Mr. AI Will See You Now

One early test of this is financial services. Financial advisors scour publicly available information to recommend investment options to their clients. It stands to reason that AI systems like GPT-4 might be ideally suited to perform such a task—in particular given that the financial advice industry has already moved to automated, digital offerings for the mass market.

CESifo Fellow Lars Hornuf and his colleagues Christian Fieberg and David Streich decided to explore, in their latest CESifo working paper, the suitability of the financial advice provided by GPT-4. They used GPT-4, a more advanced version than the better-know ChatGTP, because GPT-4’s deep learning approach leverages more data and more computational muscle to create increasingly sophisticated responses and is already in use by the financial industry. And they chose financial advice because financial decision-making is a highly consequential domain in our ageing societies, with large chunks of society approaching retirement and concerned about their financial security in the years to come. The thing is, retail investors like these soon-to-be retirees tend to make costly mistakes when investing their savings.

As it turns out, GTP-4 is very good at matching information on the client’s individual situation to a suitable portfolio of financial products. Its financial advice is on par with that provided by professional low-cost automated financial advisory services and, while the portfolios it suggests display considerable home bias, risk-adjusted returns were no worse than benchmark portfolios. On top of that, GPT-4 saves investors the compounded advisory fees they would otherwise have to pay.

GPT-4 provides its advice by the book, i.e., according to the long-established portfolio theory, even explaining the reasoning behind its advice. In doing so, it takes account of the prospective investors’ sustainability preferences and, most importantly, their self-expressed risk preference.

But this is precisely where AI comes a bit short. GPT-4 cannot handle risk profiling, a key part of regulatory guidelines for financial advisors. While risk tolerance is the first factor GPT-4 suggests considering when making an investment decision, even providing guiding questions to help the investor, assigning a risk profile is ultimately left to the investor herself.

Furthermore, while specific products, portfolio shares, and exchange tickers are included in the recommendation, GPT-4 cannot assist in implementing the portfolio, i.e., opening an account, or purchasing and rebalancing portfolio components.

Thus, while GPT-4 does well in matching investor profiles to specific portfolios, it will likely not make the entire financial advisory process redundant in the near future. Flesh-and-blood financial advisors can breathe a sigh of relief.

That said, any improvements in the quality of generative AI responses that are to be expected with future releases should definitely keep them awake at night. Poor souls.