# VWAP Pull-In Split — Diagnostic Report

**Date:** 2026-06-03
**Author:** Claude Code (read-only diagnostic)
**Scope:** Compare `vwap-pull-in -> trend` (15 trades, +$367.67) vs `vwap-pull-in` (127 trades, -$364.47) to identify rule candidates that reduce losing plain Pull-Ins and improve Pull-In→Trend selection.
**Data source:** `main.trade_context` on `trsports_tradingjournal_dev`, all canonical_strategy_slug-resolved trades through 2026-06-03.

---

## 1. Headline split

| Metric | `vwap-pull-in` (plain) | `vwap-pull-in -> trend` |
|---|---:|---:|
| Trade count | 127 | 15 |
| Total net P&L | **-$364.47** | **+$367.67** |
| Avg / trade | -$2.87 | +$24.51 |
| Median / trade | -$2.25 | +$12.25 |
| Win rate | 35.4% | **93.3%** |
| Winners / Losers / Scratches | 45 / 80 / 2 | 14 / 1 / 0 |
| Min trade | -$216.15 | -$11.76 |
| Max trade | +$150.76 | +$136.29 |

**Net effect:** Plain Pull-In erases everything Pull-In→Trend earns and then some. Total `vwap_pullback_continuation` family P&L is +$3.20 across 142 trades (essentially flat) — the conversion-to-trend is the entire edge.

---

## 2. Symbol breakdown

### Pull-In → Trend (all 12 winning symbols, 1 losing)

```
AAOI:   1 trade,  +$136.29 (single-trade homerun)
DAL:    3 trades, +$96.34   (consistent: 3 winners)
CAR:    1 trade,  +$53.76   (short)
WULF:   1 trade,  +$18.48
BB:     1 trade,  +$18.40
JOBY:   2 trades, +$13.15
HL:     1 trade,  +$12.25
CRWV:   1 trade,  +$9.21
RGTI:   1 trade,  +$8.92
SPY:    1 trade,  +$8.58
INTC:   1 trade,  +$4.05
LUNR:   1 trade,  -$11.76   (only loser)
```

**Diverse symbol distribution.** No single symbol carries the slug — 12 different tickers contribute winners.

### Plain Pull-In — top 10 most-traded symbols

| symbol | n | total | avg | win % |
|---|---:|---:|---:|---:|
| QQQ | 11 | -$9.55 | -$0.87 | 45.5% |
| **IBIT** | **10** | **-$185.51** | **-$18.55** | **30.0%** ⚠ |
| AMZN | 9 | +$43.93 | +$4.88 | 44.4% |
| META | 9 | -$19.13 | -$2.13 | 55.6% |
| NVDA | 7 | +$19.49 | +$2.78 | 42.9% |
| TSLA | 6 | +$43.39 | +$7.23 | 66.7% |
| GOOGL | 5 | +$32.67 | +$6.53 | 60.0% |
| **CVX** | **5** | **-$44.63** | **-$8.93** | **20.0%** ⚠ |
| MSFT | 4 | +$133.54 | +$33.39 | **100%** ✓ |
| CRWV | 4 | -$6.29 | -$1.57 | 75.0% |

### Plain Pull-In — biggest single-symbol losses

| symbol | n | total | win % |
|---|---:|---:|---:|
| **IBIT** | 10 | **-$185.51** | 30% |
| **AMD** | 3 | -$69.65 | **0%** |
| **ETHU** | 3 | -$69.35 | **0%** |
| CVX | 5 | -$44.63 | 20% |
| RGTI | 2 | -$35.68 | — |
| KOPN | 1 | -$28.80 | — |
| MXL | 1 | -$28.30 | — |
| RKLB | 1 | -$28.24 | — |
| WULF | 1 | -$23.30 | — |

**The IBIT story:** 10 plain Pull-In attempts across 10 different days, net -$185, 30% win rate. **Single biggest leak in the dataset.** AMD and ETHU are 0% win rate on small samples — pattern of "Pull-In on leveraged/levered-product names doesn't work."

---

## 3. Time-of-day breakdown

| Hour ET | plain PI n | plain PI total | plain PI avg | plain PI win % | PI-Trend n | PI-Trend total | PI-Trend avg |
|---:|---:|---:|---:|---:|---:|---:|---:|
| 04 | 1 | -$13.92 | -$13.92 | 0% | — | — | — |
| 07 | 2 | -$58.93 | -$29.47 | 50% | — | — | — |
| 08 | 6 | -$2.37 | -$0.40 | 50% | — | — | — |
| **09** | **62** | **-$209.92** | **-$3.39** | **37.1%** | 4 | +$19.87 | +$4.97 |
| **10** | **29** | **-$191.22** | **-$6.59** | **24.1%** ⚠ | **7** | **+$233.06** | **+$33.29** (100% wins) |
| 11 | 8 | +$2.02 | +$0.25 | 50% | 1 | +$18.40 | +$18.40 |
| 12 | 8 | +$58.91 | +$7.36 | 12.5% | — | — | — |
| 13 | 4 | +$39.90 | +$9.98 | **100%** | 3 | +$96.34 | +$32.11 |
| 14 | 5 | -$2.02 | -$0.40 | 20% | — | — | — |
| 15 | 1 | +$15.53 | +$15.53 | 100% | — | — | — |
| 16 | 1 | -$2.45 | -$2.45 | 0% | — | — | — |

