A Computational Systems Perspective on Risk-Aware Portfolio Optimization with Actor–Critic Learning
Main article
Abstract
This study reframes portfolio optimization as a computational systems problem rather than a purely financial forecasting task. We argue that the effectiveness of reinforcement learning in dynamic capital allocation depends on the co-design of data engineering, state representation, actor–critic interaction, and execution-aware risk governance. On this basis, the paper develops a modular architecture, Computational Systems Actor–Critic Learning (CS-ACL), that combines data harmonization, discriminative state compression, clipped actor–critic learning, and a risk controller that penalizes drawdown, turnover, and unstable policy shifts. Two public datasets are used: a large market panel derived from Yahoo Finance and a structured intelligent-finance dataset with allocation signals and asset categories. The framework is evaluated under walk-forward backtesting against mean–variance allocation, LSTM-assisted allocation, PPO, SAC, and TD3 baselines. Results show that CS-ACL achieves the strongest overall balance among return, volatility, drawdown, turnover, and convergence stability. Under identical transaction-cost assumptions, the model delivers an annualized return of 18.7%, a Sharpe ratio of 1.67, and a maximum drawdown of -10.8%, while maintaining lower policy oscillation than competing RL baselines. The key analytical contribution is that actor–critic learning performs best when it is treated as part of a computational infrastructure rather than as an isolated prediction engine. The paper therefore contributes to data-driven AI and computational discovery by offering a reproducible systems perspective on risk-aware capital allocation in non-stationary financial environments.
