Portfolio Optimization through a Multi-modal Deep Reinforcement Learning Framework
Abstract
Wong JiaJie and LIU LiLi*
In today’s increasingly complex and volatile stock markets, leveraging advanced machine learning and quantitative techniques is becoming indispensable for enhancing trading strategies and optimizing returns. This study introduces a so- phisticated Multi- modal framework that combines Deep Rein- forcement Learning (DRL) with Algorithmic Trading Signals and Price Forecasts to improve risk-adjusted returns in equity trad- ing. Utilizing the Proximal Policy Optimization (PPO) algorithm within a custom trading environment built on the FinRL library, our approach integrates advanced algorithmic signals−such as moving average crossovers and oscillator divergence−and incorporates enriched price forecasts from Long Short-Term Memory (LSTM) networks. The proposed framework was rigorously evaluated using a diverse set of 29 out of 30 constituent stocks within the Dow Jones Industrial Average (DJI). The empirical results highlight the effectiveness of the Multi-modal DRL approach, demonstrating significant outperformance over traditional benchmarks, with an annualized return of 16.24%, an annualized standard deviation of 17.49%, a Sharpe Ratio of 0.86, and a Sortino Ratio of 1.27. These findings underscore the potential of Multi-modal DRL frameworks to offer consistent, robust performance and contribute to advancing trading strategies in dynamic market environments.