Xiao, Sun, Luo & Wang (2024) — TradingAgents: Multi-Agents LLM Financial Trading Framework
Citation. Xiao, Y., Sun, E., Luo, D., & Wang, W. (2024). TradingAgents: Multi-Agents LLM Financial Trading Framework. arXiv:2412.20138.
One-line takeaway. Simulate a trading firm as a swarm of specialised LLM agents — fundamental, sentiment and technical analysts, researchers, traders and a risk manager — deliberating in natural language toward a trade, and report backtests that beat single-agent baselines. The nearest thing to my proposal in the literature, and the design I most want to argue against.
What the paper claims
Decomposing an LLM trading system into specialised role-agents that deliberate produces better and more “human-like” decisions than one monolithic LLM. The framework mirrors a real desk — analysts feed researchers, who feed traders with varied risk profiles, overseen by a risk-manager agent — and the reported backtests show superior cumulative return, Sharpe ratio and risk management versus single-agent and simple baselines.
How they show it
A multi-modal dataset (prices, news, social-media sentiment, insider transactions, filings, technical indicators) drives the agents; evaluation is a backtest over a short, recent window (data from early 2024, trading over a several-month 2024 span in the reported setup), compared against single-agent and passive baselines.
What I’d push on
- The evaluation is exactly the kind Li et al. (2025) dismantle. A months-long, single-period, recent (largely bull) window over a narrow universe is precisely where LLM strategies look best and generalise worst; FINSABER shows such advantages evaporate under broad-universe, multi-decade, bias-controlled testing. There is no deflated Sharpe or multiple-testing correction.
- Look-ahead is hard to rule out. A frozen LLM whose training data postdates the trading window may effectively “know” outcomes when it reads 2024 news in 2024 — a contamination that a point-in-time, provider-sourced pipeline is built to prevent and a chat model cannot easily guarantee.
- Free-text deliberation is the wrong primitive for capital. Natural-language agent debate is non-deterministic, unbounded, hard to audit, and error-compounding, and it re-opens the prompt-injection surface at every hop — the concerns my multi-agent coordination entry argues make this design unsafe near money.
- A risk-manager agent is a soft control. An LLM overseeing other LLMs is gameable and unverifiable; it is not the hard, deterministic risk gate that an execution boundary requires.
- Regime-blindness. Consistent with FINSABER’s finding that such systems are conservative in bulls and reckless in bears, there is no explicit regime conditioning — the adaptation the whole approach most needs is left to emerge, and doesn’t.
How it connects to the proposal
TradingAgents is simultaneously my nearest precedent and my clearest foil, and citing it precisely is how I explain what the proposal is not. I adopt its one genuinely right idea — decomposition into specialised agents — and reject essentially every other choice: structured, typed shared state instead of free-text debate; an LLM that is never an execution authority; a deterministic risk gate rather than a risk-manager agent; regime-awareness by construction; and deflated, out-of-sample, honest-null evaluation. Where TradingAgents builds for autonomy and an impressive demo, the proposal builds for control, auditability and evidence — the axis its successors’ critics say is missing.