J3P — Options, Filters & Permutations to Test
The full menu of what we can vary, what each tests, the source it's grounded in (Dale / Shannon / VCET), and the recommended order. Tracked in the
runs registry. Updated 2026-06-28.
Where we are: No play has a tradeable edge on the available ~14-day window. We've ruled out "it's the entry model" — late-pullback (v2.2), Dale-faithful pullback (v2.3), and breakout (CBO) all lose ~−0.2 to −0.4R. The window was choppy/low-trend; the leading hypothesis is now "trade only when the market is actually trending, in that direction" — i.e. regime filtering — plus we need more data than kti's 14 days to judge a trend system.
1 · Setups (the plays)
| Play | Dir | Source | Status |
| VWAP Pull-In (recapture) | long | Shannon (don't chase gap, wait for VWAP) | in |
| Trend Trade (Dale 1st-dev pullback) | long | Dale | in (v2.3 = faithful) |
| CBO Expansion (channel breakout) | long | VCET · Shannon-consistent (he enters breakouts) | in (v2.3) |
| Backside Scalp (fade extension) | short | SMB | in (rarely fires) |
| Short Trend + Short CBO | short | Dale/VCET (mirror of long) | TO ADD |
"Are we getting shorts?" — Almost none today: only Backside is short (~3% of trades); Trend & CBO are long-only. Adding the short side (Dale trend below −1σ, downside channel break) is item #1 below.
2 · Filters / regime gates (the variations to test)
A Add the SHORT side (mirror Trend & CBO to the downside)
Dale trend short = price holds below −1σ, sell the touch from below, trail; CBO short = downside channel break. Tests: whether we're leaving half the opportunity (downtrends) on the table. Dale/Shannon both trade shorts ("short weakness after the rip").
B VCET rotation-block / true-break-only (your past VCET gate)
Block ALL entries while vcet_state == ROTATION or price is still inside the channel (cbo_inside_channel); only allow entry on a confirmed channel break with vwap_width_expanding & expansion_score > 0. Tests: the exact gate that cut VCET bleed 57% in Phase-0a — does forcing "true break only" stop the chop losses? VCET v0.1.1 expansion gate.
C Market-regime filter — trade in the direction of SPY & QQQ
Only take longs when SPY and QQQ are in an uptrend that day (price > rising VWAP / above rising 5-DMA), only shorts when both are down; stand down when they disagree. Tests: does aligning with the broad tape remove the chop losses? Shannon ("more likely successful when trends are aligned with the trend of the overall market").
D Shannon short-term-MA filter — trade in the direction of the 5-DMA
Bias from the
5-DMA slope (Shannon's short-term trend MA, daily): longs only when the symbol's 5-DMA is rising, shorts only when falling; 50-DMA slope as a second confirmation.
Tests: a per-name trend gate (vs the market-wide gate in C).
Shannon ("shorter-term analysis using the 5-DMA … benefit of doubt to buyers when the trend is higher").
Shannon Fig 10.4 (Multiple Timeframes book) — TF→MA lengths, the concrete mapping: 1-min short=50; 2-min=20; 5-min short=40, int=100, long=200; 10-min=20/50/100; 30-min=7/17/33; Day=10/20/50/200. So on our 5-min backtest the short-term trend MA = 40-period SMA (use its slope for direction). This table is already encoded in your mtf_trend.py / MTF Trend Matrix.
E AVWAP-slope bias (Shannon's PRIMARY rule — partly in already)
Longs only when VWAP slope > 0, shorts only when < 0. J3P already requires this on the long side; formalize it both ways. Shannon ("rising AVWAP → long bias, avoid shorts").
3 · Other permutations I'd add (not yet on your list)
- Grade gate (A/B only). Selectivity is the one thing that's been positive (A-grade +0.17R vs C −0.34R). Test "trade A/B only" on every variant — likely the biggest single lever.
- Time-of-day filter. Midday (11:00–13:30 ET) is consistently the worst. Test "skip midday" / "open + power-hour only."
- Name-character filter. J3P gets chopped in hot, volatile names (TSLA/COIN worst; GOOGL/MU best) — the opposite of what we assumed. Test restricting to moderate-volatility clean trenders (e.g. ATR%/price band), not the wildest movers.
- Combined gate (the likely winner): rotation-block (B) + market-regime (C) + grade A/B (#1). Each filter alone may be weak; the combination — only A/B setups, on a true break, with the tape trending — is the real test of "trade only when everything lines up."
- Exit-model sweep. Trend trail too tight (18% win = whipsaw). Test trailing on the 5-DMA / a 2-bar swing / chandelier-3ATR instead of the +1σ-loss.
4 · The data problem (your "can't go past 14 days" point)
You're right — kti/karb caps at ~14 days of minute bars, which is far too little to judge a trend system (we keep landing in one choppy window). The engine already abstracts the data source (bar_source), so the fix is a drop-in adapter, not a rewrite. Options, by job:
| Path | Depth | Best for | Verdict |
| New historical data adapter (Polygon.io / Databento / Alpaca / restore Liberator) | months–years, full basket | Deep systematic backtesting | ✓ best for the edge question — same engine, swap bar_source, get years × the whole basket |
| Kore / g8trade backtest (KITE) | longer, server-side | Porting the strategy to execution / forward-test | The path after we have edge — it's the exec layer, not the research layer |
TradingView Pine strategy() | years on TV data | Single-name visual validation | Good cross-check, but 1 stock at a time & Pine backtest limits — not a basket edge tool |
My recommendation: for the research/edge question, add a Polygon (or Databento) minute-bar adapter → re-run every variation above over 6–12 months × the full basket. That's the only way to know if any of these filters actually work across regimes. Use Kore/g8trade later for forward-test/execution once a variant earns it, and TV as a per-name visual sanity check. (Restoring Liberator does the same job as Polygon if its entitlement comes back.)
5 · Recommended order
- Add shorts (A) — we're blind to half the market.
- Rotation-block (B) + market-regime (C) — the core "only trade true breaks when the tape trends" test.
- Layer grade A/B (#1) + skip-midday (#2) on top.
- In parallel: stand up a deep-history data adapter so steps 1–3 are judged on 6–12 months, not 14 days.
- Shannon 5-DMA (D) and the exit-model sweep (#5) as the next round.
Every run lands in the
registry with its filter combo, so we can compare all of them head-to-head.