Pull-In → Trend ROTATION Winners — Trajectory Review

Generated 2026-06-04 · Source: VCET_PULL_IN_AT_ENTRY_PILOT §2 + fresh entry → +25 min trajectory data · Method: In-memory VCET v1.0 at slope_trend_thresh=2.0, kti 14-day window, snap at −5, 0, +5, +10, +15, +25 min around each entry · Scope: The 5 vwap-pull-in -> trend trades that entered in ROTATION (per the pilot).

⚠ Headline finding

Only 1 of 5 is a true ROTATION → TREND conversion. The other 4 split into:

1. Quick scoreboard

#TradeP&LConversion?What it really was
1DAL 5/26 13:56+$68.25❌ NOTD VWAP rotation scalp (stayed ROTATION, slope ≈ 0, compression building)
2DAL 5/27 13:56+$22.50❌ NOCounter-trend buy on big gap-down (−5.4%); stayed ROTATION; slope negative
3JOBY 5/28 09:31+$12.45✅ YESTrue ROTATION → TREND_UP at +15min; bw 1.54 → 6.42, slope +1.6° → +49°
4DAL 5/28 13:56+$5.59❌ NOTight compression, score climbing (96 → 100), modest +$0.08 drift
5JOBY 5/29 09:31+$0.70❌ NO (wrong side!)LONG into TREND_DOWN crash; slope −3.9° → −66.9°; barely escaped scratch

So the "5 ROTATION winners" framing from the pilot was misleading. The real read: 1 confirmed conversion, 3 mislabeled rotation scalps, 1 lucky escape.

2. Per-trade trajectories

1. DAL 2026-05-26 13:56 LONG +$68.25 — rotation scalp (mislabeled) rotation scalp

RVOL 0.73 (calm) · Gap −1.3% · ATR −0.39 · MTF 83 (strong) · no catalyst

tstatereasonscoreslope°bw%expcompclose
−5ROTATIONFLAT_BANDS60−0.391.490078.32
0ROTATIONFLAT_BANDS90−0.351.490078.42
+5ROTATIONFLAT_BANDS91−0.351.4915078.54
+10ROTATIONFLAT_BANDS93−0.341.490078.53
+15ROTATIONFLAT_BANDS94−0.291.480078.50
+25ROTATIONFLAT_BANDS96−0.201.4803078.52

Verdict: Stayed in ROTATION the whole 25 min. Slope ≈ 0 throughout. Bands barely moved. Score CLIMBED 60 → 96. Price drifted +$0.10/share. Textbook TD VWAP rotation scalp profile. The slug "Pull-In → Trend" is wrong for this trade — it never converted.

2. DAL 2026-05-27 13:56 LONG +$22.50 — counter-trend gap-down buy (mislabeled) rotation scalp

RVOL 1.31 · Gap −5.4% (big gap down) · ATR −1.6 (very oversold) · MTF 82 · no catalyst

tstatereasonscoreslope°bw%expclose
−5ROTATIONSTICKY_DEFAULT53−1.152.823581.73
0ROTATIONSTICKY_DEFAULT52−1.342.861581.71
+5ROTATIONSTICKY_DEFAULT50−1.632.923581.49
+10ROTATIONSTICKY_DEFAULT48−1.762.971581.46
+15ROTATIONSTICKY_DEFAULT47−1.923.011581.50
+25ROTATIONSTICKY_DEFAULT45−1.403.05081.56

Verdict: Slope NEGATIVE entire window. Bands EXPANDING slightly. Score DECREASING. Price moved AGAINST the LONG entry (−$0.15). Trade was profitable despite unfavorable trajectory — likely small share count, tight management, or scratched within minutes. Not a conversion.

3. JOBY 2026-05-28 09:31 LONG +$12.45 — TRUE CONVERSION ✅ true conversion

RVOL 1.59 · Gap +0.74% (calm open) · ATR +0.11 (flat) · MTF 62 · no catalyst

tstatereasonscoreslope°bw%expclose
−5ROTATIONSESSION_INACTIVE41+1.611.542011.43
0ROTATIONSESSION_INACTIVE0+17.992.987511.59
+5ROTATIONSESSION_INACTIVE0+31.375.037511.76
+10ROTATIONSESSION_INACTIVE0+42.926.167511.87
+15TREND_UPSLOPE_STRONG100+49.066.423511.83
+25TREND_UPSLOPE_STRONG100+44.386.722011.98

Verdict: THE conversion pattern. Entered with bands narrow (1.54) and slope barely positive (+1.6). Within 5 min: bands DOUBLED, slope EXPLODED, expansion_score jumped 20 → 75. By +15 min: state TREND_UP with score 100. Price moved 11.43 → 11.98 (+4.8%). The compression-before-expansion signature is unmistakable — but NOT visible at entry. Conversion happened in the first 5 minutes AFTER entry.

4. DAL 2026-05-28 13:56 LONG +$5.59 — compression scalp (mislabeled) rotation scalp

RVOL 0.94 (calm) · Gap −8.1% (huge gap down) · ATR −2.5 (very oversold) · MTF 74 · no catalyst

tstatereasonscoreslope°bw%expcompclose
−5ROTATIONFLAT_BANDS96+0.212.84153081.97
0ROTATIONFLAT_BANDS96+0.272.82353081.93
+5ROTATIONFLAT_BANDS97+0.202.8103081.89
+10ROTATIONFLAT_BANDS98+0.202.7903081.95
+15ROTATIONFLAT_BANDS99+0.172.7803081.88
+25ROTATIONFLAT_BANDS100+0.112.7503082.01

Verdict: Slope very close to zero. Bands CONTRACTING. Compression_score = 30 the entire time — true compression. Score climbed 96 → 100. Price drift +$0.08. Compression scalp — tight, low-vol, modest gain. Compression NEVER resolved into expansion within 25 min. Not a Pull-In → Trend conversion.

