VWAP Pull-In Gate — Validation Audit

Read-only validation: does the Green/Yellow/Red gate improve decision quality vs the old generic vwap_pull_in label? · Built 2026-06-09 · Machine summary: vwap_pull_in_gate_validation_summary.json
Data: Liberator-validated backfill (vwap-band-entry-surface, 323 trades; VWAP validated 0.0028% vs kti). Plain vwap-pull-in n=132, net −$298.13. No database query — computed from on-disk validated artifacts.
Proven separates winners/losers in the data Working not yet validated per-trade Rejected tested and NOT supported as specified
PARTIALLY SUPPORTED — REVISE before promotion

The gate concept improves expectancy — but almost entirely through the RED exclusion (driven by gap > 1.5%) and the post-entry checkpoints, not through the RVOL thresholds. Bucketed RVOL does not separate winners from losers (the calmest <1.0 bucket is the worst). Instrument type is a stronger separator than RVOL. Promote the proven parts in stages; rework the RVOL axis first.

Executive summary

⚠ The reframe this audit forces

The original framing was "calm tape good, hot tape bad." The bucketed data says the enemy is more specific: big gaps (>1.5%), crypto-proxy/small-cap names, and uncut fades — not high RVOL. High-RVOL (>2.5) trades lost only −$58 on n=13; the calmest (<1.0) trades lost −$320. Keep the gate's spirit; change which dials it turns.

What is data-proven

  1. Pass when |gap| > 1.5% (both directions). −$341.93 / 24.1% win / n=54. The strongest, fully-populated pre-trade filter.
  2. Cut the 3-minute fade. Plain-PI fade state wins 6.7% at 3 min (median −$7.29). Should be a hard rule.
  3. Only trend-hold after 10–15 min expansion confirms. Early-expansion = 77.8% win / 100% reached-next-band at 15 min.
  4. Instrument type filters. Plain Pull-In is a megacap-tech setup; it fails on crypto-proxies and small/high-beta names.
  5. IBIT and CVX are robust repeat offenders.
  6. The RED exclusion improves expectancy (−$298 → ~+$59 if RED passed).

What remains a working rule working

What is rejected (tested, not supported) rejected

Q1 · RVOL buckets (plain Pull-In) not supported

RVOL populated on 108/132 (82%). Buckets are non-monotonic — note the <1.0 row.

RVOL bucketnWin %Total P&LAvg/tradeAvg winAvg loss
< 1.0 (calmest)5332.1%−$320.62−$6.05+$18.21−$17.51
1.0–1.5 (green)2634.6%−$23.80−$0.92+$17.26−$10.54
1.5–2.5 (yellow)1643.8%+$14.27+$0.89+$18.10−$12.49
> 2.5 (red)1330.8%−$58.20−$4.48+$18.39−$14.64
NA (not populated)2441.7%+$90.22+$3.76+$26.14−$12.23

Read: the gate predicts <1.0 should be safest and >2.5 worst. The data shows the opposite at the low end. RVOL does not separate as designed — likely confounded by instrument type (low-RVOL crypto-proxies sit in <1.0).

Q2 · Gap buckets (plain Pull-In) supported

Gap populated on 132/132 (100%). Bucketed by |gap| to match the gate.

|Gap| bucketnWin %Total P&LAvg/trade
< 0.5% (green)4040.0%−$36.40−$0.91
0.5–1.5% (yellow)3847.4%+$80.20+$2.11
> 1.5% (red)5424.1%−$341.93−$6.33

Signed split of the >1.5% bucket — both directions lose

DirectionnWin %Total P&L
Gap up > +1.5%3426.5%−$200.09
Gap down < −1.5%2020.0%−$141.84

Read: the |gap| > 1.5% RED rule is the cleanest, most reliable filter in the whole gate. The green/yellow split below 1.5% does NOT separate (yellow actually beat green). Gap's value is the RED line, not the gradation.

Q3 & Q4 · First-3 / 5 / 10 / 15-minute behavior (plain Pull-In) supported

State at window3 min5 min10 min15 min
Fade — win %6.7% (n30)9.4% (n32)10.3% (n39)17.5% (n40)
Flat rotation — win %44.2%35.0%31.2%34.6%
Early expansion — win %62.5% (n8)72.7% (n11)70.6% (n17)77.8% (n18)
Early-exp % reached next band100%81.8%100%100%

Q3 read: a fade at 3 minutes wins 6.7% — the "respect the 3-minute failure" rule is decisively supported and should be a HARD rule. Q4 read: early expansion confirmed by 10–15 min wins ~70–78% and reaches the next band; trend-hold should only be allowed after this. Plain-PI vs vwap-pull-in → trend: PI→Trend trades are favorable at 5/15 min ~73% (early-expansion ~47–60%) vs plain-PI ~34% favorable / ~8–14% expansion — the conversion shows up fast.

Q5 · Symbol damage (plain Pull-In) IBIT + CVX

Losers

SymnWin %P&L
IBIT robust1030%−$185.51
CVX robust520%−$44.63
AMD tiny30%−$69.65
ETHU tiny30%−$69.35
RGTI tiny30%−$38.98
META955.6%−$19.13

Winners

SymnWin %P&L
MSFT4100%+$133.54
AMZN944.4%+$43.93
TSLA666.7%+$43.39
GOOGL560%+$32.67

Read: only IBIT (n=10) and CVX (n=5) are robust-sample losers. AMD/ETHU/RGTI are 0% but n=3 — warnings, not a proven blacklist. Don't hard-code the tiny-sample names.