**Headline:** Plain Pull-In is **catastrophically concentrated in 09:00–10:00 ET** (91 of 127 trades = 72%, total -$401, ~30% win rate). Pull-In→Trend in the same window prints +$253 across 11 trades at 100% win in the 10:00 hour.

Same hour, different outcomes. The hour isn't the problem. **Selection within the hour** is.

---

## 4. Weekly breakdown

| Week | plain PI n | plain PI total | PI-Trend n | PI-Trend total |
|---|---:|---:|---:|---:|
| 2026-04-06 | 2 | +$27.15 | 1 | +$136.29 |
| 2026-04-13 | 11 | -$19.81 | 4 | +$90.37 |
| 2026-04-20 | 16 | -$58.07 | — | — |
| 2026-04-27 | 13 | +$157.46 | 1 | +$8.58 |
| 2026-05-04 | 21 | +$9.35 | — | — |
| 2026-05-11 | 24 | -$32.35 | — | — |
| 2026-05-18 | 5 | -$39.56 | — | — |
| **2026-05-25** | **21** | **-$159.78** ⚠ | 7 | +$116.13 |
| **2026-06-01** | **14** | **-$248.86** ⚠ | 2 | +$16.30 |

**Pattern:** The recent 2 weeks are the worst for plain PI (-$408 combined). The same weeks produce $132 from PI-Trend. **In the same week, the strategy works on some setups and not others — selection is the discriminator.**

---

## 5. Top 10 best/worst trades

### Plain Pull-In — best 10

```
2026-04-21  IBIT  LONG  +$150.76
2026-05-07  MSFT  LONG  +$83.95
2026-04-30  AMZN  LONG  +$83.60
2026-04-30  IBIT  LONG  +$60.53
2026-06-03  TSLA  LONG  +$50.42
2026-05-14  MSFT  LONG  +$48.36
2026-04-20  MRVL  LONG  +$37.55
2026-05-12  NVDA  LONG  +$35.87
2026-05-28  GOOGL LONG  +$32.73
2026-04-10  NVDA  LONG  +$27.15
```

### Plain Pull-In — worst 10

```
2026-06-03  IBIT  LONG  -$216.15  ⚠ outlier
2026-04-24  IBIT  LONG  -$65.50
2026-05-27  AMD   LONG  -$59.30
2026-05-19  IBIT  LONG  -$51.43
2026-05-12  ETHU  LONG  -$49.30
2026-05-28  META  LONG  -$47.73
2026-04-22  IBIT  LONG  -$42.18
2026-06-03  NVDA  LONG  -$36.86
2026-06-02  RGTI  LONG  -$33.58
2026-05-13  KOPN  LONG  -$28.80
```

**4 of the 10 worst are IBIT.** The biggest single trade in the entire diagnostic (-$216 on 6/3) is IBIT plain Pull-In.

### Pull-In → Trend — best 10

```
2026-04-10  AAOI  LONG  +$136.29
2026-05-26  DAL   LONG  +$68.25
2026-04-16  CAR   SHORT +$53.76
2026-05-27  DAL   LONG  +$22.50
2026-04-13  WULF  LONG  +$18.48
2026-05-28  BB    LONG  +$18.40
2026-05-28  JOBY  LONG  +$12.45
2026-06-03  HL    LONG  +$12.25
2026-04-13  CRWV  LONG  +$9.21
2026-04-15  RGTI  LONG  +$8.92
```

### Pull-In → Trend — worst (only 1 negative trade)

```
2026-05-26  LUNR  LONG  -$11.76   (only loser)
```

(All other PI→Trend trades are wins — the "worst 10" trails into wins below $10.)

