VCET Pull-In At-Entry Pilot — Phase 4a

Generated 2026-06-04 · In-memory VCET computation (no main.vcet_bars) · Engine: v1.0 at slope_trend_thresh=2.0 · Bars: kti five_min, 14-day lookback · Snap rule: bar_ts ≤ entry_utc AND > entry_utc − 5min (per original VCET brief).

⚠ Headline verdict

The original hypothesis was NOT supported on this pilot sample. Pull-In → Trend winners did not cluster in EARLY_EXPANSION / EXPANSION_CONFIRMED / TREND_UP at entry — they entered in ROTATION. The biggest single Pull-In loss in the sample (-$216 IBIT 6/3) entered in EXPANSION_CONFIRMED. The state label alone does not separate winners from losers at v1.0 / slope=2.0 on this sample.

However — VCET numeric fields (slope, band_width, vcet_score, compression_score) do show systematic differences between winners and losers within the same slug. The signal is in the geometry, not the categorical label. See §4.

Verdict on next step: RVOL + gap (the existing Quality Gate) remains the primary actionable filter. VCET v1.0 at this sample size is too noisy to displace it. See §7 for the recommendation.

1. Data availability

Snap rate by slug

SlugTotalWith VCETNO_BARSnap rate
vwap-pull-in leak127359227.6%
vwap-pull-in -> trend edge159660.0%
chase-extended sentinel56114519.6%
TOTAL1985514327.8%
PI → Trend has the highest snap rate (60%) because those trades cluster in late-May (2 in 06/01 week + 7 in 05/25 week = the bars-available window). Plain PI is spread across the full April-June range; chase-extended similarly.

2. Pull-In split by VCET state at entry

vwap-pull-in — the leak slug (n=35 OK)

State at entrynTotal P&LAvgWin %
ROTATION22-$91.77-$4.1722.7%
TREND_DOWN5-$38.87-$7.770.0%
TREND_UP4-$41.40-$10.350.0%
FAILED_EXPANSION_TENTATIVE2-$18.35-$9.180.0%
COMPRESSION1-$2.10-$2.100.0%
EXPANSION_CONFIRMED1-$216.15-$216.150.0%
Every state bucket lost money. The single EXPANSION_CONFIRMED row is the IBIT 6/3 -$216 outlier (see §6). TREND_UP and TREND_DOWN both 0% WR — plain PI in trending environments lost every time, contradicting the intuitive "enter Pull-In when trend is established" read.

vwap-pull-in → trend — the edge slug (n=9 OK)

State at entrynTotal P&LAvgWin %
ROTATION5+$109.49+$21.90100.0%
TREND_UP3+$10.69+$3.5666.7%
TREND_DOWN1+$12.25+$12.25100.0%

This is the surprise.

5 of 9 PI → Trend winners entered in ROTATION, not in EXPANSION_CONFIRMED as the hypothesis predicted. The biggest PI → Trend winners (+$68 DAL 5/26, +$22 DAL 5/27) both entered in ROTATION_FLAT_BANDS.

Interpretation: the conversion to trend happens after entry, not at entry. VCET-at-entry captures the moment before the trend fires — that moment is rotation. The state evolves to TREND_UP in the 5-15 bars following entry. The right question may not be "what state at entry" but "what state at exit" or "what state transition occurred between entry and exit."

chase-extended (n=11 OK)

State at entrynTotal P&LAvgWin %
ROTATION11+$96.15+$8.7454.5%
All 11 chase-extended trades with bars entered in ROTATION. Net positive at 55% WR. The "extended chase" label produced by gen_scorecard.py (entry >2 SD from VWAP avg) does NOT match VCET's view, which sees these entries as inside the rotation band. The two classifiers are looking at different things.

