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

FinCon organizes its agents as a manager-analyst communication hierarchy modeled on the structure of an investment firm, and adds a risk-control component that episodically self-critiques and updates “conceptualized beliefs” — propagating them selectively to the nodes that need them rather than broadcasting, which keeps peer-to-peer communication cost down. It is evaluated on single stock trading and portfolio management. Published on the NeurIPS 2024 main conference track (Advances in Neural Information Processing Systems 37, pp. 137010–137045).

There is no code. The-FinAI/FinCon holds a single README and no LICENSE file, and has never contained source. The README, last edited in February 2026, states that the data agent relies on several commercial APIs, which makes releasing the full system and its data difficult, and promises a release “within the next 3-4 months”; that window has since elapsed with no further repository activity. The trading results the paper reports are the authors’ own backtests, and without code or data they cannot be independently reproduced.

The paper is indexed here for its lineage and its pointers rather than for anything runnable. Seven of its authors — including Yangyang Yu, Haohang Li and Jordan W. Suchow — also wrote FinMem, whose layered-memory agent it extends. In place of code the README directs readers to two released systems from the same group: InvestorBench (ACL 2025, MIT licensed) and Agent Market Arena, which carries no license and has had no commits since it was created in January 2026 and whose “accepted by WWW 2026” claim is self-reported and unverified.