---

## 6. Context comparison — the real signal

### ATR + RVOL + Gap (ALL trades per slug)

| | plain PI | PI → Trend |
|---|---:|---:|
| Avg ATR from 20d SMA | 1.59 | 2.07 |
| Avg ATR from prior close | **+0.25** | **-0.25** ⚠ |
| Avg RVOL pre-entry | **3.08** | **1.18** ⚠ |
| Avg gap % | **+1.45%** | **+0.30%** |

### Same fields split by winners vs losers (the actionable view)

| | plain PI losers (n=80) | plain PI winners (n=45) | PI→Trend winners (n=14) |
|---|---:|---:|---:|
| ATR from prior | **+0.37** | +0.08 | **-0.38** |
| **RVOL** | **4.01** | **1.38** | **1.14** |
| **Gap %** | **+2.11%** | +0.35% | **-0.88%** |

**This is the signal.** Plain PI losers come in with:
- **RVOL 4x normal** (4.01 vs ~1 baseline)
- **2%+ gap up** (means stock is "in play" with overnight news/hype)
- **ATR-extended already** (+0.37 ATR above prior close at entry)

Plain PI winners and Pull-In→Trend winners are in **calm tape**: RVOL ~1.1–1.4, near-flat gap, near-flat from prior close.

### MTF context

| | plain PI | PI → Trend |
|---|---:|---:|
| n with MTF data | 127 | 15 |
| Grand total avg | 19.23 | **43.60** (2.3× stronger) |
| Position total | 12.14 | 24.13 |
| Swing total | 10.13 | 21.80 |
| Intraday total | 4.50 | 13.07 |

PI→Trend setups happen in **2× stronger MTF environments** across every timeframe.

But **MTF doesn't separate winners from losers within plain PI** — losers have MTF grand avg 21.74, winners 14.64 (basically equal). MTF is a useful macro filter but not a per-trade filter.

### Stock vs MAs

| | plain PI | PI → Trend |
|---|---:|---:|
| Above 5d MA | 71.7% | 93.3% |
| Above 20d MA | 76.4% | 86.7% |
| Above 50d MA | 78.0% | 93.3% |
| Above 200d MA | 65.4% | 73.3% |

PI→Trend stocks are in stronger longer-term uptrends. Plain PI winners-vs-losers MA stack basically equal (similar to MTF — useful macro filter, weak per-trade filter).

### Market alignment (QQQ/SPY)

Both slugs: **100% of trades happen with QQQ above 20d AND SPY above 20d.** Your trading universe is already filtered to bull market days. Index trend is not the differentiator.

### TICKQ at entry

| | plain PI | PI → Trend |
|---|---:|---:|
| n populated | 95 | 10 |
| Avg TICKQ | 49 | **450** |
| Min | -2198 | -820 |
| Max | 1917 | 1906 |

PI→Trend entries happen at **9× higher TICKQ average** — broader-market breadth is buying alongside.

### Catalyst presence

| | plain PI | PI → Trend |
|---|---:|---:|
| Total | 127 | 15 |
| With catalyst (non-none) | 9 (7%) | **0 (0%)** |
| No catalyst | 118 (93%) | 15 (100%) |

PI→Trend is **never catalyst-driven** in this dataset. Plain PI is mostly not, but the small 7% with catalysts is worth a separate look.

### Direction (side)

| | LONG | SHORT |
|---|---:|---:|
| plain PI | 114 | 13 |
| PI → Trend | 14 | 1 |

Long bias overwhelming on both. PI→Trend short side has insufficient sample to evaluate.

### Concentration check (overtrading per symbol-day)

| | total trades | distinct symbol-days | avg per sym-day | max per sym-day |
|---|---:|---:|---:|---:|
| plain PI | 127 | 127 | 1.0 | **1** |
| PI → Trend | 15 | 15 | 1.0 | **1** |

**Not overtrading at the symbol-day level.** Every Pull-In entry is one-per-symbol-per-day. The concentration problem isn't "10 NVDA Pull-Ins today" — it's "1 IBIT Pull-In on each of 10 different days, all losing."

### Fields NOT in `trade_context` (would need VCET integration)

These would be the next-level diagnostics but require VCET-at-entry data that doesn't exist on these rows yet:
- VWAP band at entry (LB1, LB2, etc.)
- Distance from VWAP in SD at entry
- Whether trade reached +2 SD post-entry (the trend-confirmation trigger)
- VCET state at entry (ROTATION / EARLY_EXPANSION / EXPANSION_CONFIRMED / TREND_UP)
- MFE / MAE per trade

The VCET-at-entry validation work (Phase 4) is what populates these. Until then, **RVOL and gap are the strongest signal** in available data.

