This paper presents a structured synthesis of the statistical arbitrage literature, tracing the evolution of the field from classical mean-reversion and cointegration frameworks to contemporary machine learning and reinforcement learning architectures. Through a comparative analysis of prior empirical studies across equity, ETF, and cryptocurrency markets, the paper finds that although rising market efficiency and structural complexity have weakened the persistence of traditional signals, statistical arbitrage retains meaningful potential when supported by adaptive modeling and robust risk management. To address challenges related to tail-risk exposure, model fragility, and declining interpretability, this study proposes a human–AI collaborative execution framework. In this system, machine intelligence focuses on high-dimensional signal extraction and pattern recognition, while human oversight enforces risk constraints, contextual judgment, and interpretability requirements. This complementary structure is operationalized through CVaR-based pre-execution screening, multi-horizon drawdown controls, and explanability-guided intervention, forming a resilient and practically deployable architecture for risk-aware statistical arbitrage.
Baiying Lei (Tue,) studied this question.