Low Vol— how it works

The low-volatility anomaly is one of the most durable facts in equities: low-volatility stocks have historically delivered better risk-adjusted returns than the high-flyers, not worse. This book is a clean expression of it — long the calmest names, rebalanced daily, no shorts.

The pick engine is a 20-model momentum ensemble: four XGBoost ensembles (5 seeds each), each trained on a different asymmetric “will this stock spike?” label — 30d/+20%, 45d/+30%, 60d/+40%, 90d/+50%. Each name's four probabilities are rank-averaged into one cross-sectional score per day.

The book then goes long the lowest-scored 20 names — the stocks the momentum model judges least likely to spike. Those are precisely the low-volatility, stable names: calm price action, no breakout structure, mostly trading comfortably within their range. Inverting a momentum score is a clean, adaptive way to isolate them.

Originally this book went long the model's top picks — the predicted spikers. An honest walk-forward diagnosis (research 151–154) showed that configuration was anti-predictive: negative rank-IC, a SPY-hedged Sharpe of −0.28, and it had been greenlit by a backtest that booked each trade's P&L on its exit date with no daily marks (hiding all drawdown). The signal, inverted, is the opposite — a robust low-volatility book that is positive in every walk-forward fold. So the book was flipped rather than retired.

Daily entries: the 20 lowest-scored names, subject to a 7-day per-ticker reentry cap and no overlap with currently-held names. Capital across books is set by risk-parity weighting — Low Vol is the calmest book, so it carries the largest weight (~35% of the aggregate); the per-leg figure is provisional while the rebuilt book ramps. 10% trailing stop on the highest close since entry, 90-day maximum hold otherwise. Long-only, no shorts.

• Sharpe 1.51 · CAGR +15.6% · max drawdown −13.6%
• Annualized volatility ~10% — genuinely low-vol
• Trade win rate 55%
• Positive in all five folds (fold Sharpes +1.8 / +1.4 / +1.1 / +2.3 / +1.3)
• SPY beta 0.49; SPY-hedged Sharpe +0.76 — real edge beyond market exposure

Numbers are walk-forward across 2024-01 → 2026-05 with true daily portfolio marks and 10 bps round-trip cost — the same honest recompute methodology used to audit every book.

Weekly retrain Saturday 09:30 UTC; rolling 30-month training window; exponential 6-month half-life sample weighting so the model adapts to regime shifts.