BUILDING A MULTI-AGENT ALGORITHMIC TRADING SYSTEM BASED ON DEEP REINFORCEMENT LEARNING
Abstract
The paper addresses the problem of creating an adaptive algorithmic trading system capable of fusing heterogeneous analytical signals and managing risks. A multi-agent architecture is proposed that includes agents for fundamental, technical and sentiment analysis, a price forecasting agent, trading agents with different risk profiles, a risk management agent and a deciding agent. Agents are trained using deep reinforcement learning algorithms (PPO, A2C, DDPG) implemented with Ray RLlib. The formal model is cast as a Markov decision process with decentralized execution and centralized training. Experimental evaluation on historical data of the Dow Jones 30 index over 2010-2020 demonstrates that the multi-agent system outperforms single-agent analogues: cumulative return increased by 47%, the Sharpe ratio rose from 1.52 to 1.78, and the maximum drawdown during crisis periods decreased by 12-18%. The proposed approach shows high robustness and can serve as a foundation for industrial algorithmic trading solutions.