Hey, I’m Matt. I’m a former Bloomberg News reporter, and you’re reading AI Street, where I report on how Wall Street uses AI. On Tuesdays, I highlight novel research, emerging use cases, and expert interviews.
If you tell ChatGPT “Take on the persona of Warren Buffett,” you know you’re not getting the real Oracle of Omaha.
But you might think, as I did, that instruction helps get some of the billionaire’s investing judgment to come through in its answers.
New research is calling that assumption into question. Instructing an AI model that it is a “therapist, a rocket scientist or an elite money manager” does little to shape the model’s actual forecasts. The results may sound cogent, but they don’t really reflect actual judgment but “a continuation of the prompt rather than a stored belief.”
That last quote comes from a recent paper from the University of Chicago that tested how AI forecasts compared to those from humans. The short answer: not good.
JFYI: We’ve previously covered other shortcomings of AI in finance. Some examples: positional bias (the same information, presented in a different order, can produce different outputs), framing effects (the same odds stated two ways lead to different decisions), and, of course, biases inherent in the data the models are trained on. For more info see:
NOTE FROM OUR SPONSOR
The Best Dataset Isn’t For Sale — Yet
Months before Google agreed to pay $10 million for Spirit Airlines’ data, Brickroad had already flagged it.
The company’s AI agents scan news, earnings reports, app-store rankings and changes to terms of service for signs that a company is producing valuable data — and might be willing to license it. When a source looks promising, Brickroad contacts the company, secures a sample and helps the buyer evaluate the data before committing to a deal.
You can’t really call up Warren Buffett and say “hey, Warren, pal, should I be buying banks in this high-rate environment?” But you can test AI predictions against humans who already provide their forecasts to the public: prediction markets and analysts’ earnings forecasts.
Here’s how University of Chicago professors constructed their research. (Full disclosure: I write for Chicago Booth Review, a publication of the business school.)
In prediction markets, they used event contracts, which are binary. Either the event happened or it did not. The authors don’t specify the exact kinds of events, but provide general topics such as economics, politics, sports and weather.
The AI model is asked the same questions dozens of times, but what changes is the persona. The AI is instructed to be younger or older, and more or less willing to take on financial risk. Here’s their prompt:
The brackets could be filled out from fixed choices—White, Black, Hispanic or Asian for race; ‘some college or less’ or ‘college or more’ for education; and low, moderate or high tolerance for financial risk, along with set choices for age, gender, income and politics.
All of these different permutations created 1,296 possible fake AI personas. The researchers didn’t test every single persona but rather 54 profiles for each event on a given day, drawing from the larger pool over time.
The researchers compared the model’s forecast with the current market price of the event, which shows the odds implied by the trades. In other words, how likely humans think an event will happen.
And as for the results: The different personas mostly agreed with one another. The AI assigned the outcome that eventually happened a 60.6% chance on average, compared with 84% implied by market prices—a gap of more than 23 percentage points. The AI stand-in remained much closer to a coin flip than the traders.
Behind the paywall: what happened when the researchers fed the AI a deliberately false market price—and how its forecasts stacked up against Wall Street earnings analysts.