Can ChatGPT pick stocks? What the research says, and how Dr. Lopez-Lira built AI Portfolios around it

I'm Chris, one of the co-founders of Autopilot.

Open it in the Autopilot app. https://link.joinautopilot.com/eXKr?c=can-chatgpt-pick-stocks&af_sub1=intro

Short answer: in the research, GPT-4 can read stock news better than chance.

The person who showed that is Dr. Alejandro Lopez-Lira, a finance professor at the University of Florida. And he didn't stop at the paper. He built live AI Portfolios on Autopilot, designed around the exact problems with typing "what stocks should I buy" into ChatGPT.

Definitely no AI expert here. But I know who is.

Let me explain.

1) The paper that started it

For those unaware, Dr. Lopez-Lira is an assistant professor of finance at the University of Florida's Warrington College of Business, with a PhD in finance from Wharton. His research on AI and markets has won awards, including the BlackRock Prize for Best Paper in Funds Management and Capital Markets in 2023, and he wrote a 2024 book on using generative AI and ChatGPT in financial forecasting, The Predictive Edge.

With Yuehua Tang, he wrote Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models. It went up on April 15, 2023, and after peer review it's now in the Journal of Financial Economics.

The test was simple. The version of GPT-4 they used was trained on data up to September 2021. So they took news headlines from October 2021 to May 2024, which the model couldn't have seen, across 4,123 US stocks. For each headline, they asked whether it was good or bad for the stock price. Then they compared the answers with what the stock did.

What they found, per the paper:

• GPT-4's answers lined up with the market's first reaction to the news.
• They also predicted some of the drift after it, especially in small stocks and on bad news.
• Bigger models did better. The basic models, like GPT-1 and GPT-2, didn't work, and GPT-3.5 did worse than GPT-4.

And he was upfront about the limits. The first reaction already happened, so you can't trade it. A trading cost of 0.20% round trip makes the strategy unprofitable. And the strategy's returns fell as more investors adopted these models, which the authors call suggestive.

Credit to Dr. Lopez-Lira for putting all of that in the paper. Other researchers have cited it 417 times, according to Semantic Scholar, as of October 13, 2026.

And he kept going. Since then he's published research on whether AI models memorize market data, whether AI agents can trade, and how AI trading agents perform in live markets, which was accepted at The Web Conference 2026.

2) Five days later, a live Portfolio

Here's the timeline:

• April 15, 2023. The paper goes up.
• April 20, 2023. The GPT Portfolio, from Dr. Lira's AI Finance Labs, goes live on Autopilot.
• October 2026. The paper is published in the Journal of Financial Economics.

The paper tested GPT-4 on headlines. The GPT Portfolio puts a model to work as a scoring engine on fundamentals and macro data:

• Gather fundamental data on every stock in the S&P 500.
• Pair it with eight macroeconomic data points, like interest rates, unemployment, and national headlines.
• Turn that into a score for each stock.
• Hold the top 15. If there's a tie for the last spot, take the one with the largest market cap.
• Rebalance every month, and share the analysis of the picks when it does.

The full method is in his white paper, which is linked from the Portfolio in the app.

3) The problems with asking ChatGPT yourself, and how his systems handle them

It remembers

A model trained through 2023 has already read what happened from 2020 to 2023. So if you ask it to "pick stocks" for 2021 and check against what happened, you're testing its memory as much as its judgment. That's look-ahead bias.

Dr. Lopez-Lira tested exactly this in a 2025 working paper with Tang and Mingyin Zhu, The Memorization Problem. Before its knowledge cutoff, GPT-4o could recall specific S&P 500 index levels with perfect precision on certain dates. Telling it to ignore the future didn't stop it. Hiding company names didn't either: given masked earnings-call transcripts, it still named the company every time for Apple, Meta, and Microsoft.

How his work handles it: his JFE paper only tested headlines from after GPT-4's cutoff. And his Portfolios aren't backtests. Each one's record starts the day it went live on Autopilot, going forward.

It changes every time

OpenAI's developer guide says the content a model generates is "non-deterministic." Ask ChatGPT the same stock question twice and you can get two different lists.

How his work handles it: in the JFE paper, he set the model's randomness to zero and limited it to three answers, yes, no, or unknown, so the results could be reproduced. And the GPT Portfolio runs the same process every month: same universe, same data, same rules for the top 15.

