claude
de92b4283d
Six confirm/refute families (vol-normalized momentum, mean-reversion z-fade, vol-regime-gated breakout, H4 EMA trend, H4 breakout port, channel-position fade) authored by engine-blind agents from a distilled reference card plus the CLI validation loop, adversarially design-reviewed, then run over 4 instruments x 3 stop regimes. Five honest nulls; one candidate (vol-normalized momentum on US500, WF-OOS P<=0.05 in all three regimes) forwarded to the Arc 4 breadth test.
89 lines
5.0 KiB
Markdown
89 lines
5.0 KiB
Markdown
# Arc 3 — six new data-only signal families, designed engine-blind
|
||
|
||
**Date:** 2026-07-13
|
||
**Engine:** aura @ `84e1075` (release build)
|
||
**Blueprints:** `blueprints/{mom-vnorm-h1,meanrev-z-h1,bo-h1-volregime,trend-h4-cross,bo-h4-trend,channelpos-fade-h1}.json`
|
||
**Campaigns:** `research/campaign-arc3-*.json`
|
||
|
||
## Method
|
||
|
||
Two experiments in one arc.
|
||
|
||
**The research question:** which signal families expressible in the std
|
||
vocabulary alone carry directional information on H1/H4 bars? Six families,
|
||
each a confirm/refute hypothesis (a base signal or a conditioner-over-base),
|
||
were run over the Arc 1 four-instrument matrix × three stop regimes
|
||
(vol{3,k2} = Arc 1, vol{60,k6}, vol{60,k12} = the Arc 2c ladder) with
|
||
vol-proportional slippage, through the Arc 1 pipeline (deflated screen →
|
||
90/30 WF → pooled-OOS bootstrap → generalize).
|
||
|
||
**The meta-experiment:** every blueprint was designed by an agent that had
|
||
*no access to engine sources* — only a distilled authoring reference card
|
||
(node vocabulary with port tables, op-script grammar, document schemas, unit
|
||
semantics, design rules) plus the CLI validation loop (`graph build`,
|
||
`introspect --params`, three-tier `campaign validate`). Each design passed an
|
||
independent adversarial design review (look-ahead, unit-consistency, wiring,
|
||
grid sanity, hypothesis shape) before anything ran. Result: 6/6 families
|
||
built and validated green on the first run attempt — evidence that the
|
||
deployment posture (docs + self-describing CLI, no sources) is sufficient
|
||
for the authoring role *if* the semantic gaps named in Arc 2's engine-gap
|
||
list are closed by documentation.
|
||
|
||
## The families and their verdicts
|
||
|
||
| family | hypothesis (compressed) | verdict |
|
||
|---|---|---|
|
||
| `mom_vnorm_h1` | sign of (close − close₋ₙ)/realized-vol carries direction | **candidate** on US500; null elsewhere |
|
||
| `meanrev_z_h1` | H1 deviations from SMA in std units revert | null |
|
||
| `bo_h1_volregime` | vol-expansion regime reshapes the H1 breakout | null (no reshaping that survives) |
|
||
| `trend_h4_cross` | H4 EMA-spread in vol units carries direction | null, OOS strongly negative |
|
||
| `bo_h4_trend` | Arc 1 pair stabilizes on H4 bars | null — textbook IS-overfit signature |
|
||
| `channelpos_fade_h1` | position in the recent range mean-reverts | null, extreme churn |
|
||
|
||
## Findings
|
||
|
||
1. **Vol-normalized momentum on US500 is the only regime-robust OOS positive
|
||
of the whole program.** Pooled WF-OOS gross bootstrap: mean +0.171
|
||
(P(E[R]≤0) = 0.026, n=8951) at vol{3,k2}; +0.102 (P = 0.036) at vol{60,k6};
|
||
+0.071 (P = 0.044) at vol{60,k12}. Winner parameters are stable
|
||
(mom_lag=72 ≈ 3 days, vol_avg=24) across regimes, and full-window net (at
|
||
slippage stress) turns positive at wide stops (+0.012 at k6, +0.028 at
|
||
k12). Deflated scores stay slightly negative — within its own 9-member
|
||
family the winner is not distinguishable from luck; the OOS consistency
|
||
across three (correlated) regimes is what makes this a candidate rather
|
||
than a null. GER40 leans the same way gross (OOS +0.10, P 0.15) without
|
||
getting there; EURUSD and XAUUSD are OOS noise.
|
||
2. **The H4 breakout is a lesson in why the pipeline exists.** Full-window
|
||
nets look excellent (XAUUSD: 7–9/9 members net-positive per regime, best
|
||
+0.37; US500 deflated +0.028 at k6) while the WF-OOS bootstraps are noise
|
||
(P 0.37–0.79) and generalize floors are the worst of the arc (−0.55 to
|
||
−1.40, sign 0/4). 90-day refit windows on H4 bars leave ~540 bars per
|
||
window — the refits chase noise the full-window fit averages away.
|
||
Without the WF/MC stages this family would have been a false discovery.
|
||
3. **Conditioners keep failing on this breakout.** Arc 1's trend gate and
|
||
arc 3's vol-expansion gate both fail to produce a reshaping that survives
|
||
the pipeline — the base H1 breakout's instrument-dependence (Arc 1's
|
||
core finding) is not explained by trend or vol regime.
|
||
4. **Mean-reversion families select at the grid edge** (`meanrev_z_h1` picks
|
||
window=24 and fade=0.5, both extreme values; `channelpos_fade_h1` runs
|
||
14 000–26 000 trades) — the H1 mean-reversion hypothesis wants a faster
|
||
timescale than this grid, and pays for it in churn. A refinement would
|
||
need finer bars and a cost-realistic churn budget, not more grid points.
|
||
|
||
## Verdict
|
||
|
||
Five honest nulls and one candidate. `mom_vnorm_h1` earns a breadth test
|
||
(`research/campaign-arc4-breadth-mom-vnorm.json`, 22 instruments) against
|
||
the always-long beta control (`research/campaign-control-always-long.json`)
|
||
— US500 2018–2026 is a secular bull market, and a 3-day momentum long-lean
|
||
on an index that mostly rose must beat beta, not zero, before it is
|
||
interesting. See `docs/arc-4-breadth-and-controls.md`.
|
||
|
||
## Engine friction
|
||
|
||
None new beyond the Arc 2 list. The campaign path carried six new blueprint
|
||
families, a degenerate one-member control sweep, and 22-instrument matrices
|
||
without a single executor complaint; every authoring error was caught at
|
||
`graph build` / three-tier validate time with a message that named the
|
||
problem.
|