Can large language models trade? What a simulated market full of AI traders showed

I'm Chris, one of the co-founders of Autopilot. Most research on AI and stocks asks whether a model can pick well. Alejandro Lopez-Lira asked a different question: what happens when a whole market is full of AI traders?

Here's the study, what it showed, and why it matters to anyone asking a chatbot for picks.

One thing first, because we're an SEC-registered investment adviser and I have to say it. Autopilot is the app. The actual investment advice comes from Autopilot Advisers, LLC. If you want the full legal version of anything in here, it's on our legal page.

1) The study

It's called "Can Large Language Models Trade? Testing Financial Theories with LLM Agents in Market Simulations," by Alejandro Lopez-Lira. It was first posted in April 2025 and, per Lopez-Lira's CV, has been submitted to a journal.

Alejandro Lopez-Lira is an assistant professor of finance at the University of Florida. The bio for Dr. Lira's AI Finance Labs, an independent Pilot on Autopilot, says its strategies are designed by Prof. Lopez-Lira, who holds a Ph.D. in Finance from the Wharton School.

2) What it did

The paper builds an open-source simulated market where the traders are language model agents, each given a trading strategy to follow. Then it watches what the market does.

It's a simulation on purpose, because you can test ideas about markets in a simulation that you can't test with real money.

3) What it showed

The agents followed the strategies they were given. And the simulated market behaved like markets do: prices found their level, bubbles formed, and prices underreacted to news at times.

The paper also flags a risk. The prompts given to AI agents can make their behavior correlated, and correlated traders can affect how stable a market is.

4) Why that matters to you

If a lot of people ask similar models similar questions, they may get similar answers, and trade the same way at the same time. That's crowding, and crowded trades can move against everyone at once.

It's one more reason a chat answer isn't an edge by itself. A published process with its own data, rules, and schedule is at least a different process from the one everyone else is running. More on that in Asking ChatGPT for stock picks at home vs following an AI Portfolio.

5) What it doesn't mean

It doesn't tell you how any Portfolio will do. A research finding isn't a promise about a Portfolio. Studies look at historical or simulated data. A Portfolio is a different process, run forward in real accounts, and it can do worse than the market. Some Portfolios go up for years. Some are flat for years.

6) How following works

Pick a Portfolio, connect the brokerage you already use, and allocate money. You give Autopilot Advisers limited authority to send orders to the account you already have. When the Portfolio changes, we send them and your broker fills them. The money stays put.

A flat cash subscription, not a percentage of your balance. Per our Form CRS dated February 2026, Basic Tier has no Base Advisory and Licensing Fee, and Premium Tier is $29.99 to $199.99 a quarter or $99.99 to $699.99 a year. Pilot subscriptions are separate and add on for each Pilot you subscribe to. Because it's a flat dollar amount, it's a bigger share of a small account than a big one, so do the math at your balance.

TLDR: Lopez-Lira built a simulated market full of language model traders. The agents followed their strategies, and the market showed price discovery, bubbles, and underreaction. The paper warns that similar prompts can make AI traders act alike. It's a simulation, not a forecast, and a reason to be wary of everyone asking the same chatbot.

Frequently asked questions

Can large language models trade?

In a simulation, yes. Lopez-Lira's paper built an open-source market where language model agents trade by assigned strategies, and they followed them. The simulated market showed price discovery, bubbles, and underreaction. The paper also warns that similar prompts can make AI traders behave alike.

What happens when AI agents trade in a market?

In Lopez-Lira's simulated market, the market behaved much like real ones, with prices finding their level, bubbles, and underreaction to news. The paper flags that prompts can correlate AI agents' behavior, which can affect market stability. It's a simulation, not a prediction of real markets.

Can AI trading agents follow a strategy?

In Lopez-Lira's study, the language model agents followed the trading strategies they were assigned inside a simulated market. That's research, not a Portfolio. On Autopilot, AI models are tools inside a Pilot's own process, and Autopilot Advisers applies the Portfolio to your account.

Start Investing

Since Autopilot Advisers is an SEC-registered investment adviser, we have to put disclaimers on stuff like this. They're below, and the full version is on our legal page.