Q6 · Instrument type (plain Pull-In) supported

Instrument typenWin %Total P&LAvg/trade
Megacap tech5151.0%+$151.91+$2.98
ETF / crypto proxy2740.7%−$232.26−$8.60
Small-cap / high-beta / other5418.5%−$217.78−$4.03

Read: the cleanest single separator after gap. Plain Pull-In is a megacap-tech play. Classification is heuristic (symbol→type map) and cells are moderate-n, so direction is proven, exact thresholds are working.

Q7 · Gate vs old generic label supported via RED exclusion

Old generic plain Pull-In: n=132, −$298.13, 35.6% win. Partitioned by the combined RVOL+gap gate:

Gate bucketnWin %Total P&LAvg/trade
GREEN (rvol 1.0–1.5 & |gap|<0.5)1030.0%+$41.38+$4.14
"green" but rvol <1.0 & calm gap1741.2%−$229.31−$13.49
YELLOW3250.0%+$124.45+$3.89
RED5925.4%−$356.72−$6.05
Ungradable (rvol NA)1442.9%+$122.07+$8.72

Read: the gate works by isolating RED (−$356.72) — pass it and the book flips to ~+$59. YELLOW was net positive, so half-sizing it would have cut profit (−$62), not reduced drawdown. The RVOL≥1.0 floor matters: the "very calm" <1.0 sub-bucket lost −$229. GREEN proper is positive but tiny (n=10).

Q8 · Do the mistake tags explain the leak? partly

Mistake tagComputable now?Finding / missing field
traded_red_gateYES59 trades / −$356.72 — explains the entire leak
chased_hot_tapepartialoverlaps red gate (same population here)
ignored_3min_failureYESfade at 3 min wins 6.7% — real and costly
held_rotation_like_trendpartialneeds intended_management_mode / planned_hold
oversized_yellowNOmissing position_size_vs_plan
no_clear_invalidationNOmissing pre_trade_invalidation_defined
good_executionderivednone of the above + followed checkpoints

Read: the tags DO explain the leak — primarily traded_red_gate / chased_hot_tape (−$356.72) and ignored_3min_failure. But 2 of 7 tags need fields that aren't logged yet.

Recommendation

REVISE, then promote in stages. Do not promote the gate as currently specified (the RVOL axis is wrong). After one confirmatory walk-forward pass, promote the proven parts; rework RVOL; defer the weak parts.

Promote (after walk-forward)

  • |gap| > 1.5% → RED (primary)
  • 3-minute fade-exit (hard rule)
  • 10–15 min expansion gates trend-hold
  • Instrument-type filter (megacap-tech eligible)
  • IBIT + CVX watch

Rework / defer

  • Rework the RVOL axis — demote to context; resolve the RVOL×instrument confound
  • Defer green/yellow gradation (weak)
  • Defer AMD/ETHU/RGTI blacklist (tiny n)
  • Defer oversized_yellow + no_clear_invalidation tags (need logged fields)

Exact next changes before promotion

Before Strategy Library promotion

  1. Run a walk-forward / out-of-sample pass — everything here is in-sample on ~2 months of one trader's trades. ✓ DONE 2026-06-09 → walk-forward / OOS validation: the post-entry fade-exit + gap-RED + avoid-bad-type hold their sign OOS; "megacap = profitable" and the entry-gate-alone FLIP negative in the test half (relative filters, not absolute edges). Test regime- and outlier-dominated → HOLD promotion; forward-test live.
  2. Replace the RVOL gate axis with an instrument-type axis + the |gap|>1.5% RED rule in vwap_pull_in_strategy_library_draft.json; keep validation_status: working_rule_needs_backtest on every threshold until the pass confirms.
  3. Resolve the RVOL×instrument confound: cross-tab RVOL bucket × instrument type to see whether RVOL adds anything once type is controlled. ✓ DONE 2026-06-09 → RVOL × instrument-type cross-tab: RVOL does not survive type control (in megacap-tech, <1.0 is the best bucket; the <1.0 loss was IBIT/crypto). Recommendation: remove RVOL from the gate (context-only). Use gap → instrument type → post-entry behavior.
  4. Reconcile the RVOL definition vs the original diagnostic (loser-mean 4.01) — they disagree; one query settles it.

Before Scorecard integration

  1. Prioritize the proven fields: gate_color (gap-driven), setup_classification, gap_pct, NEW instrument_type, first_3min_result, first_10_15min_confirmation.
  2. Ship the computable mistake tags now: traded_red_gate, chased_hot_tape, ignored_3min_failure.
  3. Add the missing logged fields before their tags can work: position_size_vs_plan (for oversized_yellow) and pre_trade_invalidation_defined (for no_clear_invalidation).
  4. Design-only — no gen_scorecard.py / _scientific_overlay.js edits until approved.

Before VCET / live-classifier integration

  1. Surface gap_pct + instrument_type now (computable live, reliable). Do not lead with RVOL — 82% populated and doesn't separate.
  2. Wire the live 1-min / VCET-v1.1 feed to compute first_3min_result / first_10_15min_confirmation at the ticket — the proven post-entry checkpoints can't fire live without it.
  3. Emit the gate as gap-RED / type-RED / clean rather than an RVOL-centric label.

Data gaps & sample-size warnings

Source & related

Read-only validation audit · computed from on-disk Liberator-validated artifacts, no DB query · no production Strategy Library / Scorecard / VCET / DB changes · all thresholds remain working rules pending an out-of-sample pass.