---

## 7. Loss concentration verdict

**Where the leak lives:**

| Concentration axis | Concentrated? | Evidence |
|---|---|---|
| Symbol | **YES** — IBIT alone is half the total loss | IBIT -$185 of -$364 total |
| Time of day | **YES** — 09–10 ET window is -$401 | 91 of 127 trades, ~30% win rate |
| Week | **PARTIAL** — last 2 weeks are -$408 combined | But recent weeks aren't structurally different from earlier weeks; it's RVOL-elevated trades clustering |
| RVOL bucket | **YES** — high RVOL is losing | Losers avg 4.01 RVOL, winners 1.38 |
| Gap % | **YES** — high gap is losing | Losers avg +2.11%, winners +0.35% |
| Per-symbol-per-day count | **NO** | All 127 are 1-per-symbol-per-day |
| MA stack | **NO** | Winners and losers have similar MA alignment |
| Catalyst flag | **NO** | Only 7% catalyst-driven; can't differentiate |
| Index alignment | **NO** | All 100% above QQQ/SPY 20d |

---

## 8. Rule candidates — actionable

### To reduce losing plain Pull-Ins

**Rule A (strongest single filter — RVOL gate):**

> **Skip plain Pull-In setups when RVOL pre-entry > 2.5.**

Estimated effect on current dataset: would eliminate trades whose loser-avg RVOL was 4.01 (the bulk of losers). Plain PI winners had RVOL 1.38, so this filter preserves the wins.

**Rule B (companion filter — gap gate):**

> **Skip plain Pull-In setups when gap % > 1.5%.**

Plain PI losers gapped +2.11% avg, winners +0.35%. PI→Trend winners actually gapped **negative** (-0.88%) — the calm/down-open tape is where this strategy works.

**Rule C (combined — recommended live filter):**

> **Skip plain Pull-In if (RVOL > 2 AND gap > 1%) OR (RVOL > 3) OR (gap > 2%).**

Layered with rule B (gap > 1.5% alone): this is your "hyped tape" filter. The setup loses money in hyped tape, period.

**Rule D (symbol blacklist):**

> **Skip plain Pull-In on: IBIT, AMD, ETHU, RGTI.**

These four symbols carry the bulk of the losses on small-to-medium samples. Either (a) blacklist for plain Pull-In, or (b) require a higher quality bar (lower RVOL threshold, e.g. RVOL < 1.5) on these names specifically.

**Rule E (time-of-day refinement):**

> **Within the 09:00 ET hour, demand RVOL < 2 AND gap < 1%. Skip otherwise.**

The 09 ET window holds 62 plain PI trades at 37% win rate. Apply Rule C strictly inside that hour.

### To increase Pull-In → Trend selection

**Rule F (hold for conversion):**

> **Pull-Ins entered in calm tape (RVOL ~1.0–1.5, gap < 0.5%, MTF grand > 30) should be held for the trend-conversion trigger (+2 SD break) rather than exited at the +2 SD target.**

The PI→Trend slug captures the management decision when you let a Pull-In ride through the +2 SD breakout. Today this conversion happens only 15 times in 142 PI-family trades. Per the data, when conditions are calm and MTF strong, hold longer.

**Rule G (universe selection):**

> **Pre-market screening for PI→Trend candidates: TICKQ > 200 by 10:00 ET, gap between -1% and +1%, RVOL between 0.8 and 1.5, stock above 5d/20d/50d MAs.**

These are the empirical signatures of the 14 PI→Trend winners (n=15 total, so high signal-to-noise).

### What you cannot rule on yet (Phase 4 / VCET work)

- **+2 SD reach** as a hard precondition (not stored)
- **VCET state at entry** (EXPANSION_CONFIRMED vs ROTATION) — strongest theoretical filter, requires Phase 4
- **MFE/MAE per trade** — would let you size by expected favorable excursion

---

## 9. Estimated impact if Rules A+B+D applied retroactively

This is a back-of-envelope on the 127 plain PI trades (rough — actual would need re-querying):

| Filter applied | Trades excluded | Est. P&L recovered | New total |
|---|---:|---:|---:|
| **Rule A only** (RVOL > 2.5) | ~50–60 | ~+$300 | ~-$65 |
| **Rule A + B** (RVOL > 2.5 OR gap > 1.5%) | ~70 | ~+$400 | ~+$35 |
| **Rule A + B + D** (+ blacklist IBIT/AMD/ETHU/RGTI) | ~85 | ~+$500 | ~+$135 |

Combined with the existing PI→Trend +$367, total `vwap_pullback_continuation` family could go from +$3 to **~+$500** with no new setups, just rejection.

These are estimates from grouped averages — exact backtest would need per-row re-application.

---

## 10. What I'm NOT doing in this diagnostic

- Not modifying DB
- Not refreshing Scorecards HTML
- Not touching VCET core code
- Not running VCET backfill
- Not changing taxonomy
- Not building rule enforcement (this is recommendations only — the rules need your sign-off + implementation in `gen_scorecard.py`'s classifier or in a separate pre-entry filter)

---

## Stop conditions

This diagnostic stops here. Suggested next steps when ready:
1. Sample-review 10 of the IBIT plain Pull-In losers and 5 of the PI→Trend winners — does the data narrative match your read of those trades?
2. Decide which of Rules A–G to enforce (in classifier code, in your morning watchlist, or as a hard pre-trade check)
3. After 2 weeks of Rule-A enforcement, re-run this diagnostic to measure delta
