A Deep Reinforcement Learning and Predictive Architecture for Stock Portfolio Management

Abstract

This paper presents a deep reinforcement learning framework for stock portfolio management that integrates timeseries forecasting with advanced graph representations. We employ a DeepAR module to provide predictive signals on future price movements and a Temporal Portfolio Graph (TPG) to capture inter-asset correlations. These enriched features are fed into a Proximal Policy Optimization (PPO) agent, enabling robust portfolio reallocation across diverse market conditions. Experimental evaluations from 2012 to 2024 demonstrate that our approach outperforms vanilla PPO and traditional market benchmarks, delivering higher returns and favorable riskadjusted performance. The results underscore the effectiveness of combining predictive modeling and graph-based state representations for more informed, adaptable trading strategies.