Abstract
Large language models (LLMs) have demonstrated notable potential inconducting complex tasks and are increasingly utilized in various financialapplications. However, high-quality sequential financial investmentdecision-making remains challenging. These tasks require multiple interactionswith a volatile environment for every decision, demanding sufficientintelligence to maximize returns and manage risks. Although LLMs have been usedto develop agent systems that surpass human teams and yield impressiveinvestment returns, opportunities to enhance multi-sourced informationsynthesis and optimize decision-making outcomes through timely experiencerefinement remain unexplored. Here, we introduce the FinCon, an LLM-basedmulti-agent framework with CONceptual verbal reinforcement tailored for diverseFINancial tasks. Inspired by effective real-world investment firmorganizational structures, FinCon utilizes a manager-analyst communicationhierarchy. This structure allows for synchronized cross-functional agentcollaboration towards unified goals through natural language interactions andequips each agent with greater memory capacity than humans. Additionally, arisk-control component in FinCon enhances decision quality by episodicallyinitiating a self-critiquing mechanism to update systematic investment beliefs.The conceptualized beliefs serve as verbal reinforcement for the future agent'sbehavior and can be selectively propagated to the appropriate node thatrequires knowledge updates. This feature significantly improves performancewhile reducing unnecessary peer-to-peer communication costs. Moreover, FinCondemonstrates strong generalization capabilities in various financial tasks,including single stock trading and portfolio management.