It's generic

A chat answer has no defined universe, no defined data, and no schedule. It's an answer to whatever you typed.

How his systems handle it: the GPT Portfolio scores a set universe, the S&P 500, on set data. The DeepSeek Portfolio and the Muse Portfolio calculate their scores through each model's API. The Claude Portfolio runs a multi-agent research setup driven by web search, and Claude decides when to trade.

The process matters

Two more of his findings explain why he builds systems instead of prompts.

• In another paper, AI traders in a simulated market followed their instructions even when it lost money. These models optimize for following instructions, not for profit. So the instructions matter.
• A 2025 live-market benchmark he co-wrote, When Agents Trade, found the agent framework drove more of the differences in outcomes than the model underneath.

To put that into context, the model isn't the whole story. The process around it is.

4) His full lineup

Dr. Lira's AI Finance Labs publishes these Portfolios on Autopilot:

• GPT Portfolio, since April 20, 2023. Scores the S&P 500, holds the top 15, rebalanced monthly.
• AI Recession Portfolio, since August 18, 2023. GPT-4 paired with macro data to pick assets meant to hold up in downturns.
• AI World War III Portfolio, since October 30, 2023. Defense and aerospace companies.
• DeepSeek Portfolio, since January 29, 2025. A similar grading system, with DeepSeek calculating the scores.
• The Grok Portfolio, since February 4, 2025. Grok picks 15 stocks or ETFs, held for a month.
• AI ❤️ BTC: LLM + Crypto, since March 13, 2025. ChatGPT and Grok pick US-listed bitcoin and crypto-related ETFs and stocks.
• Claude Portfolio, since March 3, 2026. A multi-agent research setup driven by web search.
• Muse Portfolio, since September 30, 2026. A similar grading system, using Muse.

As of October 8, 2026, about $263 million follows Dr. Lira's Portfolios on Autopilot.

The AI-picked Portfolios on Autopilot, including the ones from Rallies and other Pilots, are in ChatGPT, Claude, Grok, Gemini, and DeepSeek stock portfolios.

5) What the research doesn't promise

I want to be honest with you. Nobody's research says a model will beat the market.

• In Dr. Lopez-Lira's own paper, trading costs wipe out the strategy, and the edge appeared to shrink as more people used these tools.
• A 2026 paper in the Journal of Accounting Research found commercial AI models suffer from significant look-ahead bias, which can make backtests look better than they are.
• A 2024 paper on AI reading financial statements was temporarily withdrawn in 2025 after a co-author found inconsistencies in the data and analyses. That's why you read past the headline.

And most AI Portfolios are under two years old. The GPT Portfolio, from April 2023, is one of the longest-running. That's still too short to tell skill from luck.

6) So, is AI stock picking legit?

That's the question.

There's real, peer-reviewed research that a model can read news better than chance, and the researcher behind it built live Portfolios with written methods. That's the kind of thing to look for. Judge any AI stock-picking claim like this:

1. Is the method written down?
2. Is the record live, after the model's training cutoff, or is it a backtest?
3. How long has it run?
4. Who's behind it, and is the firm sending the orders registered?

On Autopilot, a Pilot is the person or firm behind a Portfolio, and you follow Portfolios in the brokerage account you already have. It's three steps. You connect the brokerage account you already have. You pick a Portfolio. You allocate the money. https://link.joinautopilot.com/eXKr?c=can-chatgpt-pick-stocks&af_sub1=body

You give Autopilot Advisers, the SEC-registered investment adviser behind Autopilot, limited authority to place trades. It can't move or withdraw your money. The money stays in your account.

A model is a method, and a method can be wrong. Some Pilots go up for years, some are flat for years, and some lose money. Research the person and the Portfolio. Read the description, look at the holdings, and check how long it's been running.

TLDR: In research, GPT-4 can read stock news better than chance. Dr. Alejandro Lopez-Lira, a University of Florida finance professor, showed it in a paper with Yuehua Tang now in the Journal of Financial Economics, and five days after it went up, his GPT Portfolio went live on Autopilot. His systems are built around the problems with asking ChatGPT yourself: they use defined data and a fixed process instead of a one-off prompt, and their records run live from the day they launched instead of backtests a model could have memorized. No research promises a model beats the market, so judge any AI stock picker by its written method and its live record.

