Source-linked AI summary
FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making
Yangyang Yu, Zhiyuan Yao, Haohang Li, Zhiyang Deng, Yupeng Cao, Zhi Chen, Jordan W. Suchow, Rong Liu, Zhenyu Cui, Zhaozhuo Xu, Denghui Zhang, Koduvayur Subbalakshmi, Guojun Xiong, Yueru He, Jimin Huang, Dong Li, Qianqian Xie
TL;DR
Sequential financial investment decisions require integrating diverse information while managing risk across repeated interactions with volatile markets. FINCON addresses this gap with a hierarchical LLM multi-agent system and dual-level verbal risk control, and it reports strong performance across stock trading and portfolio management, while multi-asset contexts remain challenging.
Problem
High-quality sequential investment decisions require multi-source information synthesis and risk management, but existing financial language-agent systems have limited long-term risk control and portfolio-management coverage.
Method
FINCON combines a synthesized manager-analyst hierarchy with within-episode CVaR monitoring and over-episode verbal reinforcement that updates investment beliefs.
Results
FINCON consistently exhibits superior decision-making quality across bullish, bearish, and mixed market conditions, while achieving low MDD across most trading assets and the highest investment returns.
Takeaways & Limitations
The framework generalizes across single-stock trading and portfolio management while reducing redundant communication through selective propagation of updated investment beliefs.
Takeaways & Limitations
Multi-asset portfolio management increases hallucination risk because longer, more complex inputs can produce incorrect memory-event references.
Abstract
from arXiv · showhide
Large language models (LLMs) have demonstrated notable potential in conducting complex tasks and are increasingly utilized in various financial applications. However, high-quality sequential financial investment decision-making remains challenging. These tasks require multiple interactions with a volatile environment for every decision, demanding sufficient intelligence to maximize returns and manage risks. Although LLMs have been used to develop agent systems that surpass human teams and yield impressive investment returns, opportunities to enhance multi-sourced information synthesis and optimize decision-making outcomes through timely experience refinement remain unexplored. Here, we introduce the FinCon, an LLM-based multi-agent framework with CONceptual verbal reinforcement tailored for diverse FINancial tasks. Inspired by effective real-world investment firm organizational structures, FinCon utilizes a manager-analyst communication hierarchy. This structure allows for synchronized cross-functional agent collaboration towards unified goals through natural language interactions and equips each agent with greater memory capacity than humans. Additionally, a risk-control component in FinCon enhances decision quality by episodically initiating a self-critiquing mechanism to update systematic investment beliefs. The conceptualized beliefs serve as verbal reinforcement for the future agent's behavior and can be selectively propagated to the appropriate node that requires knowledge updates. This feature significantly improves performance while reducing unnecessary peer-to-peer communication costs. Moreover, FinCon demonstrates strong generalization capabilities in various financial tasks, including single stock trading and portfolio management.
1 Introduction
Sequential financial decisions require integrating diverse, time-varying information across repeated market interactions while balancing returns and risk. FINCON addresses this challenge with a hierarchical multi-agent framework and risk-controlled verbal reinforcement for stock trading and portfolio management.
- Financial investment tasks require repeated market interactions and integration of diverse information streams to maximize profit while managing risk.
- Human investment firms use hierarchical functional roles to synthesize resources, but human cognitive limitations can hinder rapid market-signal processing.
- LLM-based language agents offer flexible communication and prompt-driven structures for reasoning, planning, decision-making, and financial applications.
- Effective financial agent systems must organize teamwork and communication while continuously learning and refining actions.
- Existing financial language-agent systems face challenges in controlling long-term risk and extending beyond single-asset trading.
- FINCON introduces a manager-analyst hierarchy with risk control, extends financial language agents to portfolio management, and updates investment beliefs through CVaR and verbal reinforcement.
2 Preliminaries
FINCON formulates financial decision-making as a risk-sensitive sequential problem and assigns distinct information-processing and trading roles to analyst and manager agents. Its portfolio workflow combines agent decisions with mean-variance optimization, while textual prompt updates optimize the multi-agent policy.
- FINCON processes multimodal market information through analyst agents, then uses a manager to make trades and risk-control prompts to optimize decisions.
- For portfolio tasks, analysts construct a stock pool using return correlations and the manager determines portfolio weights with an external mean-variance optimizer.
- Quantitative trading is modeled as an infinite-horizon POMDP with observable and unobservable market components, sequential actions, rewards, and reflections.
