Title: Regime-Adaptive Continual Learning for Portfolio Management

Abstract:

Traditional Portfolio Management (PM) strategies often falter in financial markets due to their inherent non-stationarity, which is characterized by frequent structural changes and regime shifts. While existing solutions like naive online fine-tuning and rolling-window retraining have been proposed, they suffer from significant drawbacks: the former fails to utilize accumulated knowledge effectively, while the latter incurs prohibitive computational expenses. Consequently, these methods yield limited adaptability and suboptimal returns. Continual learning (CL) presents a viable alternative, allowing trading agents to accumulate and transfer insights across sequential tasks. To tackle the complexities of dynamic financial landscapes, this study introduces Regime-aware Continual Adaptive Portfolio management (ReCAP), a novel framework that embeds CL into PM. ReCAP utilizes an adaptive regime detection module to partition historical market data into regimes of varying lengths, thereby facilitating the learning of regime-specific policy vectors and the creation of a comprehensive policy library. During the trading process, a regime-gate module dynamically synthesizes policy vectors from this library according to the prevailing market conditions, ensuring swift adaptation to newly identified regimes. Crucially, ReCAP updates only the regime-gate and the policy vector associated with the current regime, thereby safeguarding previously acquired knowledge. Empirical evaluations across five real-world datasets confirm that ReCAP consistently surpasses established baselines, delivering enhanced long-term investment returns and demonstrating rapid responsiveness to regime transitions.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC