ID: S030
Slug: trend-continuation-pullback-v1
Run date: 2026-07-01
Outcome: FAILED
Headline metric: portfolio max drawdown 43% (tail gate ≤ 25%) — unchanged from S029
Fail reason: The pullback entry improves per-trade expectancy over the S029 breakout (+0.36% vs +0.26%, nearly closing the gap to the eligible-leader baseline) — confirming the breakout timing was the leak — but the momentum-crash drawdown (43%) is unchanged, and CVaR-95 (−10.8%) still fails. The drawdown is a property of holding correlated leaders, not of the entry; no entry rule touches it.
What we tested
Everything frozen identical to S029 — the survivorship-free universe, the Trend-Template screen (stacked rising 50/150/200-DMAs, near-high, RS-rank ≥ 70), the exit (1.5×ATR hard stop + 50-DMA trail, gap-honest, data-gap guard), equal-weight, 30 bps — with one varied element: the entry. Buy a Trend-Template leader when it reclaims its 21-day EMA (close > EMA21 having closed ≤ EMA21 the prior day), at the next open. 34,948 trades (≈2× the breakout, reclaims are more frequent), avg hold 17 days.
What we found
| Gate (Stage 1) | Required | Observed | Verdict |
|---|---|---|---|
| Edge beats baseline | mean > 0 and > baseline | +0.36% vs baseline +0.43% | FAIL (but nearly closed) |
| Significance | p < 0.05, right sign | p = 0.62 | FAIL |
| Tail — CVaR-95 | ≥ −8% | −10.8% | FAIL |
| Tail — worst trade | ≥ −20% | −71% | FAIL |
| Tail — portfolio max drawdown | ≤ 25% | 43% | FAIL |
| Win rate / payoff (reported) | — | 26% / 3.3× | — |
Exit split flips vs S029: 66% now trail out on the 50-DMA (vs 32%), 29% stop, because a pullback entry sits closer to the trailing MA. The improvement is real and localised to the entry; the drawdown and the gap-through CVaR are untouched.
The entry was the leak — but not the wall. S029 said the breakout subtracted value; S030 confirms it — the pullback lifts the mean by ~40% and lands the edge essentially at the eligible-leader baseline. So the screen carries whatever drift there is, and a sensible entry neither adds nor destroys much beyond it. But the binding failures — the 43% momentum-crash drawdown and the −71% overnight gap-through — are exactly as deep as S029's, because they are properties of holding leaders, not of when you bought.
Why this matters
Two A1 entries tested; both cut on the same entry-independent wall. Combined with the B1 evidence (exposure overlays barely move a correlated-crash drawdown), this isolates the momentum-crash portfolio drawdown as A1's structural wall — the trend-family analogue of B1's inseparable tail. A better entry improves the trade; it does not make the strategy survivable.
What this doesn't tell us yet
- The baseline uses positional matching (a shared-helper limitation), so "edge ≈ baseline" is directional; the robust, binding failures are the tail co-gates (CVaR, drawdown), baseline-independent.
- It does not prove no A1 can pass — but it shows the wall is the drawdown, and the two obvious levers (better entry here; exposure control from the B1 diagnostics) don't clear it.
What happens next
No Stage 2. The evidence points at the drawdown, and the B1 diagnostics already showed exposure overlays (vol-target, market-regime) barely move a fast correlated-crash drawdown. The honest reading is that A1, like B1, is structurally blocked — the momentum-crash round-trip (the S004 lesson, now reconfirmed twice) is not fixable by entry or per-trade risk. A market-regime exposure overlay specific to A1 could be tried, but with tempered expectations; otherwise A1 closes and the budget moves to C1 (cross-asset RS rotation).
S030 is recorded failed; family A1's cumulative trial count stays at 2.
What we tested — the recipe
Slice & dice
For the specialist — methodology details (click to expand)
- Net: mean +0.36%/trade, median −2.3%, win 25.6%, payoff 3.30×, n = 34,948, avg hold 16.8 trading days. Matched baseline +0.43% (positional caveat); bootstrap p = 0.62.
- Tail: CVaR-95 −10.8%; worst −71% (overnight gap through the stop); portfolio max DD 43%.
- Exits: trail 22,999 (66%) · stop 10,136 · gap-through 1,358 · held-to-end 436 · gap-halt 19.
- Delta vs S029: only the entry (21-EMA reclaim vs 50-day-high/volume breakout); screen, exit, stop, sizing, friction, baseline, gate all identical.
- Artifacts:
trades.parquet,ab_summary.json,charts.json. Runner:lab/quickkill/trend-continuation-pullback-v1/run.py.