Architecture and Evaluation of a Multi-Agent System for Market Monitoring

Abstract:

The contemporary financial environment is characterized by a rapidly increasing volume of heterogeneous data, which necessitates automated solutions for their integration and analysis in the context of market dynamics. In this paper, a multi-agent system for automated monitoring and analysis of capital markets is presented, developed using the CrewAI framework and combining structured financial data with sentiment analysis of financial news. The proposed system integrates numerical data with textual content from relevant news sources, enabling the identification of information with a potential impact on stock price movements. The system architecture comprises two specialized agents: MarketScanner, responsible for analyzing market indicators and generating concise analytical reports, and NewsAggregator, which collects relevant news and performs sentiment analysis using large language models and the FinBERT model for financial text processing. The system was evaluated using data from Apple, Bank of America, and Exxon Mobil. Experimental results indicate that the integration of a multi-agent architecture with sentiment analysis supports systematic examination of market-relevant information. Sentiment classification accuracy reached above 80% in selected cases when evaluated against expert annotations. The findings suggest that sentiment–market relationships vary across companies and depend on the specific market context.