Disclosures

© Autopilot Holdings Corporation 2026 — All rights reserved. “Autopilot” refers to Autopilot Holdings Corporation and its wholly-owned, and separately managed subsidiaries, including Autopilot Advisers, LLC, an SEC-registered investment adviser. All investing is subject to investment risks, including possible loss of the principal invested, and past performance is not indicative of future results. View our Customer Relationship Summary and other important information at joinautopilot.com/legal.

This information is provided by Autopilot for educational and illustrative purposes only and is not a recommendation, an offer to sell or the solicitation to buy any security. Autopilot relies on information from various sources believed to be reliable, including information from Clients, Pilots and other third parties, but cannot guarantee the accuracy or completeness of that information. Autopilot does not provide tax or legal advice and you are encouraged to consult with professionals before making investment decisions.

Autopilot is not affiliated with, sponsored, or endorsed by the companies listed, described, or featured on its site. Company logos or trademarks used do not imply endorsement and are the property of their respective owners.

Investing involves risk, including the possible loss of principal. Past performance does not guarantee future results. You should carefully consider your investment objectives, risk tolerance, and time horizon before investing through Autopilot.

Autopilot Advisers, LLC is an SEC-registered investment adviser. Registration as an investment adviser does not imply a certain level of skill or training. Additional information about Autopilot is available in our Form ADV Part 2A, available at adviserinfo.sec.gov or joinautopilot.com/adv.

Autopilot provides limited investment advisory services focused on Portfolio selection and suitability assessment. Our services are not intended to replace comprehensive financial planning. We recommend consulting with a qualified financial advisor regarding your complete financial situation.

This content is provided for educational and informational purposes only. It does not constitute investment, legal, or tax advice and should not be relied upon as the basis for any investment decision. You should consult with qualified professionals regarding your specific circumstances.

Autopilot does not provide tax, legal, or accounting advice. You should consult with qualified tax and legal professionals regarding your specific circumstances.

Autopilot does not hold client assets. All investments are held at the brokerage firm you select and connect to Autopilot. Your brokerage firm is responsible for trade execution, custody, and reporting. Autopilot is not responsible for your broker-dealer’s services, fees, or execution quality.

Autopilot’s services depend on technology systems, third-party APIs, and internet connectivity. System outages, API disruptions, or connectivity issues may temporarily affect service availability or trade execution timing.

Client account holdings are designed to track the Portfolio you’ve selected, but your actual holdings and performance may differ from the Portfolio due to execution timing, market conditions, brokerage capacity constraints, fractional share availability, and other factors outside Autopilot’s control.

For Portfolios inspired by publicly disclosed trading activity, there may be a delay between when a Pilot executes a trade and when that trade becomes publicly available and is incorporated into the Portfolio. This timing delay may affect the suitability or performance of trades when executed in client accounts.

Third-Party Pilot Portfolios are created and managed by independent third parties who are not affiliated with Autopilot, are not acting as your investment adviser, and owe you no fiduciary duty. Third-Party Pilots receive a share of the fees attributable to their Portfolios, which creates conflicts of interest for them and for Autopilot, and they may trade the same securities their Portfolios hold. Autopilot may convert a Third-Party Pilot Portfolio to an Autopilot Portfolio at any time.

Research findings describe academic studies of historical or simulated data and are not predictions of any Portfolio's results. Biographical and publication details are from the University of Florida faculty directory, Lopez-Lira's CV and website, the publishers' and preprint servers' records, and the Pilot's bio in the app, as read on September 24, 2026. The University of Florida, the Wharton School, the journal, and the publisher are not affiliated with Autopilot and have not endorsed it. Autopilot does not use AI to construct Portfolios, advise clients, or select investments; model outputs are inputs to the Pilot's own process and are not predictions of future results. OpenAI, Anthropic, xAI, Google, Meta, and DeepSeek are not affiliated with Autopilot, do not manage these Portfolios, and have not endorsed them. Process descriptions come from each Portfolio's description in the app and the white papers linked from it, and may change. Listing is not an endorsement of a Pilot or strategy. Fee amounts are quoted from Form CRS dated February 2026 and may change.