5. JOBY 2026-05-29 09:31 LONG +$0.70 — wrong-side, barely escaped wrong-side

RVOL 1.76 · Gap −5.6% (big gap down) · ATR −0.88 · MTF 22 (weak) · no catalyst

tstatereasonscoreslope°bw%expclose
−5ROTATIONSESSION_INACTIVE0−3.942.057512.12
0ROTATIONSESSION_INACTIVE0−36.595.547511.92
+5TREND_DOWNSLOPE_STRONG100−52.747.027511.73
+10TREND_DOWNSLOPE_STRONG100−62.227.593511.74
+15TREND_DOWNSTICKY_DEFAULT70−66.917.513511.77
+25TREND_DOWNSTICKY_DEFAULT64−37.097.242011.78

Verdict: Slope already −3.94° at −5min. Within 5 min after entry: slope −52.7°, bands wide open (7.02), state TREND_DOWN. VCET correctly identified the down-direction trend within 5 minutes. The LONG trade was wrong-side. Price fell 11.92 → 11.73 (−1.6%) in first 5 minutes after entry. The +$0.70 P&L was either a tiny share count, scratched immediately, or saved by late drift back. Not a successful conversion — a tape lucky to break even.

3. Answers to the four questions

Q1. Was ROTATION actually constructive compression before expansion?

Mostly NO. Only 1 of 5 (JOBY 5/28) showed compression-before-expansion: narrow bands at entry (1.54) → rapid expansion (1.54 → 6.42 in 15 min) → state transition ROTATION → TREND_UP at +15min.

The other 4: 3 DAL trades had bands stay flat, slope stay ≈ 0, no expansion ever fired (genuine rotation scalps, classifier mislabeled them). 1 JOBY 5/29 had bands expand DOWNWARD into TREND_DOWN — wrong direction for a LONG.

Q2. Is state-at-entry less useful than state trajectory?

YES — definitively. JOBY 5/28 is the proof: state at entry was ROTATION (SESSION_INACTIVE); state at +15 min was TREND_UP (SLOPE_STRONG). A state-at-entry filter would have flagged it as "rotation, pass" or "no fresh signal." The conversion was visible only in the trajectory — band width DOUBLING in 5 min, slope going from +1.6° to +17.99° to +31.37°, expansion_score jumping 20 → 75 in 5 min.

The numeric DERIVATIVES (rate of change of slope, rate of band expansion) distinguish a real conversion from a rotation scalp. State labels are too lagging.

Q3. Should the next VCET test be entry → +25 min trajectory instead of entry-only?

YES. Track: band-width delta at +5min, slope acceleration entry-5 → entry+5, expansion_score climbing past 50 within 10 min, state transition into EARLY_EXPANSION / EXPANSION_CONFIRMED within 15 min.

Caveat: the gen_scorecard.py classifier's labeling of "Pull-In → Trend" appears inconsistent with this signature. Of 5 trades it labeled PI→Trend ROTATION winners, only 1 actually showed the conversion. The classifier may be tagging trades on outcome, not on the conversion mechanism — a taxonomy refinement question for later.

Q4. Should VCET numerics be added to the Quality Gate as secondary context?

YES with specific thresholds from this review:

NumericPre-trade preferPre-trade avoid
vwap_band_width_pct at entry< 2.0 (narrow = compression possible)> 6.0 (already-expanded = past the move)
vwap_slope_deg at entryWatch sign — POSITIVE for LONGStrongly NEGATIVE for LONG
vcet_score at entry> 55< 30 (often SESSION_INACTIVE / low-conviction)
range_compression_score at entry> 25 (compression building)< 5 (no compression structure)

Secondary filter, not replacement. RVOL + gap caught the Pull-In split cleanly on the full 127-trade sample. VCET numerics add another layer of CONTEXT but don't displace the primary filter.

4. The bigger insight

The classifier rule that produces vwap-pull-in -> trend is unreliable

From this 5-trade sample, only 1 (JOBY 5/28) was a true conversion. The other 4 were mislabeled — they didn't trade through a state transition; they were rotation/compression scalps that profited.

Implication: the +$367 cumulative P&L credited to vwap-pull-in -> trend is real, but the slug name overstates the conversion mechanism. Some of those wins were rotation scalps that happened to exit at +2 SD.

This is NOT a taxonomy problem — the canonical slug system is fine. It's a classifier accuracy problem in gen_scorecard.py. The fix would be to re-define Priority 3 to require a state transition (VCET data at entry → exit), not just a wide exit. Phase 5+ scope.

5. Recommendations

  1. Update the Pull-In Quality Gate cheatsheet to add an optional VCET-numeric secondary filter when bars are live.
  2. Re-frame Phase 4b: test ENTRY → +5 / +15 MIN TRAJECTORY, not state-at-entry. The conversion signature is in the first 5-15 minutes after entry.
  3. Flag the classifier accuracy concern for a separate audit (don't fix now). Priority 3 should be tightened to require a state transition. Discussion for after Phase 4b confirms the trajectory pattern.
  4. Sample-review the remaining 9 historical PI → Trend trades without VCET bars in the pilot (AAOI +$136, CAR +$53, WULF +$18, BB +$18, CRWV +$9, RGTI +$9).
  5. Do NOT build main.vcet_bars yet — strengthens the pilot's verdict that v1.0 entry-state isn't the right layer. Trajectory analysis with v1.1 + CBO would be the proper next investment.

6. What I did not change

Phase board unchanged.

Links