Brummel/Aura
1
0
Fork 0
Code Issues 31 Pull Requests Actions Packages Projects Releases Wiki Activity

Research lead: slow-momentum positive on GER40 but not FX — an equity-index trend effect? #141

New Issue
Closed
opened 2026-06-25 09:15:13 +02:00 by Brummel · 1 comment
Brummel commented 2026-06-25 09:15:13 +02:00
Owner
Copy Link
Copy Source

The lead

The first in-sample momentum screen (frictionless Stage-1 R, full history,
slow SMA-cross signal + vol-stop defining R) found a positive E[R] on
GER40 at the 4h–8h / 32h cross timescales — the first positive E[R] aura has
produced, and at the timescale theory predicts (Phase 0 showed fast signals
are chop-dominated noise on M1).

But the cross-symbol killer test failed for a universal edge:

 fast/slow    EURUSD    GBPUSD     GER40    USDCAD   #pos(of 4)
 240/1920   -0.0262   -0.0142   +0.0687   -0.0468     1
 240/7680   -0.0431   -0.0325   +0.0088   +0.0060     2
 480/1920   -0.0170   +0.0104   +0.0722   -0.0509     2
 480/7680   -0.0144   -0.0060   -0.0129   +0.0339     1
1920/7680   +0.0736   -0.0127   -0.1061   -0.0163     1

No single (fast/slow) cell is positive across all four symbols. The GER40
momentum cells (+0.069, +0.072) are negative on the FX majors. Had we
stopped at GER40, we'd have "found" an overfit, symbol-specific result — the
exact 5-year trap; the cross-symbol test caught it. (lengths in M1 bars;
stop_length 1920, stop_k 2.0; frictionless.)

The reframed hypothesis (grounded, not post-hoc rescue)

"GER40 trends, FX majors mean-revert" is consistent with a well-documented
prior: equity indices exhibit time-series momentum / trend; FX majors are
more mean-reverting intraday. So the candidate is not "universal momentum"
but equity-index trend-following. Treat this as a new hypothesis and
test it on evidence not yet looked at:

  • Other indices: FRA40 (CAC, already wired — test now) and NAS100 (US
    region, after the instrument-spec issue). If GER40 + FRA40 + NAS100 all show
    positive slow-momentum in-sample, the "index trend" effect is credible; if
    only GER40, it stays a single-symbol artifact.
  • Temporal OOS on GER40: split the history (e.g. IS 2014–2020, OOS
    2021–2026) — is the positive robust, or one-regime (e.g. the 2020–21 bull)?
  • Only if it survives both: Monte-Carlo on the OOS trades, then Stage-2 costs.

Honesty guard

A single-symbol, in-sample positive driven by rare tail winners (win ~14 %,
avg-win ~10 R, many small losses) is a hypothesis, not an edge. It counts
only after temporal OOS + cross-index consistency + MC + costs. The asset-class
reframe must be tested forward, not assumed.

Depends on

The OOS validation harness issue (for the temporal split / MC) and the
memory BLOCKER (for full-history runs at scale). FRA40 can be screened now,
throttled.

