Research map / Multi-agent systems and games / LLM agents
FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FinCon, a GPT-4-Turbo-driven multi-agent framework with a manager-analyst hierarchy, CVaR-based within-episode risk control, and conceptual verbal reinforcement across episodes, evaluated on single-stock trading and portfolio management.
Problem
Sequential financial investment decision-making (single stock trading and portfolio management) under volatile markets, where existing LLM agent systems lack risk control, multi-asset support, and efficient communication/optimization.
Approach
A synthesized manager-analyst hierarchy in which specialized analyst agents process uni-modal data sources and a manager makes decisions, combined with a dual-level risk control: CVaR-triggered within-episode risk alerts and over-episode conceptual verbal reinforcement that updates investment beliefs via textual gradient descent, selectively propagated to relevant agents.
Key results
- On single-stock trading, FinCon reports CR of 82.871% and SR of 1.972 on TSLA, versus 34.624% CR for the best LLM baseline (FinMem) on TSLA.
- On portfolio management, FinCon achieves CR 113.836% and SR 3.269 on the TSLA/MSFT/PFE portfolio, versus 14.699% CR without CVaR control.
- Ablations show removing CVaR within-episode risk control drops GOOG CR from 25.077% to -1.461%, and removing belief updates drops portfolio CR from 113.836% to 28.432%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add within-episode CVaR risk control (ablation) | FinCon · GOOG | CR %: -1.461 -> 25.077 | Helped |
| add within-episode CVaR risk control (ablation) | FinCon · NIO | CR %: -52.887 -> 17.461 | Helped |
| add within-episode CVaR risk control (ablation) | FinCon · Portfolio management (TSLA, MSFT, PFE) | CR %: 14.699 -> 113.836 | Helped |
| add over-episode belief updates (CVRF) (ablation) | FinCon · GOOG | CR %: -11.944 -> 25.077 | Helped |
| add over-episode belief updates (CVRF) (ablation) | FinCon · NIO | CR %: 8.197 -> 17.461 | Helped |
| add over-episode belief updates (CVRF) (ablation) | FinCon · Portfolio management (TSLA, MSFT, PFE) | CR %: 28.432 -> 113.836 | Helped |
| add over-episode belief updates (CVRF) (ablation) | FinCon · GOOG | SR: -0.496 -> 1.052 | Helped |
| add over-episode belief updates (CVRF) (ablation) | FinCon · NIO | SR: 0.156 -> 0.335 | Helped |
| add over-episode belief updates (CVRF) (ablation) | FinCon · Portfolio management (TSLA, MSFT, PFE) | SR: 1.181 -> 3.269 | Helped |
Surprising findings
- Without within-episode CVaR risk control, FinCon yields negative cumulative returns on GOOG and NIO, underperforming buy-and-hold on NIO.
Limitations
- Portfolio management is limited to small portfolios of three assets; scaling to tens of assets is untested.
- Multi-asset decisions show hallucination, such as references to non-existent memory event indices.
- DRL baselines failed to converge on recently listed COIN due to limited data.
Experimental setup
Datasets: Yahoo Finance stock prices Refinitiv news SEC 10-K and 10-Q filings Zacks analyst reports earnings conference call audio
Benchmarks: single stock trading (TSLA, AMZN, NIO, MSFT, AAPL, GOOG, NFLX, COIN) portfolio management (TSLA/MSFT/PFE, AMZN/GM/LLY)
Metrics: Cumulative Return Sharpe Ratio Max Drawdown
Baselines: A2C PPO DQN Generative Agents FinGPT FinMem FinAgent Markowitz MV FinRL-A2C
Builds on: FinMem FinAgent Generative Agents Reflexion TextGrad
Closest papers by meaning
- FinBen: A Holistic Financial Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.87)
- Trade in Minutes! Rationality-Driven Agentic System for Quantitative Financial Trading (ICLR 2026 · similarity 0.87)
- TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets (NeurIPS 2025 · similarity 0.86)
- HiveMind: Contribution-Guided Online Prompt Optimization of LLM Multi-Agent Systems (AAAI 2026 · similarity 0.86)
- FinSearchComp: Towards a Realistic, Expert-Level Evaluation of Financial Search and Reasoning (ICLR 2026 · similarity 0.86)
- Win Fast or Lose Slow: Balancing Speed and Accuracy in Latency-Sensitive Decisions of LLMs (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.