Regime-Adaptive Continual Learning for Portfolio Management

Battle Count: 63

This paper is directly and highly relevant to quantitative trading. It addresses the fundamental challenge of non-stationarity in financial markets (alpha decay) and proposes a principled framework for adaptive portfolio management. The work covers multi-asset allocation, regime detection, continual adaptation, and risk management across diverse markets (US equities, Japanese equities, commodity ETFs). The experimental setup with 17 years of data, realistic transaction costs, and comparison against 13+ baselines makes it immediately applicable to real-world quantitative trading systems. The continual learning paradigm directly addresses the practical need for trading strategies to adapt over time without full retraining.

The framework involves multiple interacting components: CUSUM-based regime detection, policy fine-tuning per regime, policy vector computation and library management (merging/pruning), a Regime-Gate Module with dynamic output head expansion/pruning, and the overall continual trading loop. While each component is individually implementable, the orchestration of the full pipeline (Algorithm 1 and 2) requires careful engineering. The use of standard RL (PPO) and PyTorch/FinRL reduces some complexity, but the adaptive library management, regime-gate attention mechanism, and online regime detection add significant implementation overhead. Hyperparameter sensitivity analysis shows robustness, which helps practical deployment.

Code is publicly available at https://github.com/Dumail/ReCAP. Detailed hyperparameters are provided (κ=0.5σ, h=2.5σ, δs=0.5, learning rate=10^-4, batch size=256, 10 random seeds). Datasets are from Yahoo Finance (publicly accessible). Implementation uses PyTorch and FinRL. However, some details about the exact CUSUM implementation, feature engineering specifics, and the full training pipeline could benefit from more documentation. The paper provides two algorithms (Algorithm 1 and 2) for the complete workflow.