Do ChatGPT stock picks track market prices?
Welcome back to Agentic Equities – tracking the stocks ChatGPT is telling people to buy (not investment advice). Our ChatGPT ratings for all 530 stocks were updated yesterday. It took the OpenAI API longer than expected to complete, which is why this email is appearing on Saturday morning instead of Friday afternoon – will fix next week.
In this newsletter I present the new data, analyze the week-over-week changes, plot GPT price target changes v. market price changes, and present a hypothesis on the inherent volatility of ChatGPT’s stock advice. To expand our coverage and unlock future analysis, become a subscriber:
The New Top 10
Here’s the latest top 10 by upside as of market close on Friday, August 15 – from the website:
Last week, the top 10 were Eastman Chemical Company, Fiserv, Atlassian, Fortinet, Molina Healthcare, Moderna, Caesars Entertainment, International Flavors & Fragrances, Cooper Companies, and Fair Isaac. Six of the top ten highest-upside stocks remained in the top ten this week.
Trade Desk, which took the number one spot, is an interesting case. The stock is actually down about 37% since last week, but the GPT target is only down 20%. As a result, even though Trade Desk has fallen on both measures, it shot to the top of the list by upside because the GPT downgrade lags the market downgrade.
This raises the question, how do ChatGPT price targets track market price movements?
Week-Over-Week Changes
The top 20 market price gainers among the tracked stocks were the following:
The list of top 20 GPT target gainers, however, looks different:
The two top 20 lists only share seven stocks. For the seven stocks that appeared on both top 20 lists, the reasons don’t necessarily match. For example, UnitedHealth Group rallied in the past couple days largely because it was revealed that Warren Buffett recently purchased about 5 million shares of the company (Reuters). None of the recent GPT analyst memos mention this, instead focusing on each analyst persona’s primary factors (e.g. macro trends, quant analysis, and fundamentals).
I wondered if the current market price is a meaningful factor in the stock analyses performed by my GPT “analysts.” That is, if a stock price goes up, does that usually mean that the GPT price targets go up?
GPT Target Changes v. Market Price Changes
Here is a plot of GPT target price changes versus market price changes over the past week, for the 530 stocks I track:
There is indeed an upward trend line, but the R² is only 0.353, meaning stock price movements are a moderate predictor of GPT target movements, not a definitive one.
This is just from a week of data. It might be that GPT ratings don’t change much week-to-week. While the GPT model I query (gpt-4o-search-preview-2025-03-11) does search the web when generating its rating – meaning it should cite new public material information – there may only be so much variation in GPT stock ratings week over week, perhaps because the underlying model is fixed.
But that actually doesn’t seem to be the explanation. When I look at the absolute value of market price changes and GPT target changes, the median change is roughly the same: 2.65% for market prices and 2.52% for GPT targets. Gross changes were 1910% for market prices and 2101% for GPT targets. So GPT targets have changed roughly as much as market prices have changed. Why, then, are the GPT price changes not strongly correlated to market price changes on a per-stock basis?
My best explanation is that ChatGPT responses have built-in randomness. Large Language Models (LLMs) often use a method called “sampling,” which means they pick the next word1 probabilistically from a set of candidate words rather than always choosing the single most likely option. The model sometimes selects a less likely but still reasonable next word, adding variety and creativity to its responses. This is why if you ask ChatGPT the same question twice, you may get a different answer. This technique is, perhaps counter-intuitively, one reason ChatGPT responses are so compelling and human-like.
But in stock ratings, this element of randomness poses interesting problems. Each time you ask ChatGPT for financial advice, you may get a different answer – even if you ask the same question about the same stock. There’s essentially baked-in volatility to ChatGPT stock advice, volatility that would compound at scale as users phrase their chats differently.
A consequence of this randomness may be to mitigate the herd effect that Matt Levine has written about with respect to ChatGPT and retail traders. If all retail traders were getting the same exact advice from ChatGPT about the same stocks, then they might make the same trades as a unified bloc. In reality, there’s significant randomness to the financial advice that ChatGPT dispenses. ChatGPT-inspired retail activity may actually be quite variant.
What’s next
OpenAI released GPT-5 last week, which is now the default model in ChatGPT. I’m probably going to switch all the existing analyst personas to use this new model, which fortunately has web search available out-of-the-box and happens to be about half the price of the GPT-4o search model I’m currently using. Hopefully GPT-5 runs faster too 🤞
I’m also going to do a dedicated exploration of the sampling hypothesis described above. How much variance is there in GPT-generated ratings and price targets within a short period of time for a given stock, compared to the market volatility of that stock within the same time period? We’ll find out next week.
Finally, I’d like to expand stock coverage to all public equities and generate more ratings per stock – especially in light of the sampling-based volatility hypothesis. More coverage and more ratings increase my costs, so the expansion is dependent on the amount of paid subscribers to this newsletter. If you find this project valuable, become a paying subscriber:
Alternatively – and even more helpfully – purchase a group subscription for your company.
Thanks for reading.