Cross-asset RS rotation (multi-sleeve ETF momentum)

ID: S031 Slug: cross-asset-rs-rotation-v1 Run date: 2026-07-01 Outcome: FAILED Headline metric: best-cell Sharpe 0.59 vs equal-weight baseline 0.78 (and max-DD 30% vs 20%) Fail reason: The rotation adds drawdown and volatility without adding return. No lookback beats the equal-weight sleeve baseline on a risk-adjusted basis; the best (12-month) beats it only on raw CAGR while running a deeper drawdown (30% vs 20%), a lower Sharpe (0.59 vs 0.78), and an insignificant excess (Bonferroni p = 1.0). And the whole cross-asset premise loses to simply owning SPY. The "less correlated" thesis does not survive honest costs and a tail-aware gate.

What we tested

A frozen 26-sleeve ETF universe (9 sectors · 5 commodities · 4 bonds · 2 currencies · 6 countries, all with full 2010→2026 history), ranked monthly by trailing L-month total return, holding the top 3 equal-weight, rebalanced monthly, 30 bps on turnover. Lookback sweep L ∈ {4, 6, 9, 12} months — four pre-registered cells. Baseline: equal-weight buy-and-hold of all 26 sleeves; secondary sanity: SPY. Best-cell significance deflated (Bonferroni) over the four cells; S004's prior counts in the C1 family.

What we found

Gate (best cell, L=12): return beats baseline PASS (8.9% > 8.1%) — but drawdown FAIL (30% > 25% ceiling and > the 20% baseline), Sharpe FAIL (0.59 < 0.78), deflated significance FAIL (Bonferroni p = 1.0). Cut.

Why this matters

Three canonical swing edges — short-term reversion (B1), trend-continuation (A1), cross-asset RS rotation (C1) — each mechanised honestly, run survivorship-free with real costs and a tail-aware gate, and each cut on its own structural wall. C1's wall is the one its own sales pitch glosses: cross-asset momentum is a single factor that re-correlates in a crash, so concentrating on it buys drawdown, not diversification. This is exactly why the pitch must be sold honestly — and here it isn't sold at all, because it doesn't clear the gate.

What this doesn't tell us yet

• The result is sample-dependent (2010→2026 was a strong US-equity regime that penalised cross-asset diversification); a different regime could differ — but that is precisely the un-robustness a walk-forward would punish, and Stage 1 already fails.
• It condemns this top-3 momentum rotation, not every cross-asset idea — a dual-momentum (absolute filter), a risk-parity weighting, or a wider-hold variant are separate S-NNNs. But the burden is now high: the base rotation lost to buy-and-hold on every risk-adjusted measure.

What happens next

No Stage 2. With B1 and A1 closed and C1 cut on its first configuration, the swing pillar's three a-priori families are all answered no at Stage 1 under honest gates. The disciplined reading is to treat the swing pillar as substantially answered — a transparent fail record across the canonical edges — and to redirect research budget to the live pillars (0DTE forward collection, the iron-fly drift line), unless the owner wants to spend a trial on a specific C1 variant (dual-momentum / risk-parity) with eyes open to the high bar.

S031 is recorded failed; C1 cumulative family trials = 4 cells + S004.