Statistical Arbitrage with Market Regime Detection

Introduction

Statistical arbitrage strategies have long been a cornerstone of quantitative trading — exploiting mean-reverting relationships between co-integrated assets. However, their performance is highly sensitive to prevailing market conditions. During trending regimes, mean reversion signals generate significant drawdowns. During ranging markets, they thrive.

This research explores how Hidden Markov Models (HMMs) and regime-switching frameworks can dynamically adapt position sizing, entry thresholds, and strategy parameters based on detected market states.

Market Regime Classification

We define three primary regimes:

Low Volatility Mean-Reverting: Characterised by narrow spreads, high co-integration stability, and reliable convergence patterns. Optimal for aggressive mean reversion.

High Volatility Trending: Wide spreads, regime breakdowns, and momentum-driven moves. Mean reversion signals should be suppressed or inverted.

Transitional: Regime uncertainty where conservative position sizing and wider stop-losses are appropriate.

Methodology

Our approach employs a two-stage framework:

1. Regime Detection Layer: A Hidden Markov Model trained on rolling volatility, correlation stability, spread dynamics, and volume patterns classifies the current market state in real-time.

2. Strategy Adaptation Layer: Based on detected regime probabilities, the execution system dynamically adjusts:

Entry z-score thresholds (tighter in low-vol, wider in high-vol)

Position sizing (full in mean-reverting, reduced in transitional)

Stop-loss placement (regime-conditional)

Holding period expectations

Results

Backtesting across 10 years of Australian equity pairs data shows:

- Sharpe Ratio improvement: 0.85 → 1.42 (regime-adaptive vs. static)

Maximum drawdown reduction: 18.3% → 9.7%

Win rate improvement: 52% → 61%

Profit factor: 1.3 → 1.9

The most significant improvement came during volatile trending periods, where the regime filter prevented the strategy from taking mean reversion positions against strong directional moves.

Implementation Considerations

Production deployment requires careful attention to:

- Regime detection latency (our model updates every 15 minutes)

Transition probability calibration to avoid excessive switching

Out-of-sample validation across multiple market cycles

Transaction cost modelling for the adaptive position sizing

Conclusion

Regime-aware statistical arbitrage represents a meaningful evolution over static parameter approaches. By conditioning strategy behaviour on market state, we achieve significantly improved risk-adjusted returns while maintaining the theoretical foundation of co-integration-based trading.

At Neuground, we integrate these adaptive frameworks into our quantitative research platform, enabling systematic traders to deploy regime-conditional strategies with institutional-grade backtesting and live execution infrastructure.