TradingAgents Review: The AI Bot's 26% Claim, Checked
TradingAgents review: the AI trading agent's 26.6% Apple return came from one quarter on 3 stocks. What we found in the code.
TradingAgents is the most popular open-source AI trading project on GitHub, with roughly 80,000 stars. Its paper reports a 26.62% return on Apple stock in three months, in a quarter when Apple itself fell 5.23%. We read the whole codebase and the paper before running a dollar through it. The 26% is real as a backtest result. It’s also one quarter, three stocks, and a pipeline that later needed fixes for data it could see too early.
So it’s interesting. It isn’t proof.
We went looking because we’d already tested a pile of trading ideas the boring way, and most died on costs or bad data. An AI agent team arguing about stocks sounded like a different kind of idea. Different, yes. Better, not proven.
What TradingAgents actually does
It splits stock analysis into roles, each played by a large language model. Analysts read fundamentals, news, social media and price charts. A bull researcher and a bear researcher then argue over those reports. A trader agent turns the debate into a buy, sell or hold call, and a risk team plus a portfolio manager sign off.
That debate step is the clever part. Every recommendation comes with the case against it attached, which is more than most stock newsletters give you.
It does not trade. There’s no broker connection in the code, so it only produces a decision you’d act on yourself.
How good was the 26% result, really?
The headline comes from Table 1 of the paper (arXiv 2412.20138). The test ran January 1 to March 29, 2024, on Apple, Google and Amazon.
| Stock | Buy and hold | TradingAgents | Sharpe ratio | Max drawdown |
|---|---|---|---|---|
| Apple | -5.23% | 26.62% | 8.21 | 0.91% |
| 7.78% | 24.36% | 6.39 | 1.69% | |
| Amazon | 17.1% | 23.21% | 5.60 | 2.11% |
Look at that Sharpe ratio. 8.21. Most professional funds would celebrate anything above 2 over a full cycle. A number like that over about 60 trading days says less about skill and more about how small the sample is.
Three stocks. One quarter. All large tech names in the same market. You can’t separate luck from skill on that, and the paper doesn’t claim a confidence interval that would.
The four leaks users found
This is the part we’d want any reader to know. Between August and September 2026, the project’s changelog shows four look-ahead fixes, all reported by users, not the authors:
- Company fundamentals were stamped with the current date, so a backtest could see later numbers (issue 1300).
- Economic data from FRED came from today’s revised series instead of what was published at the time (issue 1275).
- The bot’s memory of past decisions had no point-in-time guard, so lessons could leak backward (issue 1251).
- StockTwits and Reddit sentiment was fetched with no date window at all.
The paper came out in December 2024. Those fixes landed almost two years later. We can’t say the 26% depended on any one leak. We can say it was measured on a pipeline that was still leaking, and the README now warns that results may not match any published figure.
We’ve made this exact mistake ourselves. In our own congressional trading backtest, every error we found pushed results up, never down.
What it costs to run
Each ticker, each day, means a dozen or so AI calls. On a top model we estimated $0.30 to $0.50 per ticker. Ten tickers daily runs $60 to $100 a month. Smaller or local models cut that by about 90%, but they also argue less well.
That bill shows up before you know whether the signal beats an index fund. Run the drag through our fee drag calculator and a $100 monthly cost on a $20,000 account is a 6% annual hurdle. The bot has to beat SPY by 6 points a year just to break even with doing nothing.
Should you try it?
If you like reading code and want to see how AI agents reason about stocks, yes. It’s well built. We read every file, found no hidden network calls and no way for it to touch your brokerage.
If you want it to make money, be honest with yourself about the evidence. One quarter, three stocks, look-ahead fixes after publication, and the authors’ own warning that it’s a research scaffold.
We did keep one idea. The bull-versus-bear debate works without the rest of the machine. Before we act on any stock idea now, we write out the strongest case against it in a few sentences, the same way the bear researcher would. It costs nothing. And an idea that can’t survive its own counter-case wasn’t much of an idea, no matter how confident the first draft sounded.
Paper trade it for months. Compare every pick against SPY over the identical dates, not the whole year. And count your AI bill as a cost, because it is one.
We wrote up the general traps in five backtest mistakes that fake an edge. Every one of them applies to AI agents too, just with better prose attached.
Frequently asked questions
Is TradingAgents safe to install?
We read the whole codebase in September 2026 before running anything. It had 145 Python files, no hidden binaries, no install hooks and no code that sends your data out with POST requests. It also has no broker connection, so it can't place an order. The one outbound call at startup is a small news banner, which you can turn off. Safe to read and run, in our view, but it will spend money on AI calls.
How much does it cost to run TradingAgents?
It runs a dozen or so AI agents per ticker per day, so the bill depends on the model. Using a top-tier model, our estimate was roughly $0.30 to $0.50 per ticker per run. Cheaper or local models cut that by about 90%. Ten tickers a day on a frontier model is around $3 to $5 daily, or $60 to $100 a month, before you know whether it adds anything over holding an index fund.
Did TradingAgents really make 26% on Apple?
In the paper's backtest, yes: 26.62% from January 1 to March 29, 2024, while Apple itself fell 5.23%. But that's one quarter, three stocks, and a code base that later needed several fixes for data the agents could see too early. The authors now say results may not match any published figure. We'd treat 26% as a research result, not an expected return.
What is lookahead bias in an AI trading bot?
It means the bot sees information during a backtest that didn't exist yet on the simulated date. In TradingAgents, users found company fundamentals stamped with today's date, economic data that had been revised after the fact, and social media posts fetched with no date limit. Each one lets a backtest peek at the future, which makes results look better than any live run could be.
Can I use TradingAgents with real money?
It doesn't connect to a broker, so you'd have to place every trade yourself from its output. The project's own README calls it a research framework and says it isn't trading advice. If you try it, paper trade first, for months, and compare every result against simply holding SPY over the exact same dates. That benchmark is where most strategies quietly lose.
RELATED READING