Abstract: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.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to utilize large - language models (LLMs) and multi - agent systems to enhance the ability to make high - quality and continuous financial investment decisions in financial investment decision - making. Specifically, the authors propose improvement plans for several limitations of existing systems: 1. **Inadequate risk management**: Existing systems rely on risk preferences for short - term market fluctuations and cannot effectively control the overall risk exposure during long - term trading periods. This may lead to the neglect of fundamental factors driving investment profits, resulting in potential losses. 2. **Limited adaptability**: Most existing systems are only applicable to single - asset trading tasks and show limited adaptability in financial applications involving multiple assets (such as portfolio management). 3. **High information - processing pressure**: These systems rely on the information - understanding and - extraction capabilities of a single agent, which puts heavy pressure on the language model, limits its ability to process information within a limited context window, and thus affects the quality of decision - making. 4. **High communication costs**: Although some multi - agent systems attempt to make decisions through extensive discussions, this method results in high communication costs and extended processing times, and lacks clear optimization goals, reducing the decision - making effect. To solve the above problems, the authors propose FINCON, an LLM - based multi - agent framework aimed at enhancing financial decision - making through conceptual verbal reinforcement. The main contributions of FINCON include: - Designing a synthetic manager - analyst hierarchical communication structure inspired by the structure of real - world investment teams, as well as a risk - control component. This structure allows financial data from different sources and modalities to be assigned to corresponding functional analyst agents, enabling them to focus on distilling key investment insights and indicators from a single information source. Then, the manager agent integrates these insights and interacts with the risk - control component to make trading decisions. - Proposing a general framework that can not only conduct stock trading but also portfolio management. The latter is a task that other financial language - agent frameworks have not yet claimed to be able to solve. - Innovatively constructing the personality characteristics of agents through a two - layer risk - control mechanism to achieve risk updates across cycles. Within each cycle, the quantile risk measure in quantitative finance - conditional value - at - risk (CVaR) - is used for risk supervision; in cross - cycle risk control, a unique verbal - reinforcement method is designed to update the investment beliefs of decision - makers through conceptualizing investment insights to guide future decisions. These investment beliefs are summarized from the functions of all analyst agents and can be selectively back - propagated to the corresponding functional analyst agents as needed. Through the above improvements, FINCON significantly improves the quality of financial decision - making while reducing unnecessary point - to - point communication costs.

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

FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design

FinVision: A Multi-Agent Framework for Stock Market Prediction

FinRobot: An Open-Source AI Agent Platform for Financial Applications using Large Language Models

A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist

INVESTORBENCH: A Benchmark for Financial Decision-Making Tasks with LLM-based Agent

Data-Centric Financial Large Language Models

FinVerse: An Autonomous Agent System for Versatile Financial Analysis

SNFinLLM: Systematic and Nuanced Financial Domain Adaptation of Chinese Large Language Models

TradingAgents: Multi-Agents LLM Financial Trading Framework

Designing Heterogeneous LLM Agents for Financial Sentiment Analysis

Open-FinLLMs: Open Multimodal Large Language Models for Financial Applications

A Data-Centric Approach for Financial Large Language Models with Abductive Augmentation Reasoning

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

LLM-Powered Multi-Agent System for Automated Crypto Portfolio Management

SilverSight: A Multi-Task Chinese Financial Large Language Model Based on Adaptive Semantic Space Learning

The FinBen: an Holistic Financial Benchmark for Large Language Models

FinLlama: Financial Sentiment Classification for Algorithmic Trading Applications

Enhancing Financial Question Answering with a Multi-Agent Reflection Framework

Optima: Optimizing Effectiveness and Efficiency for LLM-Based Multi-Agent System