FinArena: A Human-Agent Collaboration Framework for Financial Market Analysis and Forecasting

In one sentence

Summary

This paper addresses the limitations of traditional financial forecasting models and standalone LLMs, which often struggle with multimodal data integration, hallucinations in unstructured news, and the lack of personalized risk consideration. The authors propose FinArena, a Human-Agent collaboration framework inspired by the Mixture of Experts (MoE) approach. The system comprises three specialized agents: a Time Series Agent for historical price trends, a News Agent utilizing adaptive Retrieval-Augmented Generation (RAG) to mitigate hallucinations by dynamically deciding when to search external sources, and a Statement Agent that performs iterative reasoning on financial reports. A universal expert agent synthesizes these inputs alongside user-defined risk preferences to generate personalized investment recommendations, bridging the gap between automated analysis and human intuition. The experimental evaluation utilizes a bespoke, small-scale multimodal dataset comprising stock prices, news articles, and financial statements for five U.S. companies (Amazon, Google, Microsoft, Nvidia, Tesla) and five Chinese A-share companies (BYD, CATL, etc.) from January 2023 to March 2024. This dataset was constructed to reflect information accessible to retail investors, avoiding the data leakage issues common in large-scale public datasets. The study compares FinArena against traditional statistical models (ARIMA, GARCH) and machine learning baselines (LSTM, XGBoost) using stock trend prediction accuracy and trading simulation metrics. The adaptive RAG mechanism is specifically tested for its ability to reduce irrelevant responses and improve news analysis efficiency by filtering out low-uncertainty queries. Results indicate that FinArena surpasses both traditional and state-of-the-art benchmarks in stock trend prediction across both U.S. and Chinese markets. The framework demonstrates promising performance in trading simulations, effectively aligning strategic insights with individual investor risk profiles, ranging from conservative to aggressive. The study highlights the practical utility of the system for retail investors by providing a low-cost, interpretable decision-support tool. However, limitations include the small sample size of selected stocks, potential biases in web-crawled news data, and the reliance on specific LLM capabilities that may vary with model updates. The paper concludes that human-agent collaboration enhances investment outcomes by incorporating nuanced human risk preferences that purely algorithmic systems often overlook.

Research problem

The paper addresses three key challenges in financial LLM applications: (1) LLMs struggle with uncertainty and hallucinations when processing dynamic, unstructured news data not present in pre-training corpora. (2) Traditional models fail to effectively integrate diverse multimodal data sources (time series, text, tabular financial statements) required for comprehensive market analysis. (3) Existing frameworks often ignore human-in-the-loop interaction, specifically the need to incorporate individual investor risk preferences into automated decision-making, leading to generic strategies that may not suit all users.

Core contributions

• Proposes FinArena, a novel Human-Agent collaboration framework inspired by Mixture of Experts (MoE) for multimodal financial analysis.
• Introduces an uncertainty-driven adaptive RAG mechanism for the News Agent to dynamically balance pre-trained knowledge and external retrieval, reducing hallucinations and API costs.
• Designs an iterative reasoning Statement Agent that mimics human step-by-step financial report analysis, enhancing interpretability and confidence estimation.
• Constructs and open-sources a small-scale, retail-investor-accessible multimodal dataset covering U.S. and Chinese A-share markets to address data leakage and cost barriers in existing benchmarks.
• Demonstrates superior stock trend prediction and trading simulation performance compared to traditional statistical and machine learning baselines.

Method

• Human Module: Interactive interface capturing individual risk preferences (e.g., conservative, aggressive).
• Time Series Agent: LLM-based forecasting using historical price and volume data.
• News Agent: Uncertainty-driven adaptive RAG; Judge Module decides if external search is needed.
• Statement Agent: Iterative reasoning pipeline analyzing seasonal patterns and annual operations from financial statements.
• Aggregation System: Universal expert agent synthesizes outputs from specialized agents.
• Human-Agent Collaboration: Report Agent allows human feedback and risk preference integration before final decision.
• Data Preprocessing: LLM self-reflection for noise removal, k-means for news clustering.

Results & findings

• FinArena outperforms traditional models (ARIMA, GARCH) and ML baselines (LSTM, XGBoost) in stock trend prediction.
• The adaptive RAG mechanism effectively reduces hallucinations in news analysis by filtering low-uncertainty queries.
• Trading simulations show promising results across various investor risk profiles, validating the human-agent collaboration aspect.
• The framework successfully integrates time series, news sentiment, and fundamental analysis into a unified prediction.
• Performance is consistent across both U.S. and Chinese markets, though specific dynamics differ.
• The iterative reasoning agent provides confidence levels for financial statement analysis, enhancing transparency.
• The small-scale dataset approach avoids data leakage from LLM pre-training, offering a more realistic evaluation for retail investors.

Data & experiments

• Bespoke multimodal dataset: 10 companies (5 U.S., 5 Chinese A-share) with stock prices, crawled news, and financial statements from Jan 2023 to Mar 2024.
• Data sources: Tsanghi API for stock prices, National Business Daily (China) and Business Today (US) for news, public financial reports.
• Preprocessing: LLM-based self-reflection for news cleaning, k-means clustering for deduplication of same-day news.
• Baselines: ARIMA, GARCH, LSTM, XGBoost, and other SOTA financial forecasting models.
• Evaluation metrics: Stock trend prediction accuracy, trading simulation returns across different risk profiles.
• Markets: Comparative analysis between U.S. and Chinese A-share markets to test generalizability.

Limitations & caveats

• Small sample size: Only 10 companies were analyzed, limiting the generalizability of the findings.
• Data recency: The dataset ends in March 2024; performance in more volatile or recent market conditions is unknown.
• News source bias: Reliance on specific web sources (National Business Daily, Business Today) may introduce regional or editorial biases.
• LLM dependency: Performance is tied to the underlying LLM capabilities, which may change with model updates.
• Cost of RAG: While adaptive, the RAG mechanism still incurs API costs, which may be significant for very frequent trading.
• Interpretability vs. Accuracy: The complex multi-agent system may be harder to interpret than simpler baselines, despite the iterative reasoning design.

Curator notes

Discovery status: round2_promoted. Source links: A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist; FinBERT: A Pretrained Language Model for Financial Communications; FinBERT: Financial Sentiment Analysis with Pre-trained Language Models; InvestLM: A Large Language Model for Investment using Financial Domain Instruction Tuning; Revolutionizing Finance with LLMs: An Overview of Applications and Insights; XuanYuan 2.0: A Large Chinese Financial Chat Model with Hundreds of Billions Parameters