 | +0.10 | +0.09 | +0.10 | +0.10 | +0.09 | +0.09 | +0.10 | +0.09 | +0.09 | +0.08 | +0.09 | +0.09 | +0.08 |
|---|
| swing | +0.09 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.07 | +0.07 | +0.07 | +0.07 | +0.07 | +0.07 | +0.07 | +0.06 | +0.06 | +0.06 | +0.06 | +0.06 | +0.06 |
|---|
| drop | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 | +0.08 |
|---|
Regime filter (Shannon "trade with the daily trend")
Daily 5/10, 5/20, 10/20 MA-alignment gate (trade only when the daily trend agrees with the play), same OOS discipline:
| Play | Gate | baseline R | IN-SAMPLE R | OUT-OF-SAMPLE R | Read |
| vwap_pull_in | 10/20 aligned | −0.061 | +0.124 | −0.280 | IS lift flips hard negative OOS (overfit) |
| backside_roll | 5/20 aligned | +0.022 | +0.184 | +0.018 | IS +0.18 decays to ~breakeven OOS (noise) |
No regime gate creates a robust out-of-sample edge — the in-sample lifts don't survive. Adding a regime dimension on this sample only worsens the ~2-trades-per-parameter problem.
What this means
- Don't promote anything to the Pine. No exit ladder, trail, or split held out-of-sample on this window.
- The trail hurts the fade: a mean-reversion short shouldn't trail a runner — covering into VWAP (no runner) beat every trailed ladder.
- The real blocker is sample size (~2 weeks, kti cap). The deep-data adapter (Polygon/Databento, parked) is the unlock — 6–12 months would give the trades-per-parameter to actually optimize.