Dynamic Optimization Strategy of Financial Portfolios Using Deep Reinforcement Learning (DRL)-Based Neural Networks

In recent years, dynamic portfolio optimization has grown critical in finance, as traditional static models fail to adapt to complex market fluctuations—this is the key scientific problem. Deep reinforcement learning (DRL; learning optimal actions using neural networks and reward feedback) offers strong adaptability for dynamic allocation. Therefore, this study proposes a multi-asset DRL-neural network optimization framework that enables real-time market response via state perception, action selection, and reward feedback. Using authoritative data (e.g., Wind, Yahoo Finance) covering major indices or exchange-traded funds (ETFs; a fund that tracks a basket of assets and trades like stocks), the method tests the model across market cycles (e.g., Hong Kong and United States stocks); analyzes return, Sharpe ratio, and maximum drawdown; and compares performance with the mean-variance and Black-Litterman models. Results show the model outperforms traditional methods in risk control, return, and robustness. The study concludes that DRL enhances return stability, supporting intelligent financial decisions. Future research will integrate transaction cost models to boost real-world use.