LLM-Powered Multi-Agent System for Automated Crypto Portfolio Management
Researchers have developed a multi-agent system (MAS) for automated cryptocurrency portfolio management, integrating news sentiment, market dynamics, and trading signals. This system decomposes tasks among specialized agents, utilizing hierarchical, collaborative, and debate communication architectures. In a 52-week backtest, the best configuration achieved a 133.52% cumulative return and a Sharpe ratio of 1.502, outperforming single-agent models and deep learning baselines. The study also indicated that the benefits of multi-agent coordination are model-agnostic, performing well with GPT-4o, GPT-5, and Claude Sonnet 4.5. AI
IMPACT Demonstrates the potential for LLM-based multi-agent systems to outperform single models in complex, real-time financial decision-making.