New project: Agentic Equities

I built a new product called Agentic Equities.

The motivating observation is that more and more people are turning to ChatGPT for financial advice. Specifically, they're asking ChatGPT what stocks to buy in general, and what it thinks about certain stocks in particular.

Agentic Equities queries the underlying OpenAI models for ratings of the S&P 500 and select other stocks. The result is a list of stocks ranked by OpenAI's 12-month outlook for the share price and overall rating (buy/hold/sell).

How it works:

• I gathered an initial set of 500+ stocks—company names and stock symbols.
• I set up a script to fetch the latest prices of those stocks from AlphaVantage (first) and Yahoo Finance (as a fallback).
• I wrote a general prompt for OpenAI to search for information relevant to a certain stock and return a response in a specific format: buy/hold/sell rating, a 12-month price target, and a brief memo explaining its reasoning.
• I wrote seven specific prompts for "analyst personas," where each prompt describes a different investment philosophy. You can read those specific prompts here.
• Then I set up a Next.js app with Supabase to display the results generated by these scripts in a table that orders results by upside (descending), where upside is the simple ratio of GPT price target to current stock price.
• I run the market price update script daily after market close, and I run the GPT ratings fetch weekly on Thursdays after market close. (That GPT script generates seven ratings for 530 stocks each, or 3,710 separate OpenAI calls using gpt-4o-search-preview-2025-03-11, which adds up to about $100 per run.)

This project was partly inspired by Bloomberg writer Matt Levine's observations about how retail trading may be increasingly directed by ChatGPT. I sent the project to Matt a couple of days after I built it, and he covered it in his August 7 newsletter, which to date has driven about 6,000 visitors to the site.

When Levine covered the site, I quickly set up a Substack for people to sign up for updates. I now send that newsletter on Fridays after the new weekly GPT ratings come in. You can find some early observations based on the data there.

What's next:

• This week, I'm going to adopt GPT-5, which is the new default model in ChatGPT. I'll probably migrate the existing analyst personas over to GPT-5, though I may leave a couple on the previous model as a kind of control to compare different model behaviors. GPT-5 is about half the price of gpt-4o-search-preview-2025-03-11, which is nice.
• I'm also planning to look into the main hypothesis I wrote about in the Substack last week —that GPT ratings have built-in volatility because of the way LLMs use a technique called sampling. I'm planning to gather ratings for one stock over a discrete time period and compare those ratings' variance to the market price variance. My hypothesis is that the GPT variance will be significantly higher than the price variance.
• I've also received some inbounds from people in finance who are interested in exploring various directions for this project.

This is not investment advice.