FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making

In one sentence

Summary

The paper addresses the challenge of sequential financial decision-making, where LLMs struggle with volatile environments, multi-source information synthesis, and long-term risk management. Existing agent systems often lack hierarchical coordination or effective experience refinement, leading to suboptimal returns and high communication costs. FinCon positions itself as a synthesized framework inspired by real-world investment firms, aiming to enhance decision quality through structured collaboration and continuous learning. FinCon employs a manager-analyst hierarchy where specialized analyst agents process distinct data modalities (news, filings, audio, tabular data) to distill insights, which the manager agent consolidates for trading decisions. The system utilizes a dual-level risk-control mechanism: within-episode risk control uses Conditional Value at Risk (CVaR) for immediate risk alerts, while over-episode control employs Conceptual Verbal Reinforcement (CVRF). CVRF updates investment beliefs via textual gradient descent based on episodic performance, selectively propagating insights to relevant agents. Experiments cover single-stock trading across eight stocks and portfolio management for two portfolios, comparing FinCon against DRL agents (A2C, PPO, DQN) and LLM agents (FinGPT, FinMem, FinAgent) using metrics like Cumulative Return, Sharpe Ratio, and Max Drawdown. FinCon significantly outperforms baselines in both tasks, achieving higher returns and better risk-adjusted performance, particularly in volatile markets. Ablation studies confirm the efficacy of both risk-control components. The system demonstrates strong generalization and robustness, though it faces challenges with hallucination in complex multi-asset contexts. Limitations include reliance on specific data availability and the computational cost of multi-agent interactions, though the hierarchical design mitigates this compared to flat structures. The work highlights the potential of structured LLM agents with verbal reinforcement for financial applications.

Research problem

Sequential financial decision-making remains challenging for LLMs due to market volatility, the need for multi-source information synthesis, and effective long-term risk management. Existing LLM-based agent systems often suffer from inefficient communication, lack of hierarchical coordination, and insufficient mechanisms for updating investment beliefs based on past experiences. The paper aims to address these gaps by proposing a framework that mimics human investment firm structures and incorporates continuous learning through verbal reinforcement.

Core contributions

• Proposes FinCon, a hierarchical multi-agent framework with a manager-analyst structure for efficient information synthesis and decision-making.
• Introduces a dual-level risk-control mechanism combining within-episode CVaR monitoring and over-episode Conceptual Verbal Reinforcement (CVRF).
• Develops a verbal reinforcement method using textual gradient descent to update investment beliefs based on episodic performance.
• Demonstrates generalization of LLM agents to portfolio management, a task less explored in previous works.
• Shows significant performance improvements over DRL and other LLM agents in single-stock trading and portfolio management.

Method

• Manager-Analyst hierarchy: Analyst agents process specific data modalities; Manager agent consolidates insights and makes decisions.
• Specialized analyst agents for news, filings, audio (Whisper API), and tabular data analysis.
• Within-episode risk control: Triggers risk-averse stance when CVaR drops below a threshold.
• Over-episode risk control: Uses CVRF to compare episodic performance and update investment beliefs.
• Textual gradient descent: Updates manager prompts based on overlapping percentage of trading decisions between episodes.
• Selective belief propagation: Updates only relevant analyst agents based on the source of the insight.
• Memory modules: Working, procedural, and episodic memory with decay rates based on data timeliness.
• Portfolio optimization: Uses convex optimization for portfolio weights constrained by trading actions.

Results & findings

• FinCon achieves superior Cumulative Returns and Sharpe Ratios compared to DRL and LLM baselines in single-stock trading.
• In portfolio management, FinCon outperforms Markowitz MV, FinRL-A2C, and Equal-Weighted ETF strategies.
• FinCon shows robust performance in high-volatility market conditions (VIX > 20), where many baselines fail.
• The within-episode CVaR risk control effectively mitigates short-term losses during market downturns.
• Over-episode CVRF significantly improves long-term performance by updating investment beliefs based on past episodes.
• FinCon reduces unnecessary peer-to-peer communication costs through selective belief propagation.
• DRL agents struggle with convergence on recently listed IPOs like COIN, whereas FinCon handles them effectively.
• Ablation studies confirm that removing either risk-control component leads to performance degradation.

Data & experiments

• Uses multi-modal data including stock prices, daily news, 10-Q/10-K filings, and earnings call audio from Jan 2022 to Jun 2023.
• Evaluates on single-stock trading for eight stocks (TSLA, AMZN, NIO, MSFT, AAPL, GOOG, NFLX, COIN).
• Evaluates on portfolio management for two portfolios (TSLA/MSFT/PFE and AMZN/GM/LLY).
• Compares against DRL agents (A2C, PPO, DQN) and LLM agents (FinGPT, FinMem, FinAgent, Generative Agent).
• Metrics include Cumulative Return, Sharpe Ratio, and Max Drawdown, with statistical significance testing via Wilcoxon signed-rank test.
• Conducts ablation studies to isolate the impact of within-episode and over-episode risk control mechanisms.

Limitations & caveats

• Relies on the availability of multi-modal data, which may not be accessible for all assets.
• Computational cost of multi-agent interactions, though mitigated by hierarchical design.
• Potential for hallucination in complex multi-asset portfolio management tasks.
• Performance may vary depending on the quality of the underlying LLM (GPT-4-Turbo used).
• Training period is limited to a specific timeframe, which may affect generalization to other market regimes.
• The framework assumes a certain level of data granularity and timeliness that may not be present in all real-world scenarios.

Curator notes

Discovery status: round2_promoted. Source links: A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist; Dólares or Dollars? Unraveling the Bilingual Prowess of Financial LLMs Between Spanish and English; Learning to Trade Like an Expert: Cognitive Fine-Tuning for Stable Financial Reasoning in Language Models; LiveTradeBench: Seeking Real-World Alpha with Large Language Models; No Language is an Island: Unifying Chinese and English in Financial Large Language Models, Instruction Data, and Benchmarks; O