- Analyst agents process separate textual market-information streams, while the manager chooses buy, sell, or hold actions for single-stock trading and weighted actions for portfolios.
- The system learns analyst and manager policies to maximize cumulative trading reward while controlling risk, updating prompts through verbal reinforcement.
- Unlike DRL methods designed for POMDPs, FINCON uses a risk-sensitive optimization problem based on textual gradient descent.
3 Architecture of FINCON
FINCON coordinates specialized analysts and a sole manager through a two-level hierarchy, then applies within-episode risk monitoring and over-episode belief updates. Its agents combine role-specific processing with working, procedural, and episodic memory to refine decisions while limiting communication overhead.
- Synthesized Multi-agent Hierarchical Structure Design: The hierarchical design organizes synchronous agent communication to improve information presentation and comprehension while minimizing unnecessary communication costs.
- Architecture of FINCON: FINCON has two main components: a Manager-Analyst Agent Group and a Risk-Control component.
- Synthesized Multi-agent Hierarchical Structure Design: Analysts distill large, multimodal market data into specialized insights, while the manager consolidates those insights and serves as the sole sequential decision-maker.
- Risk Control: FINCON uses dual-level risk control: within-episode monitoring adjusts actions during market fluctuations, while over-episode updates refine investment beliefs across training episodes.
- Within-Episode Risk Control: A CVaR decline triggers a risk-averse trading stance because it indicates recent PnLs entering the worst-performing 1% of daily outcomes.
- Modular Design of FINCON Agents: Agents integrate working, procedural, and episodic memory to observe, refine, and record historical actions, outcomes, and reflections.
4 Experiments
FINCON is evaluated across single-asset trading and portfolio management, with experiments examining robustness, risk control, and belief updates. It generally outperforms comparison strategies while reducing risk, although multi-asset management remains susceptible to occasional hallucinations.
- Single-Asset Trading: FINCON outperforms LLM-based and DRL-based approaches on cumulative returns and Sharpe ratios across eight single-stock trading tasks.It also achieves one of the lowest maximum drawdowns across most trading assets.
- Single-Asset Trading: 57% cumulative return and 0.825 Sharpe ratio are achieved by FINCON on Coinbase, where limited data causes DRL algorithms to struggle with convergence.The comparison focuses on FINCON, LLM-based agents, and buy-and-hold because Coinbase has limited trading data.
- Single-Asset Trading: FINCON maintains superior decision-making quality across bullish, bearish, and mixed market conditions compared with other LLM-based agents.The paper attributes this pattern to multi-agent information synthesis and dual-level risk control.
- Portfolio Management: FINCON outperforms Markowitz mean-variance, FINRL, and Equal-Weighted ETF strategies on cumulative return, Sharpe ratio, and maximum drawdown in two portfolio tasks.The portfolios contain TSLA, MSFT, and PFE, or AMZN, GM, and LLY.
- Portfolio Management: Multi-asset management increases hallucination risk because longer, more complex inputs occasionally produce incorrect memory-event indices.Specialized agents mitigate this issue, but the limitation remains in relatively compact portfolios.
5 Conclusion
FINCON combines a hierarchical manager-analyst communication structure with dual-level risk control for financial decision-making across single-stock trading and portfolio management. The framework dynamically updates risk and market beliefs while reducing communication overhead.
- 5 Conclusion: FINCON routes financial data to specialized analyst agents, whose distilled insights are synthesized by a manager agent for trading decisions.This implements the Synthesized Manager-Analyst hierarchical communication structure.
- 5 Conclusion: The framework uses dual-level risk control to mitigate investment risks and enhance trading performance.Its risk-control component supports dynamic updates of risk and market beliefs within agent communication.
- 5 Conclusion: FINCON’s streamlined communication structure reduces overhead while supporting financial decision-making in single-stock trading and portfolio management.
- 5 Conclusion: Scaling FINCON to portfolios comprising tens of assets remains a future direction constrained by LLM input-length limits.
A.1 Related Work
Related work develops LLM agents for sequential financial decisions, flexible multi-agent communication, and prompt-based reasoning refinement. These directions motivate more capable language-agent systems for complex financial tasks.
- A.1 Related Work: LLM agents are increasingly studied for sequential decision-making involving episodic environment interactions and verbal reflections for action refinement.
- A.1 Related Work: Traditional multi-agent systems use predetermined communication, whereas LLM-based systems enable human-understandable discussions and debates.
- A.1 Related Work: Prompt optimization methods such as ReAct, Chain of Thought, Tree of Thoughts, and ART generate intermediate reasoning steps for LLM agents.
