Enhancing Investment Analysis: Optimizing AI-Agent Collaboration in Financial Research

In one sentence

Summary

This research addresses the limitation of single-agent systems in financial analysis by proposing a configurable multi-agent collaboration framework. The study investigates how varying agent group sizes (single, dual, triple) and collaboration structures (horizontal, vertical, hybrid) impact performance across three sub-tasks: fundamental analysis, market sentiment analysis, and risk analysis. The core problem is determining the optimal architectural configuration for AI agents to maximize the quality of investment research reports and decision-making accuracy when processing complex financial documents like SEC 10-K forms. The methodology involves implementing these agent structures on the FinRobot platform using GPT-4-1106-vision-preview. Agents are equipped with Retrieval-Augmented Generation (RAG) and tool functions to access financial data, news, and social media. The experiments analyze the 2023 SEC 10-K forms of 30 Dow Jones Index companies. Evaluation metrics include sub-task quality scores (fundamentals, sentiment, risk) assessed by GPT-4, as well as AIGC quality metrics like readability and coherence. The final investment decision is evaluated by predicting one-week target stock prices and binary buy/not-buy recommendations, comparing the performance of different agent configurations against actual market outcomes. Findings reveal that single agents outperform multi-agent groups in simpler tasks like fundamental and sentiment analysis, where excessive communication introduces noise. Conversely, triple-agent groups excel in complex risk analysis, with vertical structures (leader-subordinate) providing the best results by ensuring focused synthesis and reducing opinion convergence bias. The proposed ensemble structure, which selects the optimal agent configuration for each sub-task, achieves the highest performance with a 2.35% average difference in target price prediction and 66.7% accuracy in buy/not-buy decisions, surpassing all uniform multi-agent architectures. This highlights the importance of task-specific agent design in financial AI systems.

Research problem

The paper addresses the underutilization of collaborative potential in existing single-agent AI systems for financial investment research. It seeks to determine how different multi-agent collaboration structures and group sizes affect the quality of financial analysis and investment decision-making. Specifically, it aims to identify the optimal agent architecture for various sub-tasks such as fundamental analysis, sentiment analysis, and risk assessment.

Core contributions

• Proposes a novel multi-agent collaboration framework with configurable group sizes and structures (horizontal, vertical, hybrid).
• Introduces a unified RAG and tool-calling integration for autonomous financial report analysis.
• Conducts a comprehensive empirical study on 30 Dow Jones companies using SEC 10-K forms.
• Identifies that single agents are superior for simple tasks while multi-agents excel in complex risk analysis.
• Demonstrates that an ensemble structure combining optimal sub-task agents outperforms uniform architectures.

Method

• Implemented single, dual, and triple agent groups with varying collaboration structures.
• Used nested chat mechanisms to simulate vertical leader-subordinate interactions.
• Integrated RAG as a tool function for autonomous retrieval from SEC filings.
• Defined specific system prompts to enforce communication and consensus in multi-agent groups.
• Evaluated sub-task quality using GPT-4 with specific scoring criteria for fundamentals, sentiment, and risk.
• Assessed AIGC quality through readability and coherence metrics.
• Calculated investment decision accuracy by comparing predicted target prices and buy/not-buy signals to actual outcomes.

Results & findings

• Single agents outperform multi-agent groups in fundamental and sentiment analysis tasks.
• Triple-agent groups perform best in risk analysis, with vertical structures being most effective.
• Horizontal and hybrid structures are better for simple tasks requiring diverse perspectives.
• Vertical structures prevent opinion convergence bias in complex decision-making tasks.
• Ensemble structure achieves 2.35% average difference in target price prediction.
• Ensemble structure achieves 66.7% accuracy in buy/not-buy decisions, surpassing vertical (50%).
• Multi-agent collaboration improves efficiency and adaptability in complex financial environments.
• Readability and coherence scores align with functional performance trends across structures.

Data & experiments

• Dataset: 2023 SEC 10-K forms of 30 companies listed on the Dow Jones Index.
• Tools: YFinance, Financial Modeling Prep, FinnHub API, and Python Reddit API Wrapper.
• Model: GPT-4-1106-vision-preview as the foundation for all AI agents.
• Evaluation: GPT-4 scored sub-task quality (1-5 scale) and AIGC metrics (readability, coherence).
• Metrics: Target price prediction error and binary buy/not-buy decision accuracy.
• Comparison: Single, dual, triple agents; horizontal, vertical, hybrid, and ensemble structures.

Limitations & caveats

• Study limited to 30 Dow Jones companies, potentially lacking generalizability to other markets.
• Evaluation relies on GPT-4 for scoring, which may introduce bias in quality assessment.
• Experiments conducted on historical data, not real-time market conditions.
• Does not explore extensively hierarchical multi-agent systems beyond triple agents.
• Tool functions are limited to specific APIs, potentially restricting information scope.
• Future work needed to optimize structure selection dynamically based on task requirements.

Curator notes

Discovery status: round2_promoted_seed_for_round3. Source links: InvestLM: A Large Language Model for Investment using Financial Domain Instruction Tuning; Revolutionizing Finance with LLMs: An Overview of Applications and Insights