# TradingAgents Has 100k Stars and One Three-Month Backtest > Source: > Published: 2026-08-25 12:08:37+00:00 [AI](https://sourcefeed.dev/c/ai)Article # TradingAgents Has 100k Stars and One Three-Month Backtest The multi-agent trading framework is a superb reference architecture wrapped around a return figure nobody should trust. [Mariana Souza](https://sourcefeed.dev/u/mariana_souza) [TradingAgents](https://github.com/TauricResearch/TradingAgents) is about to cross 100,000 GitHub stars, which almost certainly makes it the most-starred open-source trading project on GitHub — and it has never been shown to make money outside a three-month backtest on three stocks. Both of those facts matter, and they point in different directions. The project comes out of Tauric Research, a group with roots at UCLA and MIT, and it does one thing that's genuinely clever: instead of asking a single model "should I buy AAPL?", it stages a fake hedge fund. Four analyst agents (fundamentals, sentiment, news, technicals) write reports. A bull and a bear researcher argue over them. A trader agent proposes a position. Three risk personas — aggressive, conservative, neutral — fight about sizing. A portfolio manager issues a five-tier rating from Buy to Sell. All of it runs as a [LangGraph](https://docs.langchain.com/oss/python/langgraph/overview) state machine with typed handoffs between nodes. ## The number everyone quotes, and what it actually is The headline stat you'll see repeated is "up to 30.5% annualized returns." Here's where it comes from. The December 2024 paper ([arXiv:2412.20138](https://arxiv.org/abs/2412.20138)) backtests from January 1 to March 29, 2024, and reports on AAPL, GOOGL and AMZN. Quick-thinking work went to gpt-4o-mini and gpt-4o; the deep reasoning went to o1-preview. On AAPL the system returned 26.62% cumulative while buy-and-hold lost 5.23%, with a Sharpe ratio of 8.21. A Sharpe above 3 is considered excellent. Above 8 is not a result; it's a warning light. To the authors' credit, they say so in a footnote: the ratio "exceeds our expected empirical range," they attribute it to a period with few pullbacks, and they explain the window was capped at three months because each prediction costs roughly 11 LLM calls and 20-plus tool calls. That's a candid disclosure. It's also the reason you should read "30.5% annualized" as "a quarter's worth of large-cap tech in a bull run, extrapolated." The one thing the original setup got right — and almost nobody re-running it today does — is that Q1 2024 sat after the October 2023 training cutoff of the models used. Re-run the same backtest with a 2026 model and the window is deep inside its training data. The [Profit Mirage](https://arxiv.org/abs/2510.07920) paper from late 2025 named this pattern: LLM agent returns that look dazzling in backtests evaporate once the model's knowledge window ends. A May 2026 survey of LLM trading agents (arXiv:2605.19337) went further. Of 19 empirically evaluated systems it reviewed, one reported an explicit transaction-cost model and zero met its top reproducibility tier. TradingAgents' own backtest models no slippage, no taxes, no market impact, and zero latency. So the alpha claim is unproven, and the project's own changelog tells you the team knows the plumbing was leaky: v0.2.3 (March 2026) "fixed backtesting look-ahead bias," v0.2.5 rewrote the sentiment analyst to read real Yahoo News, StockTwits and Reddit "instead of generating fabricated content," and July's v0.3.1 patched look-ahead filtering in the [Alpha Vantage](https://www.alphavantage.co/) fundamentals path. Those are the right fixes. They're also an admission that for most of the project's star-collecting life, the sentiment agent was hallucinating its inputs. ## Why it's still worth your time Now the other direction. Strip the trading claims away and TradingAgents is one of the best-engineered examples of a multi-agent pattern that most teams get wrong: adversarial role decomposition with a structured decision at the end. The 2026 releases are where it got serious. v0.2.4 replaced the BM25 "financial situation memory" with a persistent decision log at `~/.tradingagents/memory/trading_memory.md` that feeds prior calls and their realized returns back into the next run — cheap, inspectable, and far more honest than vector-store memory that nobody audits. The Research Manager, Trader and Portfolio Manager now return Pydantic-typed objects via provider-native structured output instead of regex-scraped prose. Runs can be resumed from a LangGraph checkpoint with `--checkpoint` , which matters when a 12-agent graph dies on call 9 of 11 because Alpha Vantage rate-limited you. v0.3.0 added a proper provider registry (OpenAI, Anthropic, Gemini, DeepSeek, Qwen, Bedrock, Ollama, any OpenAI-compatible endpoint), a CI gate across Python 3.10–3.13, FRED macro data and Polymarket event probabilities as vendors, and `TRADINGAGENTS_*` environment variables that override config with explicit precedence. That's a reference architecture. If you're building anything where you want a model to argue with itself before committing — due-diligence memos, incident post-mortems, vendor evaluations — the bull/bear/risk-committee shape here is worth stealing, and the code is Apache 2.0. ## Running it without fooling yourself Setup is a clone, `pip install .` , and a `tradingagents` CLI that walks you through ticker, date, analyst selection and provider keys. The two config knobs that dominate cost and quality are `deep_think_llm` and `quick_think_llm` . Put your expensive reasoning model only on the deep slot; at 11 calls per decision, a naive "use the best model everywhere" config across a year of daily decisions on a handful of tickers gets expensive fast. `max_debate_rounds` is the other lever — the paper ran its results with a single round, and every extra round multiplies the call count. If you want to evaluate it, three rules. Pick a window that starts after your model's documented training cutoff, or the result is meaningless. Pin `temperature` and use the retry budget (`llm_max_retries` ) so a flaky provider doesn't silently change your sample. And run buy-and-hold on the same window as the baseline — in the paper's own table, TradingAgents beat B&H on AMZN by only about six points, and the rule-based baselines beat it on drawdown. ## Where this lands TradingAgents is a research scaffold that people keep mistaking for a trading system, partly because its README leads with the 30.5% number and partly because 100k stars looks like validation. It isn't. The evidence that debating LLM agents produce durable alpha is, right now, a quarter of data on three mega-caps, and the surrounding literature has spent 2025 and 2026 showing why that kind of evidence doesn't survive contact with post-cutoff markets. But as a codebase for orchestrating specialized agents with typed outputs, checkpointed state and grounded tools, it's ahead of most of what's being shipped inside companies. Use it for what the disclaimer says it is: research, and a very good one to read. ## Sources & further reading - [TauricResearch/TradingAgents](https://github.com/TauricResearch/TradingAgents)— github.com - [TradingAgents CHANGELOG](https://github.com/TauricResearch/TradingAgents/blob/main/CHANGELOG.md)— github.com - [TradingAgents: Multi-Agents LLM Financial Trading Framework](https://arxiv.org/abs/2412.20138)— arxiv.org - [Profit Mirage: Revisiting Information Leakage in LLM-based Financial Agents](https://arxiv.org/abs/2510.07920)— arxiv.org - [Agentic Trading: When LLM Agents Meet Financial Markets](https://arxiv.org/abs/2605.19337)— arxiv.org - [TradingAgents v0.2.4: A Multi-Agent LLM Framework That Simulates an Entire Trading Firm](https://dev.to/_46ea277e677b888e0cd13/tradingagents-v024-a-multi-agent-llm-framework-that-simulates-an-entire-trading-firm-g2e)— dev.to [Mariana Souza](https://sourcefeed.dev/u/mariana_souza)· Senior Editor Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon. ## Discussion 0 No comments yet Be the first to weigh in.