The complete protocol behind the multi-asset rule engine: data inputs, signal construction, the 14,400-strategy grid search that picked the allocations, the seven walk-forward folds that validated them out-of-sample, the 17 rebalance rules tested for sensitivity, and the crisis decomposition.
The production book is EUR-denominated since 2026-08: VWCE (World Equity), LQEE (USD IG Bonds, EUR-hedged) and gold in EUR, benchmarked against VWCE buy & hold (EUR, spliced). History before instrument inceptions is spliced (VWCE ← VT ← 60/40 SPY/AEPGX at 2019-07 / 2008-06; LQEE ← synthetic-hedged LQD at 2017-09). The stress signals — and the validation studies on this page, which were run on the original USD book — remain computed on US market data (S&P 500, US yield curve, DXY) as validated.
Rule-Based, Not Machine-Learned
There's no training loop, no feature selection, no calibrator, no model artifact, no random seed. The strategy is a small set of hand-picked rules that take six daily inputs and produce a daily 3-asset allocation. That makes it trivial to audit, easy to reproduce, and impossible to overfit going forward.
The price of that simplicity is that we can't squeeze the last basis point of in-sample performance — but we showed (via walk-forward refutation) that doing so would have actively hurt out-of-sample results.
Six daily signal inputs: total-return adjusted closes for SPY, LQD, GLD and the DXY dollar index, plus the T10Y2Y yield-curve spread and the THREEFYTP10 term premium — all standard published US-market series. The traded book adds the EUR sleeve prices (VWCE, LQEE, gold in EUR). Nothing proprietary or hand-collected.
Six independent stress signals (price momentum, yield-curve inversion, realised vol, dollar momentum, gold momentum, term-premium Z-score) are computed point-in-time and OR'd into a single composite stress flag.
Composite OFF → 50% VWCE / 30% LQEE / 20% Gold, with a dip overlay lifting equity to 80% when SPY is 5-10% off its 21-day high (nominally +40 ppts, +30 effective once the cap binds). Composite ON → 10% VWCE / 60% LQEE / 30% Gold. Target weights are capped to [10%, 80%] and renormalised; realised weights drift between rebalances.
Each fold trains its “view” on 5 years of data and tests on the next 3 — then steps forward by 3 years. The seven folds cover 2007-07 to 2026-05, all of it out-of-sample. We tested three variants:
The allocation chosen up-front and applied identically in every fold. Best out-of-sample performer.
The in-sample winner. Marginally higher Sharpe, but worse drawdown and worse on every secondary metric — classic overfit.
Re-runs the grid search at each fold boundary. Loses ~0.16 Sharpe — re-fitting is worse than committing to a robust fixed allocation.
We enumerated 120 normal-regime allocations × 120 defensive-regime allocations in 5%-step grids (subject to summing to 100% and each leg in [10%, 80%]). The production pair sits comfortably inside the top-Sharpe basin and the top-Sortino basin — not at any sharp peak.
Among all 14,400 grid points.
Better tail risk than ~98% of variants.
Return per unit of max drawdown.
We tested daily, weekly, monthly, quarterly, semi-annual, and annual calendar rebalances, drift-only thresholds at 2%, 5%, 10%, 15%, and 20%, plus combined rules. Net of 5 bps trading costs:
Figures from the USD validation book (v1: SPY/LQD/GLD vs S&P 500 buy & hold), on which the rules were selected and validated. The live backtest page shows the EUR book.
| Episode | Strategy | S&P 500 buy & hold (USD, v1) | Note |
|---|---|---|---|
| 2008 GFC | +2.7% | −41.8% | 189d-mom + yield-curve signals fired early. |
| 2020 COVID | −0.3% | −20.0% | Realised vol & DXY signals fired in days. |
| 2022 bear | −12.0% | −19.4% | Hardest scenario: stocks & bonds fell together; gold floor helped. |
| 2023–25 rally | +67.5% | +59.3% | Captured most of the upside in the calm regime. |
The live allocation is updated every weekday at 04:30 UTC. The full backtest is on the backtest page.