A Multi-Agent Collaborative Investment Research Framework Integrating Large Language Models and Quantitative Factors for Financial Market Decision-Making

Financial investment decisions require comprehensive consideration of the enterprise's basic situation, market changes, and investment risks. Traditional prediction models based on Transformers mainly rely on structured data such as prices, while financial large language models, although capable of understanding and analyzing financial reports, still have deficiencies in quantitative analysis and risk control. This study designed a method consisting of four intelligent agents to conduct investment analysis by simultaneously utilizing enterprise financial information and market transaction data. Among them, the financial report analysis agent is responsible for extracting relevant information from the enterprise annual reports, the quantitative factor agent is used to analyze the characteristics of stock price changes, the risk control agent and the investment portfolio decision agent adjust the investment strategy according to market conditions. The study used the enterprise annual reports from 2021 to 2023, as well as the daily transaction data of Apple, Microsoft and NVIDIA from 2020 to 2023, and further extracted indicators such as stock price trends, price fluctuations and trading activity for analysis. Under the condition of using SPY as the market benchmark, the study conducted an out-of-sample test of the model for 2023. The experimental results show that compared with traditional Transformer models, this method has improved performance in terms of return and risk control, but still has certain gaps compared with the equal-weight strategy and the single-agent strategy of the FinGPT style. The findings from subsequent ablation experiments indicate that integrating enterprise information, market data, and risk management methods can help intelligent investment models develop a more comprehensive analysis process.