3. Winners vs losers across all 3 slugs by VCET state

State at entryWLWin %W P&LL P&L
ROTATION162143.2%+$504.74-$390.87
TREND_UP2528.6%+$22.45-$53.16
TREND_DOWN1516.7%+$12.25-$38.87
FAILED_EXPANSION_TENTATIVE020.0%+$0.00-$18.35
COMPRESSION010.0%+$0.00-$2.10
EXPANSION_CONFIRMED010.0%+$0.00-$216.15

Reading the cross-slug pattern

ROTATION is the only state with net-positive P&L (+$114 across 37 trades). Every other state lost money on this sample. TREND states (TREND_UP / TREND_DOWN) were 17-29% WR — the opposite of the hypothesis.

The hypothesis predicted EXPANSION_CONFIRMED + TREND states would house winners. They did not. The state machine appears to be calling regime AFTER the move has matured (and after the entry was already taken), making "VCET state at entry" a lagging indicator on this sample.

4. VCET numeric fields by slug × outcome

This is where the discriminating signal lives. Averages within each slug × outcome group.
SlugOutnvcet_scoreband_width_pctvwap_slope_degdist_from_vwap_sdexpansion_scorecompression_score
vwap-pull-inW560.41.18+3.95+0.6531.026.0
vwap-pull-inL2948.02.54-1.61+0.3336.013.1
vwap-pull-in -> trendW854.03.83-1.48+0.1526.97.5
vwap-pull-in -> trendL1100.09.60+70.07+1.0435.00.0
chase-extendedW659.31.59-3.26+0.3923.35.0
chase-extendedL555.68.00-5.10-0.1415.026.0

EDGE in the numerics — not in the state labels

Plain PI: winners enter on rising VWAP slope (+3.95°) and narrow bands (1.18); losers enter on flat/falling slope (-1.61°) and wider bands (2.54). Score 60 vs 48. The slope sign and width predict outcome more cleanly than the state label.

Chase-extended: winners have narrow bands (1.59), losers have very wide bands (8.00) — a 5× ratio. Wide bands = volatile environment = chase doesn't work.

PI → Trend: the lone loser (LUNR -$11) had extreme slope (+70°) and wide bands (9.60) and a perfect score (100) — a forced trend signal that didn't follow through. The 8 winners had calmer numerics (slope -1.48, bw 3.83).

5. Top reason codes by slug (n ≥ 1)

Most rows are STICKY_DEFAULT (state held from prior bar, no fresh trigger) or SESSION_INACTIVE (pre-market entries before 9:30 ET). This is the noise floor.
Slugreason_coden
vwap-pull-inSTICKY_DEFAULT16
SESSION_INACTIVE12
TREND_DOWN_SLOPE_STRONG2
ROTATION_FLAT_BANDS2
COMPRESSION_HOLD1
TENT_LOST_RANGE_LEVEL1
vwap-pull-in -> trendSTICKY_DEFAULT3
TREND_UP_SLOPE_STRONG2
ROTATION_FLAT_BANDS2
SESSION_INACTIVE2
chase-extendedSESSION_INACTIVE7
STICKY_DEFAULT3
ROTATION_FLAT_BANDS1

6. Top examples

Top 5 PI → Trend winners (with VCET tag at entry)

Entry timeSymbolP&LVCET statereason_codescore
2026-05-26 13:56DAL+$68.25ROTATIONROTATION_FLAT_BANDS90
2026-05-27 13:56DAL+$22.50ROTATIONSTICKY_DEFAULT52
2026-05-28 11:56BB+$18.40TREND_UPTREND_UP_SLOPE_STRONG88
2026-05-28 09:31JOBY+$12.45ROTATIONSESSION_INACTIVE0
2026-06-03 10:38HL+$12.25TREND_DOWNSTICKY_DEFAULT38

Top 5 plain Pull-In losers (with VCET tag at entry)

Entry timeSymbolP&LVCET statereason_codescore
2026-06-03 09:41IBIT-$216.15EXPANSION_CONFIRMEDRANGE_BREAK_SHORT94
2026-05-27 09:39AMD-$59.30ROTATIONSTICKY_DEFAULT0
2026-05-28 08:12META-$47.73ROTATIONSESSION_INACTIVE0
2026-06-03 09:51NVDA-$36.86ROTATIONSTICKY_DEFAULT0
2026-06-02 09:48RGTI-$33.58TREND_UPSTICKY_DEFAULT70

