chase-extended Bucket Audit

Generated 2026-06-04 · Scope: every trade where canonical_strategy_slug = 'chase-extended' (n=58) · Method: Layer 1 evidence-only pull + Layer 2 provisional reclassification with source/proxy/heuristic-tagged rules · Constraints: Read-only on production. No DB / taxonomy / classifier / VCET / Scorecards changes.

❌ Bad conclusion

"Chase / Extended works."

✓ Better conclusion

"Chase / Extended is a mixed residual label hiding better mechanisms."

⚠ Headline answer to the key question

Is chase-extended profitable because A (true chases work), B (real setups mislabeled), C (artifacts), or D (insufficient data)?

Answer: B + C, supported by D. A is rejected.

The bucket's surface +$198.28 hides: 9 mislabeled real-setup trades carrying +$333.28 of edge, one outsized pre-market NVDA outlier (+$199.07), hold-time and time-of-day skew, and a residual mass of 37 trades that net modestly negative.

1. Bucket summary

58
trades
+$198.28
total P&L
33/25/0
W / L / S
56.9%
win rate

Without 9 misclassified trades, the bucket is net −$135

9 trades (15.5% of the bucket) generate +$333.28 from mechanisms that aren't chase: 7 trend_trade candidates (MTF ≥ 80 in trade direction, +$255.30) and 2 event_driven candidates (catalyst-driven, +$77.98, 2/2 WR). The remaining 49 trades net −$135.00. The chase setup itself is NOT a working edge.

2. Answers by option (A / B / C / D)

A. True chase trades working NO

Only 2 trades match the canonical chase signature (RVOL > 2.5 AND |gap| > 1.5%); they net −$4.32 on 1/2 WR. The chase mechanism itself is not profitable.

B. Real setups mislabeled YES — DOMINANT

9 trades = +$333.28 in trend_trade (n=7, +$255) + event_driven (n=2, +$78, 100% WR). Without these 9, bucket is −$135. The bucket's "profit" is mechanistically these 9 trades.

C. Artifacts (hold-time / time-of-day / crypto) YES — STRONG

>60 min holds carry +$261.22; sub-15min trades are 1/7 WR netting −$65. 10-12 ET is the only loser window (−$55). Crypto-adjacent ETHU/IBIT/MASK = −$116. NVDA pre-market 6/2 +$199.07 alone ≈ 100% of bucket P&L.

D. Insufficient data YES — SUPPLEMENTARY

37 of 58 (64%) trades cannot be classified to a non-chase mechanism from DB fields alone. They need intraday VWAP distance / slope / candle structure to confirm any alternative. Default chase-extended residual stands honestly.

3. True chase signature breakdown

Trades matching the Quality Gate canonical chase definition (RVOL > 2.5 AND |gap| > 1.5%):

DateSymbolSideRVOLGap %Net P&L
2026-04-13 09:39IBITLONG3.84+2.4%+$1.08
2026-06-04 10:08IBITLONG2.74+1.9%−$5.40
TOTAL — n=2−$4.32

Both are IBIT, both LONG. 1W / 1L. The canonical chase mechanism is unprofitable in this sample — consistent with the Quality Gate's earlier read of chase-extended as "behavioral chase tag, not an edge."

4. Misclassified trend_trade + event_driven (+$333.28)

Trend_trade candidates (n=7, +$255.30) trend_trade

DateSymbolSideMTFRVOLGap %P&LNote
2026-05-04 09:32CRWVLONG842.23+3.5%+$26.05Clean trend confluence
2026-05-04 09:35DXYZLONG1032.62+2.9%+$31.31Strong MTF
2026-05-07 09:41DDOGLONG907.23+32.3%−$1.90Earnings-blowoff territory; barely scratched
2026-05-11 09:33NVDALONG811.79+0.2%+$33.32
2026-05-29 09:25RGTILONG970.04−1.2%−$23.20RVOL near zero — MTF up but tape dead
2026-05-29 10:02MASKLONG1060.85+30.5%−$9.35Crypto-name w/ huge gap; thin tape
2026-06-02 05:58NVDALONG813.27+2.0%+$199.07Pre-market — dominant outlier

4 of 7 winners; +$199 comes from one pre-market NVDA; median per-trade is +$26.05. MTF ≥ 80 in direction is necessary but not sufficient — confidence stays MEDIUM (no intraday VWAP slope to confirm trend).

Event_driven candidates (n=2, +$77.98) event_driven HIGH

DateSymbolSideCatalystP&L
2026-05-05 09:20INTCLONGminor_news+$27.61
2026-05-06 09:32AAPLLONGminor_news+$50.37

Both pre-9:45 open entries on named catalysts. Only HIGH-confidence calls in the audit — the catalyst fields are directly in the DB so the rule fires without proxy gaps.

