HYDRA is a professional-grade ensemble targeting institutional risk-adjusted return via uncorrelated-sleeve diversification rather than concentrated leverage. 20 independently-constructed sleeves span 8 alpha categories: equity trend, fixed income, commodity/energy, FX, volatility, crypto, cross-asset, and alternatives. Each sleeve is vol-targeted to 10% annualised; the ensemble is weighted inverse-vol (risk parity) and vol-targeted at 20% with a 5× gross cap.
Objective: Sharpe above 1.5 full-sample, above 2.0 out-of-sample, with max drawdown bounded under −20% — delivered without hindsight-biased sleeve selection, concentrated leverage, or look-ahead. Alpha comes from sleeve diversification (mean |pairwise correlation| ≈ 0.17), not from stacking a single bet.
Construction: Every sleeve uses 1-bar signal lag, 15 bps transaction cost on turnover, monthly sleeve rebalancing, daily inverse-vol weight updates, and rolling 63-day volatility scaling (5% floor, 1.5× scaling cap at the sleeve level). Walk-forward filter and regime overlays were tested and rejected — both hurt net performance (filter dropped sleeves at bottoms before recoveries; overlay was net-neutral).
Diversification math: With N=20, avg sleeve Sharpe ≈ 0.5, avg pairwise correlation ≈ 0.17, the equal-weight diversified Sharpe ceiling is ≈ 1.1. Inverse-vol weighting plus sleeve design (JPY safe-haven, dollar-neutral long-short, crisis hedges) lifts realised full-window Sharpe to 1.58 and OOS Sharpe to 2.01.
Honest ceiling: After extensive iteration, full-window Sharpe ≈ 1.6 / OOS Sharpe ≈ 2.0 is the realistic upper bound for a 21-year backtest with honest construction. Hitting Sharpe 3 over 21 years requires hindsight-biased sleeve selection, concentrated leverage (the METEOR-style approach produced −78% MDD in its 21y proxy), or sleeves exploiting regimes that won't repeat. HYDRA is the defensible professional-grade alternative.
HYDRA-Lite. A simplified execution variant is included for operators preferring monthly-only rebalancing with no dynamic vol scaling at the portfolio level: equal-weight sleeves, static 2.43× leverage, single trade per month. It achieves Sharpe 1.20 / CAGR 12% / MDD −17% — still clearly superior to SPY but trading ~0.4 SR for operational simplicity. See comparison section below.
| Sleeve Rebalance | Monthly (21 trading days). Each sleeve's signal (momentum, trend, carry, regime) is recomputed once a month from T−1 close data. Monthly cadence minimises turnover noise while preserving sleeve responsiveness. |
| Sleeve Construction | Each of the 20 sleeves is a standalone long-only, short-only, or long-short rule acting on liquid ETFs (and BTC). Every sleeve is independently vol-targeted to 10% annualised via a rolling 63-day realised vol, with a 5% floor and a 1.5× scaling cap. |
| Portfolio Weighting | Inverse-vol risk parity. Daily weight on sleeve i ∝ 1 / σ_i(63d). Sleeves not yet live (pre-inception of their ETF) have zero weight; weights renormalise to sum to 1 across active sleeves. This equalises risk contribution, not dollar exposure. |
| Portfolio Vol Target | 20% annualised vol, 5× gross cap. After the inverse-vol blend, the portfolio is scaled daily by target_vol / realised_63d_vol, clipped at 5×. In practice the leverage is typically 1.5–3×. No path-dependent throttle, no CPPI — pure vol targeting. |
| Signal Lag | 1 full bar. Signals computed from T−1 close are applied from T open onward. This is the honest no-look-ahead convention and is cross-checked against T-open execution in the backtest. |
| Transaction Costs | 15 bps on turnover. Applied at the sleeve level on each rebalance and at the ensemble level on daily weight drift. Because inverse-vol weights drift slowly, daily turnover is small (monthly turnover dominates). |
| What is NOT Done | No walk-forward sleeve filter (tested: dropped sleeves at bottoms, SR 1.57 → 1.01). No regime overlay (tested: net-neutral). No min-variance optimisation (tested: concentrated in low-vol sleeves, SR 1.58 → 1.14). No SR-tilt weighting (tested: momentum-chased weak periods, SR 1.59 → 1.07). Rejecting these is the finding. |
HYDRA-Lite is a stripped-down execution variant designed for operators who want monthly-only rebalancing with no dynamic vol scaling at the portfolio level. The sleeve rules are the same; only the ensemble layer changes.
| Weighting | |
| Rebalance | |
| Leverage | |
| Portfolio vol scaling |
What does leverage actually buy you? This section isolates the two design choices that produce HYDRA's headline numbers — leverage (the portfolio scalar that can amplify the composite by up to 5×) and dynamic inverse-vol weighting (which reshuffles sleeve weights daily) — by showing all four combinations over the same 21-year window. Comparing horizontally isolates the leverage effect; comparing vertically (HYDRA-pair vs Lite-pair) isolates the weighting effect.
How the vol-targeter decides today's gross exposure. Every day the strategy measures the trailing 63-day realised vol of the inverse-vol-weighted sleeve composite (the "raw" portfolio, pre-scale). It then sets the gross-exposure scalar to 20% / raw_vol_63d, clipped at the 5× leverage cap. That scalar is applied to day-T's return — the realised line below shows the resulting 63-day vol of HYDRA's scaled output. Both series use .shift(1) on the rolling window, so today's sizing is based only on data through T−1 close; orders fill at T open.