The IBIT 6/3 -$216 outlier

Entry: LONG IBIT at 09:41 ET, 6/3/2026. VCET state at entry: EXPANSION_CONFIRMED, reason RANGE_BREAK_SHORT, score 94. The state machine said "confirmed expansion, breaking the range to the downside" — correctly. The trade was taken LONG. The trade lost $216 because the short-side break followed through. VCET was right; the trade was wrong-side of what VCET said. One data point doesn't make a rule, but this single trade nearly accounts for the entire vwap-pull-in / EXPANSION_CONFIRMED loss bucket.

7. Verdict & rule candidates

Does VCET separate Pull-In converters from non-converters?

At v1.0 / slope=2.0 on this 55-trade pilot sample: NOT BY STATE LABEL. The hypothesis (PI → Trend winners in EARLY_EXPANSION / EXPANSION_CONFIRMED / TREND_UP, plain PI losers in ROTATION / FAILED_EXPANSION) did not hold. Winners and losers both cluster in ROTATION.

However, BY NUMERIC GEOMETRY: yes, partially. Plain PI winners have rising VWAP slope (+3.95° vs -1.61° for losers); chase-extended winners have narrow bands (1.59 vs 8.00 for losers); PI → Trend winners have moderate slope/bands, the lone loser was an extreme-slope false signal.

The Quality Gate (RVOL < 2.5 AND gap < 1.5%) likely captures most of what VCET geometry would. Until v1.1/CBO lands AND we have full historical bars, VCET v1.0 state at entry is not the right additional filter on top of the existing gate.

Practical rule candidates from this pilot (small-sample — suggestive)

RuleBasisConfidence
V1. Plain Pull-In: prefer entries with VCET vwap_slope_deg > 0 and vcet_score > 55winners avg +3.95°/60.4 vs losers -1.61°/48.0⚠ low — n=5 winners
V2. Plain Pull-In: avoid entries when state is TREND_UP or TREND_DOWN (paradoxical 0% WR)9 trades 0 wins; likely "trend already extended" condition⚠ low — n=9 but counter-intuitive
V3. Chase / Extended: require vwap_band_width_pct < 3.0winner bw 1.59 vs loser bw 8.00⚠ low-medium — n=11
V4. If VCET says EXPANSION_CONFIRMED with reason RANGE_BREAK_SHORT, do not take LONG Pull-In (counter-side break confirmed)IBIT 6/3 -$216 specifically⚠ very low — n=1
V5. Continue using RVOL+gap Quality Gate as primaryestablished earlier; VCET v1.0 doesn't displace it herehigh — from larger Pull-In diagnostic

8. Limitations

9. Next-step recommendation

What I'd do based on this pilot

  1. Keep the Quality Gate as primary. RVOL + gap is empirically actionable and explains the Pull-In split better than VCET v1.0 state at this sample size.
  2. Do not build full main.vcet_bars infrastructure yet. v1.0 alone hasn't shown the discriminating power that would justify the build. Wait for v1.1 (with CBO) AND better historical bar source before investing in the table + backfill.
  3. Add the VCET numerics (slope, band_width, score) as a secondary filter in the Quality Gate when bars happen to be available live (the regime card already has this). This is a small surface-area change to the existing pre-trade workflow, not a new database.
  4. Investigate state-trajectory not state-at-entry. If we want VCET to identify conversion-to-trend setups, we should be looking at state CHANGE between entry and entry+5/15/25 min. The current pilot has this data but the sample is too small to aggregate. A bigger sample is needed before this hypothesis can be tested.
  5. Sample-review the 5 PI → Trend / ROTATION winners (DAL 5/26, DAL 5/27, JOBY 5/28, plus 2 others). Do they share a chart pattern beyond the VCET label? This is the qualitative bridge to the next hypothesis.

10. Workflow links

Static report · 55-trade sample · in-memory VCET computation (no main.vcet_bars created) · No DB writes · No VCET engine modifications · Phase 4a complete · Phase 4b (broader sample / historical bars) deferred per recommendation.