Aegis-X: A Risk-First Regime-Adaptive Trading System

Aegis-X is a risk-first trading system that adapts to market regimes. Instead of predicting prices, it dynamically adjusts exposure to reduce drawdowns and protect capital across market cycles.

Comment

About the Project

Why Aegis-X Exists

Most trading systems fail for a simple reason: they assume markets behave the same way all the time. In reality, markets move through distinct regimes — calm, stressed, and transitional — and strategies that ignore this reality suffer large drawdowns during crises.

Aegis-X was built with a different mindset. Instead of asking “Where will price go next?”, it asks:

“Is this a safe environment to deploy capital?”

The goal is not aggressive profit maximization, but capital survival, drawdown control, and long-term robustness — the same priorities used by professional risk managers.

What the System Does

Aegis-X is a risk-first, regime-adaptive trading system for U.S. markets.

Rather than predicting prices, it uses unsupervised machine learning to detect the current market regime and dynamically adjust exposure.

  • In constructive regimes, capital is allocated to Equities (SPY)
  • In stress regimes, exposure is reduced or rotated into Defensive assets (TLT / Cash)

This allows the system to step aside during periods of elevated risk instead of riding market crashes.

Core Intelligence (Regime Detection)

The system uses a Gaussian Mixture Model (GMM) to cluster market behavior into latent regimes using structurally meaningful features:

  • Realized volatility (21-day, annualized)
  • Volatility acceleration
  • Medium-term momentum (63-day)
  • Distance from long-term trend (SMA-200)

Volatility is computed as:

$$ \sigma_{21} = \sqrt{252} \cdot \mathrm{StdDev}(r_{t-21:t}) $$

The cluster with higher average volatility is deterministically labeled as the Stress Regime, ensuring interpretability and auditability.

How Decisions Are Made (No Black Box)

Instead of binary signals, Aegis-X uses probabilistic confidence:

  • The GMM outputs regime probabilities
  • Probabilities are smoothed using a persistence filter
  • Asset allocation is adjusted continuously, not abruptly

Example (simplified):

if prob_stress > threshold:
    allocate_to = "TLT"
else:
    allocate_to = "SPY"

This reduces over-trading and prevents reaction to short-term noise.

Risk Management & Execution Discipline

The system enforces strict safety constraints:

  • No leverage
  • Kelly Criterion-based exposure scaling (with conservative caps)
  • Volatility-adjusted transaction costs
  • One-day execution lag to avoid look-ahead bias

Performance is evaluated using drawdowns, Calmar ratio, and crisis behavior, not just raw returns.

Validation Integrity

To ensure correctness:

  • An expanding-window walk-forward backtest is used
  • Models are retrained using only historical data available at the time
  • A true out-of-sample holdout is preserved

This prevents accidental future leakage — a common flaw in many trading demos.

What Makes This Different

  • Focuses on risk control, not prediction hype
  • Uses simple, explainable models instead of opaque deep learning
  • Explicitly documents limitations (e.g., stock-bond correlation breakdowns)
  • Designed with live deployment discipline in mind

Aegis-X is not a trading bot — it is a decision framework built to survive real markets.

Built With

  • language:-python-libraries:-pandas
  • matplotlib-data:-historical-u.s.-market-data-(spy
  • numpy
  • risk-based
  • scikit-learn
  • tlt)-methods:-unsupervised-learning
  • walk-forward-validation
Share this project:

Updates