There’s no shortage of breathless takes about AI transforming finance. Most of them are wrong in obvious ways. But every once in a while, someone runs a real test instead of just tweeting about “disruption.” That’s what Sangheum Cho did in a recent working paper, “Can ChatGPT Generate Stock Tickets to Buy and Sell for Day Trading?” He asked ChatGPT to pick day trading stocks based on news headlines — and then actually tested the results.

It worked.

Sort of.

By feeding ChatGPT hundreds of Bloomberg and Wall Street Journal tweets (most of which weren’t about specific companies), the model generated lists of stocks to buy and sell each day. A simple long-short strategy — buying the “buy” list, shorting the “sell” list — produced meaningful returns from open to close. The long-short spread earned over 3.7% per month, before transaction costs, using a stronger signal based on multiple rounds of prompt iteration.

It’s worth emphasizing how different this is from prior studies. Most of the earlier research (like that of Lopez-Lira and Tang) focused on firm-specific news. This paper didn’t. It mainly used macroeconomic or sector-wide headlines and let ChatGPT infer which individual stocks would react. In other words, it wasn’t “this news is about Tesla, so trade Tesla.” It was “inflation news is easing — so pick tech stocks.”

That’s a harder and more interesting test. And ChatGPT, it turns out, is good at reading between the lines.

Of course, there are some significant caveats:

• The short leg drove most of the profits. The buy side alone was weak. Shorts can be expensive and hard to locate.
• Stocks selected tended to have low liquidity.
• There were strong overnight reversals, especially on the short side. Positions that looked smart intraday often bounced sharply after hours.
• Trading costs would be significant.

Maybe most importantly: while the model identified short-term mispricings, it didn’t seem to grasp overnight risk dynamics. If you blindly held the positions overnight, your edge degraded fast.

There’s a deeper lesson underneath all the performance statistics. The real breakthrough isn’t that ChatGPT “beat the market.” It’s that it did something most traditional quant models struggle with: processing bulk macro news and translating it into firm-specific bets without explicit tagging.

In effect, the model used a messy pile of broad news to generate specific, tradable mispricing signals. It wasn’t simply summarizing or classifying. It was inferring.

That’s the piece most people hyping AI in finance miss:
The goal isn’t replacing traders. The goal is to use language models to bridge the gap between macro noise and micro opportunity.

If ChatGPT can do that even semi-competently, it’s already more useful than 90% of alternative data feeds (and at least 50% of traders).

Disclaimer

This document does not constitute advice or a recommendation or offer to sell or a solicitation to deal in any security or financial product. It is provided for information purposes only and on the understanding that the recipient has sufficient knowledge and experience to be able to understand and make their own evaluation of the proposals and services described herein, any risks associated therewith and any related legal, tax, accounting, or other material considerations. To the extent that the reader has any questions regarding the applicability of any specific issue discussed above to their specific portfolio or situation, prospective investors are encouraged to contact HTAA or consult with the professional advisor of their choosing.

Except where otherwise indicated, the information contained in this article is based on matters as they exist as of the date of preparation of such material and not as of the date of distribution of any future date. Recipients should not rely on this material in making any future investment decision.