How It Works Under the Hood
Factor Engine
Five cross-sectional factors – momentum, value, quality, low volatility, and carry – scored using sector-neutral z-score normalization with winsorizing at three standard deviations. Minimum universe coverage of 80% or 50 tickers enforced before scoring. Optional LightGBM combination with a deployment gate requiring holdout information coefficient above 0.02.
Regime Detection
Dual-layer detection. Primary: a Gaussian Hidden Markov Model trained on six features – SPY returns, realized volatility, VIX proxy, credit spread proxy, yield curve slope, and SPY-credit correlation. Three states (risk-on, risk-off, crisis) with automatic recovery detection. Fallback: rule-based volatility regime using 200-day SMA trend and 252-day volatility percentile.
Portfolio Construction
Hierarchical Risk Parity with Ledoit-Wolf covariance shrinkage and Ward hierarchical clustering. Regime-conditional factor tilts adjust weight multipliers across all five factors. Turnover constraints via linear interpolation prevent whipsaw rebalancing. Drift threshold triggers at 5% maximum absolute weight deviation.
Validation Pipeline
Three-layer graduation: minimum viability (30+ trades, Sharpe above 1.0, max drawdown below 30%), walk-forward robustness (positive out-of-sample Sharpe across all windows, robustness ratio above 0.5), and deflated Sharpe ratio above 0.95 to control for multiple comparisons. Live monitoring uses graduated decay detection across 30, 60, and 120-day windows with Fama-MacBeth cross-sectional regression for factor-level attribution.