Dynamic Portfolio Management through Deep Q-Network-Based Asset Allocation

Abstract

This study proposes an intelligent asset allocation optimization method based on Deep Q-Network (DQN) to improve the return of the investment portfolio and reduce the risk. In the financial market, the core challenge of asset allocation is to cope with the highly nonlinear and dynamic changes of the market, and traditional asset allocation methods, such as the mean-variance model and the capital asset pricing model (CAPM), are often difficult to adapt to the complex market environment. In contrast, reinforcement learning methods can continuously optimize investment strategies through interactive learning with the market. This paper adopts DQN as the core algorithm, designs a reinforcement learning framework based on market state variables, and introduces technical indicators and macroeconomic factors as state inputs to enhance the model's perception of market dynamics. The experimental part uses the MSCI World Index and its constituent stock datasets to compare the performance of different asset allocation strategies and analyze the impact of different state variables on the DQN training effect. The experimental results show that the DQN asset allocation strategy is superior to traditional methods in key indicators such as annualized return, Sharpe ratio and maximum drawdown, especially after combining multiple market information, it shows stronger robustness and adaptability. This study provides a new approach for intelligent asset management and lays the foundation for the future application of reinforcement learning in financial markets.