Emulating fundamental analysts: Analytical stage-based multi-agent framework enhanced with expert guidance and Preference-Anchored Likelihood Adjustment

With the rapid advancement of large language models (LLMs), some studies have explored their potential for predicting stock prices based on financial texts. However, previous research often overlooked the depth of analysis generated by LLMs, resulting in reasoning processes inferior to those of human analysts. In fundamental investing, which requires in-depth company analysis, conclusions from imperfect reasoning lack persuasiveness. In this study, inspired by the analysis process of human analysts, we propose an “Analytical Stage-Based Multi-Agent Framework” to enable LLMs to perform in-depth fundamental analysis. This framework divides the analysis into multiple stages, assigning an LLM agent to each. We enhance each agent’s capabilities for its specific task through expert guidance or fine-tuning, allowing them to collectively emulate the workflow of human analysts. Furthermore, we introduce Preference-Anchored Likelihood Adjustment, a new method for fine-tuning LLMs. This approach addresses the decline in likelihood of generating correct responses that occurs after using existing preference alignment methods. It employs an objective function with two terms: one to increase likelihood and another to preserve aligned preference. We conducted experiments using our framework to analyze company earnings releases. We evaluated the analysis quality based on comprehensiveness and logical soundness, while correctness was assessed by using stock prices as the ground truth to calculate the Matthews correlation coefficient and F1 score. Results demonstrate that even without expert guidance and fine-tuning, our multi-agent framework can enhance LLMs in both analysis quality and correctness. When combined with expert guidance and fine-tuning, the performance is further improved.