refs #137

## The lead The first in-sample momentum screen (frictionless Stage-1 R, full history, slow SMA-cross signal + vol-stop defining R) found a **positive `E[R]` on GER40** at the 4h–8h / 32h cross timescales — the first positive E[R] aura has produced, and at the timescale theory predicts (Phase 0 showed fast signals are chop-dominated noise on M1). But the **cross-symbol killer test failed for a universal edge**: ``` fast/slow EURUSD GBPUSD GER40 USDCAD #pos(of 4) 240/1920 -0.0262 -0.0142 +0.0687 -0.0468 1 240/7680 -0.0431 -0.0325 +0.0088 +0.0060 2 480/1920 -0.0170 +0.0104 +0.0722 -0.0509 2 480/7680 -0.0144 -0.0060 -0.0129 +0.0339 1 1920/7680 +0.0736 -0.0127 -0.1061 -0.0163 1 ``` No single (fast/slow) cell is positive across all four symbols. The GER40 momentum cells (+0.069, +0.072) are negative on the FX majors. **Had we stopped at GER40, we'd have "found" an overfit, symbol-specific result — the exact 5-year trap; the cross-symbol test caught it.** (lengths in M1 bars; stop_length 1920, stop_k 2.0; frictionless.) ## The reframed hypothesis (grounded, not post-hoc rescue) "GER40 trends, FX majors mean-revert" is consistent with a well-documented prior: **equity indices exhibit time-series momentum / trend; FX majors are more mean-reverting intraday.** So the candidate is not "universal momentum" but **equity-index trend-following**. Treat this as a *new* hypothesis and test it on evidence not yet looked at: - **Other indices**: FRA40 (CAC, already wired — test now) and NAS100 (US region, after the instrument-spec issue). If GER40 + FRA40 + NAS100 all show positive slow-momentum in-sample, the "index trend" effect is credible; if only GER40, it stays a single-symbol artifact. - **Temporal OOS on GER40**: split the history (e.g. IS 2014–2020, OOS 2021–2026) — is the positive robust, or one-regime (e.g. the 2020–21 bull)? - Only if it survives both: Monte-Carlo on the OOS trades, then Stage-2 costs. ## Honesty guard A single-symbol, in-sample positive driven by rare tail winners (win ~14 %, avg-win ~10 R, many small losses) is a **hypothesis, not an edge**. It counts only after temporal OOS + cross-index consistency + MC + costs. The asset-class reframe must be *tested forward*, not assumed. ## Depends on The OOS validation harness issue (for the temporal split / MC) and the memory BLOCKER (for full-history runs at scale). FRA40 can be screened now, throttled. refs #137
Brummel added this to the Edge research: prove a real tradeable edge milestone 2026-06-25 09:15:13 +02:00
Brummel added the idea label 2026-06-25 09:15:13 +02:00
Brummel commented 2026-06-25 11:59:28 +02:00
Author
Owner
Copy Link
Copy Source

Verdict: refuted — slow-SMA momentum is not a structural index-trend edge

Tested the index-trend reframe forward on the second wired index (FRA40, CAC)
plus a temporal IS/OOS split on both indices, frictionless Stage-1 R, same
grid as the original screen (fast ∈ {240, 480, 1920} × slow ∈ {1920, 7680},
stop_length 1920, stop_k 2.0). Split at 2021-01-01.

Cross-tab, E[R] per cell:

fast/slow GER40 IS GER40 OOS GER40 full FRA40 IS FRA40 OOS FRA40 full
240/1920 +0.074 +0.066 +0.069 −0.029 −0.110 −0.068
240/7680 +0.150 −0.107 +0.009 +0.125 +0.153 +0.140
480/1920 +0.099 +0.050 +0.072 +0.073 −0.142 −0.030
480/7680 +0.038 −0.058 −0.013 +0.121 +0.158 +0.140
1920/7680 +0.025 −0.207 −0.106 +0.156 −0.058 +0.061

Two independent failures:

1. No cross-index generalization. GER40 is temporally stable at fast/1920
(240/1920: +0.074 IS / +0.066 OOS); FRA40 is temporally stable at fast/7680
(240/7680: +0.125 / +0.153). But the timescales do not transfer: GER40's robust
240/1920 is negative on FRA40 (−0.029 / −0.110), and FRA40's robust 240/7680
fails GER40 OOS (+0.150 → −0.107). No single cell is positive across both
indices and both temporal halves. A genuine "indices trend" structural effect
would share a timescale; instead each index has its own idiosyncratic positive
cell — the signature of per-symbol overfit, not structure. Two neighbouring,
highly-correlated European indices showing opposite sign at 240/1920 (GER40
+0.069 vs FRA40 −0.068 full-history) is decisive.

2. Not statistically significant even within a symbol. The raw sqn metric
is exactly the one-sample t-statistic for E[R]=0 (sqn = √n·mean/sd,
report.rs:171). Every positive cell has |t| ≲ 1.2 even pooled full-history:
GER40 240/1920 t≈1.22 (n=17,706), 480/1920 t≈1.16; FRA40 240/7680 t≈1.01,
480/7680 t≈0.95. None rejects E[R]=0. The apparent IS/OOS sign-stability on
GER40@1920 sits inside the noise band, and the larger cells flip sign across the
split (240/7680 on GER40: +0.150 → −0.107) exactly as sampling noise around zero
would.