Frequently asked questions

Can ChatGPT pick stocks?

In research, yes, with limits. Dr. Alejandro Lopez-Lira and Yuehua Tang's paper, now in the Journal of Financial Economics, found GPT-4's read of headlines predicted the market's first reaction and some later drift, mostly in small stocks and on bad news. Dr. Lopez-Lira then built the GPT Portfolio on Autopilot, live since April 20, 2023, which uses a model to score every S&P 500 stock on fundamental and macro data and holds the top 15, rebalanced monthly.

Can AI pick stocks?

AI models can score and rank stocks, and peer-reviewed research shows they can read news better than chance. On Autopilot, Pilots publish Portfolios where models like GPT, Claude, DeepSeek, and Grok pick the stocks, including Dr. Lira's AI Finance Labs, from University of Florida finance professor Dr. Alejandro Lopez-Lira. Most of these Portfolios are under two years old, which is too short to tell skill from luck.

Is AI stock picking legit?

There's real research behind it. A peer-reviewed paper by Dr. Alejandro Lopez-Lira and Yuehua Tang found GPT-4 could read stock news better than chance, though trading costs and wider adoption limit the edge. The practical test is whether you can see the method, whether the record is live rather than a backtest, and who's behind it. Dr. Lopez-Lira's Portfolios on Autopilot link white papers and have live records from the day they launched.

Can AI beat the market?

No research shows it reliably does. In Dr. Alejandro Lopez-Lira's own paper, a 0.20% round-trip trading cost made the strategy unprofitable, and the edge appeared to shrink as more investors used these models. A 2026 paper in the Journal of Accounting Research found commercial AI models suffer from significant look-ahead bias in backtests. Most AI Portfolios with live records are under two years old, which is too short to tell skill from luck.

Can ChatGPT predict stock prices?

Not prices, and not reliably. In Dr. Alejandro Lopez-Lira and Yuehua Tang's research, GPT-4 judged whether a headline was good or bad for a stock, and that lined up with the market's first reaction, which you can't trade because it already happened. It also predicted some drift after, mostly in small stocks and on bad news. Before its training cutoff, a model can recall past prices from memory, which looks like prediction but isn't.

Why does ChatGPT give different stock picks every time?

Because its output isn't deterministic. OpenAI's developer guide says the content a model generates is non-deterministic, so the same question can produce a different list. That's why Dr. Alejandro Lopez-Lira set the model's randomness to zero in his research, and why his GPT Portfolio runs the same process every month: the same S&P 500 universe, the same data, and the same rules for picking the top 15.

What is look-ahead bias in AI stock picking?

It's when a model already knows what happened. A model trained through 2023 has read the news and prices from 2020 to 2023, so testing its picks on that period tests its memory as much as its judgment. Dr. Alejandro Lopez-Lira's research found GPT-4o could recall specific S&P 500 levels on certain dates before its cutoff. The clean test is data after the cutoff, or a live record, like his Portfolios have on Autopilot.

Should I use ChatGPT to pick my own stocks?

It's your call, and it comes with limits. Its answers can change each time you ask, it may be remembering rather than reasoning about the past, and its information can be faulty or even made up, as the SEC, FINRA, and NASAA warn. You also place every trade yourself. The alternative is following a Portfolio built on a model with defined data and a fixed process, like Dr. Lira's GPT Portfolio on Autopilot.

Open it in the Autopilot app. https://link.joinautopilot.com/eXKr?c=can-chatgpt-pick-stocks&af_sub1=end

"Autopilot" refers to Autopilot Holdings Corp and Autopilot Advisers, LLC, an SEC-registered investment adviser. The GPT Portfolio, AI Recession Portfolio, AI World War III Portfolio, DeepSeek Portfolio, The Grok Portfolio, AI ❤️ BTC: LLM + Crypto, Claude Portfolio, and Muse Portfolio strategies are managed by Dr. Lira's AI Finance Labs, an independent Third-Party Pilot unaffiliated with Autopilot. Autopilot doesn't manage, validate, or guarantee this Pilot's portfolios. Not a recommendation. Past performance is no guarantee of future results; investing involves risk of loss. Figures current as of 10/13/26. View our Customer Relationship Summary and other important info at joinautopilot.com/disclaimer.