Deep Mean-Reversion: A Physics-Informed Contrastive Approach to Pairs Trading

Abstract

Traditional pairs trading strategies often fail to identify stable mean-reverting relationships amid the complex dynamics characterizing modern financial markets, resulting in unstable spreads and reduced efficacy of conventional, rule-based execution methods. To overcome these intrinsic limitations, this study introduces ORCA (Ornstein-Uhlenbeck Reversion and Contrastive Arbitrage), a novel framework that seamlessly integrates deep representation learning and a rigorous financial dynamics model into a unified training paradigm. Central to ORCA is a physics-informed regularization approach designed to identify asset clusters that exhibit not merely similarity, but intrinsic tradability characterized by robust dynamic properties. ORCA concurrently optimizes a contrastive learning module alongside a Physics-Informed Neural Network (PINN) module, where the latter serves as a regularizer enforcing cluster formation consistent with the statistical dynamics of a stable Ornstein-Uhlenbeck process. Consequently, ORCA systematically produces asset clusters with inherently superior mean-reversion characteristics. Empirical analysis conducted on the NYSE dataset demonstrates the practical effectiveness of ORCA: applying a simple mean-reversion trading strategy with a static threshold to ORCA-generated clusters significantly outperforms strategies employing clusters derived via alternative benchmark methodologies. These findings position ORCA as a new benchmark methodology in the realm of structure-aware statistical arbitrage. A comprehensive overview of empirical results is provided in Figure 1.