5. Layer 2 reclassification — full breakdown

Proposed mechanismnTotal P&LWin rateConfidence% of bucket P&L
trend_trade (candidate) trend7+$255.304/7 (57%)MEDIUM128.7%
event_driven (candidate) event2+$77.982/2 (100%)HIGH39.3%
gap_fade (candidate)7−$3.063/7 (43%)MEDIUM−1.5%
chase_extended (confirmed) true chase2−$4.321/2 (50%)MEDIUM−2.2%
level_breakout (candidate)2−$31.511/2 (50%)MEDIUM−15.9%
counter_trend_scalp (candidate)1−$22.430/1 (0%)LOW (heuristic)−11.3%
insufficient_data unclassifiable37−$73.6822/37 (59%)INSUFFICIENT−37.2%
TOTAL58+$198.2833/58 (57%)100.0%

Confidence distribution: HIGH n=2 · MEDIUM n=18 · LOW n=1 · INSUFFICIENT n=37.

6. The hold-time / time-of-day artifact (mechanism C)

The bucket's profitability lives in hold-time, NOT in setup mechanism

Trades held >60 min carry 132% of the bucket's net profit; sub-15min trades net −$65. The 10-12 ET entry window is the only consistent loser.

Hold-time distribution

Hold bucketnTotal P&LWin rateNote
Sub-5 min4−$45.900/4 (0%)Quick-exit losers (H2 heuristic)
5–15 min3−$18.971/3 (33%)
15–60 min8+$1.933/8 (38%)Near break-even
Over 60 min43+$261.2229/43 (67%)All the profit lives here

Time-of-entry distribution

WindownTotal P&LWin rate
Pre-market (< 09:00)9+$92.066/9 (67%)
Open 09:30–09:4527+$58.7515/27 (56%)
09:45–10:007+$55.535/7 (71%)
10:00–12:0010−$54.963/10 (30%)
12:00–14:003+$39.623/3 (100%)
After 14:002+$7.281/2 (50%)

Same Quality Gate finding as the plain Pull-In bucket: crypto-adjacent leveraged products do not behave like clean-tape setups, regardless of slug. ETHU (−$60), IBIT (−$47), MASK (−$9) = −$116 across 10 trades.

7. The NVDA 6/2 +$199 outlier

One trade ≈ 100% of bucket P&L

NVDA 2026-06-02 05:58 LONG, MTF=81, RVOL=3.27, gap=+2.0%, P&L = +$199.07.

This single pre-market entry accounts for 100.4% of the bucket's net +$198.28. Without it, the bucket nets −$0.79. The audit classifies it as a trend_trade candidate (strong MTF + matching direction), but its real mechanism (pre-market positioning on overnight strength? earnings catalyst? insider news? trend continuation?) is not determinable from the DB alone — and is mechanistically distinct from the "chase" framing the slug implies.

Implication: any aggregate read of the chase-extended bucket needs to be reported with and without NVDA 6/2 to be honest.

8. Insufficient_data caveat

64% of the bucket cannot be classified from the DB alone

37 of 58 trades get insufficient_data because no candidate rule fires from main.trade_context fields. They need:

Until these are in the DB, this audit is a partial picture. The Priority 5 chase-extended residual assignment honestly stands for these 37 trades, but their underlying mechanism is unknown.

Within the 37, top winners and losers:

22 winners / 15 losers within the 37 — modest positive raw rate, fat-tailed by the few large losers.

9. Recommendations (no implementation yet)

  1. Do NOT promote chase-extended to a "working" or recommended setup. The bucket-level +$198 is artifact + mislabeling; the canonical-chase subset (n=2) is net negative.
  2. 7 trend_trade and 2 event_driven candidates deserve a closer look when classifier refinement happens (per VWAP_PULL_IN_CLASSIFIER_REFINEMENT_PLAN §8 Option A). Specifically, gen_scorecard.py Priority 5 should NOT absorb trades that meet a strong P5_catalyst or P6_trend_confluence rule.
  3. Add intraday VWAP distance to main.trade_context before any further reclassification work. Without it, 64% of this bucket is unclassifiable — and that's the structural blocker behind every MEDIUM ceiling in this audit.
  4. Mid-morning (10–12 ET) "chase" entries should be reviewed at the trader-behavior level, not the classifier level. Time-of-day skew this strong likely reflects an attention/tilt window rather than a setup category.
  5. Keep the canonical slug as-is. Like vwap-pull-in -> trend, chase-extended is an outcome-based residual label. Renaming or splitting it without VWAP geometry would be premature.

10. Honest caveats

11. What did NOT change

Phase board unchanged.

Links — audit deliverables

Links — related reports