Regime-aware
dynamic asset
allocation.

Fifty-six years of daily total returns across fourteen tradable assets. Treasury proxies are constructed from FRED yields using par-bond duration and convexity. Gold is sourced from Stooq XAUUSD daily fixings with the GLD expense ratio deducted. Real estate uses NAREIT total-return history before VNQ's 2004 inception. Style and quality factors come from AQR's research data. All proxies are vol-scaled and mean-adjusted into the live ETFs once available, preserving cumulative return continuity across the splice.

A four-component Gaussian Mixture Model on z-scored macro features — real yield, curve slope, inflation pulse, credit spread. Components are labeled Growth / Calm / Shock / Crash by sorting on the credit-spread mean (lowest CS → Growth). The raw GMM classification is smoothed by a Continuous State Jump Model with a transition penalty auto-calibrated to a 252-day target regime duration. Online Viterbi steps the regime forward each business day; full retrains happen only on a regime change.

Phase 18b — vol-conviction momentum across six sector ETFs (QQQ, SMH, XLE, XME, TLT, GLD), holding 100% in the asset with the strongest 126-day momentum gated by a 21-day-vol switching margin.
Phase 12c — top-N Hierarchical Risk Parity across nine assets, with a 50/50 blend of GARCH-shocked sample covariance and the regime prior, levered (or de-levered) to a regime-targeted volatility (12% / 10% / 8% / 6% in Growth/Calm/Shock/Crash).
Phase 23 — eight-asset MVO with an entropy-pooled posterior, regime-locked risk aversion and L2, a hard per-asset cap (35% per risky / 10% BIL), and an L1 turnover penalty against the prior day's drift.

Two scalars sit between the optimizer and the executed allocation. The crisis scalar is a deterministic function of an aggregate panic Z-score (equity vol, bond vol, correlation breakdown, curve-slope vol; EWMA-normalized) — linear ramp from 1.0 at +4σ to 0.0 at +5.5σ. The LL-velocity mitigator blends the portfolio toward 1/N when the GMM's log-likelihood collapses, signalling that current macro is unprecedented relative to the training distribution.

At every point in the backtest, the model has access only to data published before that day. The macro engine retrains on every regime change using only the historical window up to the retrain date. No future information leaks into past allocations. Without walk-forward retraining, the backtest would be an over-fit illustration rather than a track record.

Five of the ten core ETFs list inside the OOS window: TLT and IEF (July 2002), VNQ (September 2004), GLD (November 2004), DBC (February 2006), and BIL (May 2007). Pre-listing returns for each are a synthetic proxy — FRED par-bond total return for bonds, Stooq XAUUSD for gold, NAREIT for real estate, T-Bill carry for cash — vol-scaled and mean-adjusted into the live ETF using a 756-day overlap window starting at inception. Because that calibration window sits after the listing date, the splice scaling is forward-looking from the proxy era's perspective. Roughly 5.5 of the 24 OOS years partly evaluate against spliced synthetic returns; from June 2007 onward, every asset is live. CPI is shifted forward 45 days to cover the BLS publication lag, and the four Phase 18b sector extras (QQQ, SMH, XLE, XME) have no proxy and are excluded from the candidate universe before their respective inception dates.