A.2 Textual Gradient-Descent
Textual gradient-descent methods use LLM reflections on trading successes and failures to guide prompt updates. Overlapping decision sequences help regulate these updates toward stable, incremental improvement.
- A.2 Textual Gradient-Descent: A meta-prompt refines a task prompt by instructing an LLM to improve its performance on the target task.
- A.2 Textual Gradient-Descent: LLM reflections on past trading successes and failures provide semantic gradient signals for prompt optimization.
- A.2 Textual Gradient-Descent: Uncontrolled prompt updates can overshoot or oscillate, analogous to instability from an inappropriate learning rate.
- A.2 Textual Gradient-Descent: The method measures overlap between consecutive trading decision sequences and directly edits the previous task prompt for stable, incremental improvement.
A.3 FINCON Testing Stage Workflow
During testing, FINCON reuses learned investment beliefs, prompts, and policy while disabling over-episode risk control. Within-episode risk control remains active to adjust trading actions in response to short-term performance and market fluctuations.
- A.3 FINCON Testing Stage Workflow: Testing inherits the manager-analyst components, reflections, learned prompts, and trained policy from the training stage.
- A.3 FINCON Testing Stage Workflow: The over-episode risk-control mechanism is disabled during testing, while within-episode risk control continues operating.
- A.3 FINCON Testing Stage Workflow: The policy collects daily profit-and-loss, portfolio weights, and daily CVaR values throughout the testing loop.
- A.3 FINCON Testing Stage Workflow: Manager-agent self-reflection is triggered when daily CVaR decreases or daily profit-and-loss is negative.
- A.3 FINCON Testing Stage Workflow: The workflow collects an investment trajectory and calculates performance metrics from it after the testing period.
A.4 Figure of Modular Design of Agents in FINCON
FINCON organizes manager and analyst agents through role-specific prompts, memory, and communication modules. Analysts process specialized information, while the manager consolidates insights, records experience, and makes trading decisions.
- Analyst agents receive role-specific duties to distill insights from news, market data, filings, and earnings-call audio.
- The manager perceives analyst insights and risk-control updates, sends feedback, and consolidates information into trading actions.
- Manager and analyst agents maintain distinct working, procedural, and episodic memories for observations, distilled insights, actions, and trajectories.
- The configuration and profiling modules generate text queries for retrieving investment information from agent memory databases.
- Perceptual and memory modules use prompts with LLMs to extract key investment insights from retrieved information.
A.5 Experimental Setup
FINCON is evaluated in a multimodal market setting against LLM-based, DRL-based, and classical baselines. The experiments use financial data, standardized profitability and risk metrics, and separate training and testing periods.
- The dataset combines stock prices, daily news, Form 10-K and 10-Q filings, and earnings-call audio from January 3, 2022, to June 10, 2023.
- Evaluation uses Cumulative Return, Sharpe Ratio, and Max Drawdown to measure profitability, risk-adjusted returns, and risk management.
- Single-stock comparisons include seven algorithmic agents and Buy-and-Hold, while portfolio comparisons include Markowitz and reinforcement-learning baselines.
- FINCON uses GPT-4-Turbo at temperature 0.3, trains from January 3, 2022, to October 4, 2022, and tests from October 5, 2022, to June 10, 2023.
A.6 Single Stock Trading Result Graphs
The reported single-stock and portfolio experiments show strong FINCON performance and attribute gains to its two-level risk-control design. Within-episode CVaR responds to market risk, while over-episode belief updates improve decisions through iterative reflection.
- CVaR risk control improved FINCON across financial metrics in both single-asset trading and portfolio management.
- Without within-episode risk control, single-asset trading produced negative returns and higher drawdowns, including GOOG at 22.42% and NIO at −77.210% for Buy-and-Hold comparison.
- Portfolio Cumulative Return increased from 14.699% to 113.836% with within-episode risk control.
- CVRF belief updates enriched conceptual factors such as historical momentum and news insights with executable information over four training episodes.
- Trading-action overlap exceeded 80% between the last two adjacent episodes, while updated beliefs were mostly aligned.
- With belief updates, portfolio Cumulative Return reached 113.836% and Sharpe Ratio reached 3.269, compared with 28.432% and 1.181 without them.
- The experiments evaluate returns, risk, and portfolio value using multimodal financial data, risk-estimation formulas, and performance metrics.
- FINCON achieved positive Cumulative Returns and Sharpe Ratios in single-stock trading and the highest primary metrics in portfolio management.