MULTI-AGENT DEEP REINFORCEMENT LEARNING APPROACH FOR ALGORITHMIC STOCK TRADING BY VARYING RISK-AVERSION

Abstract

In the rapidly evolving domain of finance and technology, Machine Learning has become a critical tool for gaining a competitive edge in the realm of algorithmic trading. This work presents a novel approach to algorithmic stock trading by deploying Multi-Agent Deep Reinforcement Learning (MADRL) that incorporates varying levels of risk aversion. We develop algorithmic equity trading strategies that hedge financial risk while outdoing the market, utilizing real historical data of Dow 30 US stock indices for training and testing. We construct an environment where agents, reflecting diverse investor profiles with varying risk aversions, interact and execute trades based on technical indicators and risk management strategies such as Conditional Value-at-Risk (CVaR). We explore the performance of actor-critic DRL algorithms - Proximal Policy Optimization (PPO), Deep Deterministic Policy Gradient (DDPG), and Advantage Actor-Critic (A2C). The work builds as a comparative analysis of DRL algorithmic trading against passive trading strategies, offering insights into the effects of risk aversion in algorithmic trading and its real world applications. The findings suggest that our fine-tuned DRL environment can outperform the benchmark and MADRL framework can hedge against market risk during economic downturns to limit excessive losses.