Combined with the earlier FX-negative screen and the fast-SMA null (Phase 0),
this exhausts the simple moving-average momentum family on the tradeable
universe: fast MA = chop-dominated null, slow MA = per-symbol non-significant,
non-generalizing noise. The index-trend hypothesis is refuted; closing this
lead. The cross-symbol + temporal-OOS discipline did its job — it caught what
a GER40-only (or even GER40+FRA40-IS-only) screen would have sold as an edge.

Next: a structurally different trend entry (breakout / channel-extreme — the
untested flavour of the #137 lead candidate), under the same discipline. refs #137

## Verdict: refuted — slow-SMA momentum is not a structural index-trend edge Tested the index-trend reframe forward on the second wired index (FRA40, CAC) plus a temporal IS/OOS split on **both** indices, frictionless Stage-1 R, same grid as the original screen (fast ∈ {240, 480, 1920} × slow ∈ {1920, 7680}, stop_length 1920, stop_k 2.0). Split at 2021-01-01. Cross-tab, **E[R] per cell**: | fast/slow | GER40 IS | GER40 OOS | GER40 full | FRA40 IS | FRA40 OOS | FRA40 full | |---|---|---|---|---|---|---| | 240/1920 | +0.074 | +0.066 | +0.069 | −0.029 | −0.110 | −0.068 | | 240/7680 | +0.150 | −0.107 | +0.009 | +0.125 | +0.153 | +0.140 | | 480/1920 | +0.099 | +0.050 | +0.072 | +0.073 | −0.142 | −0.030 | | 480/7680 | +0.038 | −0.058 | −0.013 | +0.121 | +0.158 | +0.140 | | 1920/7680 | +0.025 | −0.207 | −0.106 | +0.156 | −0.058 | +0.061 | Two independent failures: **1. No cross-index generalization.** GER40 is temporally stable at fast/**1920** (240/1920: +0.074 IS / +0.066 OOS); FRA40 is temporally stable at fast/**7680** (240/7680: +0.125 / +0.153). But the timescales do not transfer: GER40's robust 240/1920 is **negative** on FRA40 (−0.029 / −0.110), and FRA40's robust 240/7680 **fails** GER40 OOS (+0.150 → −0.107). No single cell is positive across both indices *and* both temporal halves. A genuine "indices trend" structural effect would share a timescale; instead each index has its own idiosyncratic positive cell — the signature of per-symbol overfit, not structure. Two neighbouring, highly-correlated European indices showing **opposite sign** at 240/1920 (GER40 +0.069 vs FRA40 −0.068 full-history) is decisive. **2. Not statistically significant even within a symbol.** The raw `sqn` metric is exactly the one-sample t-statistic for E[R]=0 (`sqn = √n·mean/sd`, `report.rs:171`). Every positive cell has |t| ≲ 1.2 even pooled full-history: GER40 240/1920 t≈1.22 (n=17,706), 480/1920 t≈1.16; FRA40 240/7680 t≈1.01, 480/7680 t≈0.95. None rejects E[R]=0. The apparent IS/OOS sign-stability on GER40@1920 sits inside the noise band, and the larger cells flip sign across the split (240/7680 on GER40: +0.150 → −0.107) exactly as sampling noise around zero would. Combined with the earlier FX-negative screen and the fast-SMA null (Phase 0), this exhausts the simple moving-average momentum family on the tradeable universe: fast MA = chop-dominated null, slow MA = per-symbol non-significant, non-generalizing noise. **The index-trend hypothesis is refuted; closing this lead.** The cross-symbol + temporal-OOS discipline did its job — it caught what a GER40-only (or even GER40+FRA40-IS-only) screen would have sold as an edge. Next: a structurally different trend entry (breakout / channel-extreme — the untested flavour of the #137 lead candidate), under the same discipline. refs #137
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
BLOCKER
Core functionality blocked; preempts other selection
 bug
Observable misbehaviour
 feature
Substantial extension; no full spec required
 idea
Droppable; no commitment
No Label idea
Milestone
No items
No Milestone Edge research: prove a real tradeable edge
Projects
Clear projects
No project
Assignees
Clear assignees
claude (Claude) Brummel
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: Brummel/Aura#141
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?