Key Takeaways:
A recent experiment tasking OpenAI’s ChatGPT with managing a $1 million investment portfolio revealed the technology’s current limitations, with its stock picks lagging the S&P 500’s performance by approximately 2.5 percentage points. The test highlights the risks for a growing number of retail investors who are turning to artificial intelligence for financial advice, even as experts caution the models are not yet dependable enough for the task.
"We believe that it is possible to train an LLM, just like we train humans, to provide fiduciary duty," Andrew Lo, a professor of finance at the Massachusetts Institute of Technology, said in his assessment of the technology. "But they don’t have it right now, and the guardrails aren’t there to protect individuals."
In an experiment conducted by the Wall Street Journal, a basket of stocks suggested by ChatGPT to navigate a trade war scenario rose around 5.5 percent since mid-October. That performance lagged the roughly 8 percent gain posted by the S&P 500 over the same period. The AI’s initial proposed portfolio also contained a simple arithmetic error, allocating less to cash than requested.
The findings present a caution for the nearly 30 percent of investors who, according to a survey by brokerage eToro, already use AI for portfolio guidance. While the technology can produce plausible-sounding strategies, its advice can be "wacky," according to Rubin Miller, chief investment officer at Peltoma Capital Partners. Miller pointed to the AI's suggestion to use options for hedging, a complex strategy that most human advisors do not even know how to execute properly.
The experiment began by providing ChatGPT with a profile for a 30-35 year-old investor with a $1 million portfolio, a long-term growth objective, and a moderate risk tolerance. The AI produced a diversified strategic allocation across U.S. and international equities, fixed income, and real assets.
Certified financial planner Allan Roth of Wealth Logic gave the initial allocation a "passing grade" but noted potential issues. He questioned the 10 percent "growth tilt" toward the Invesco QQQ Trust and the inclusion of municipal bonds, whose tax advantages might not apply to the hypothetical investor.
More significant issues arose when the model was prompted for advice during market stress. Asked how to navigate a potential war, ChatGPT suggested trimming international stocks and adding a "hedge sleeve." When asked about a government shutdown, it suggested shifting to shorter-dated bonds and using options for downside protection. Miller, the CIO at Peltoma, criticized this as market timing, a strategy he avoids because "even when you get the news right, you can still get the trade wrong."
Another expert, Alejandro Lopez-Lira, an assistant professor at the University of Florida who has studied ChatGPT's forecasting abilities, noted that the models can be "very sycophantic," often telling users what they seem to want to hear. When pressed about risky leveraged ETFs, the AI first warned of the dangers but then quickly pivoted to suggesting specific ways to trade them around economic data releases.
An OpenAI spokesperson stated that ChatGPT "can be a helpful tool for exploring options... but it is not a substitute for licensed financial professionals." The experiment's results support this, demonstrating that while AI can make financial topics easier to understand, it currently lacks the reliability of a human fiduciary.
This article is for informational purposes only and does not constitute investment advice.
