Reinforcement Learning-Driven Dynamic Trading Strategies for Financial Markets

Xi-Jing Ou, Jie Huang

Abstract

This paper presents the RL-DynTrade framework by using a cutting-edge deep reinforcement learning method, Proximal Policy Optimization (PPO), with a Deep Q Network (DQN) agent to dynamically adapt to changing risk-reward dynamics. PPO enables real-time, fine-grained, risk-reward adaptation via an actor-critic design with clipped surrogate objectives, enabling stable, efficient continuous portfolio allocation. DQN offers a more interpretable, computationally lightweight trading policy by using a discrete action space, Q-learning, and deep neural networks to learn optimal buy/sell decisions from historical price and volume trends. RL-DynTrade can be used for dynamic algorithmic trading, as it leverages a simple yet well-known method (DQN) with cutting-edge technology (PPO). Results from experiments using stock and index data show that RL-DynTrade outperforms both traditional approaches and RL models that use PPO or DQN alone in terms of stability, flexibility, and profitability.

Source: semanticscholar · PDF

Read the AI summary and key takeaways for traders on WOBR Quant Research.