The Meta Hierarchic MARS (Meta-Adaptive Reinforcement Learning System) Strategy is a sophisticated quantitative trading framework designed to solve the problem of market non-stationarity. Traditional single-logic strategies (e.g., pure trend following or mean reversion) inevitably fail when market regimes shift. This framework employs a Hierarchical Mixture-of-Experts (H-MoE) architecture, where specialized agents (Trend, Mean Reversion, Volatility, Hedge) are dynamically allocated capital by a "Meta-Controller" based on the prevailing market regime.
The strategy recognizes four distinct market phases:
- Growth: High drift, low volatility (Trend Following).
- Stagnation: Mean reverting, low volatility (Liquidity Provision).
- Crisis: High volatility, negative skew (Tail-Risk Hedging/Shorting).
- Transition: Rising volatility, directionless (Breakout/Long Volatility).
By adapting to these regimes in real-time, the MARS strategy aims to capture "Crisis Alpha" while minimizing drawdowns during choppy transitions.
The codebase tracks the evolution of this system from a conservative prototype to a high-octane alpha generator, and finally to a robust, institutional-grade risk-adjusted engine.
"The Stable Baseline" Version 3 established the core stability of the hierarchical model. It prioritized capital preservation and high Sharpe ratios over raw explosive returns.
- Characteristics: Extremely low Maximum Drawdown (Avg ~5.7%) and high risk-adjusted returns (Sharpe ~1.74).
- Behavior: Quick to de-lever in uncertainty. While it underperformed in raging bull markets compared to pure long-only strategies, it offered superior protection during corrections.
"The Convexity Engine" Version 4 was engineered to maximize total capture. It loosened the risk constraints to allow the agents to ride trends more aggressively.
- Characteristics: Massive raw returns (Avg ~88.1%) and high Buy & Hold capture (112%).
- Trade-off: The pursuit of alpha came at the cost of volatility. Average Max Drawdown spiked to ~18%, with volatile assets like TSLA seeing up to 36% drawdowns. It effectively beat the market but with higher variance.
"The Smart Defense" Version 5 represents the maturation of the strategy. It introduces a multi-layered defense system to curb the drawdowns of v4 without sacrificing its alpha-generating capability. It implements three specific mechanisms:
- Regime Momentum: Instead of reacting to instant signals, v5 tracks the momentum of regime probabilities over a 5-bar window, providing a 3-5 day early warning system for regime transitions.
- VIX Fear Gauge: Incorporates external market fear data (via
^VIX). When the VIX > 25, the baseline exposure is automatically engaged, overriding local signals to protect capital. - Drawdown-Adaptive Baseline: A dynamic feedback loop that scales down position sizing as drawdowns deepen, explicitly capping maximum loss potential.
The following benchmarks illustrate the progression of the strategy across five major technology tickers.
| Ticker | v1 | v2 | v3 | v4 | v5 | Buy & Hold |
|---|---|---|---|---|---|---|
| AAPL | +14% | +31% | +41% | +48% | +51% | +52% |
| MSFT | -0% | +14% | +25% | +12% | +18% | +2% |
| GOOG | +28% | +26% | +41% | +74% | +69% | +91% |
| NVDA | -16% | +42% | +53% | +125% | +138% | +110% |
| TSLA | +36% | +62% | +135% | +181% | +103% | +139% |
Note: v5 achieves superior returns on NVDA (+138%) compared to both v4 and Buy & Hold, while significantly smoothing out the equity curve.
| Metric | v3 | v4 | v5 |
|---|---|---|---|
| Avg Return | 58.9% | 88.1% | 75.7% |
| B&H Capture | 75% | 112% | 96% |
| Avg Alpha | 24.4% | 16.1% | 16.5% |
| Avg Sharpe | 1.74 | 1.34 | 1.42 |
| Avg MaxDD | 5.7% | 18.0% | 16.0% |
- Choose v4 if you want Maximum Raw Return and can stomach ~18-36% drawdowns. It is the "risk-on" version of the strategy.
- Choose v5 if you want Risk-Adjusted Consistency. It trades a small portion of the raw upside (Avg Return 88% -> 76%) to significantly improve the safety profile and reliability of the strategy.