Valerie Baldinger

Generative AI is a general purpose technology with rapidly growing adoption among retail investors. How does the adoption of this technology change household investment decisions? I study the adoption of a generative AI–based conversational assistant by retail investors, using data that link investors’ prompts and AI answers to transaction-level trading records and portfolio holdings. Using a staggered difference-in-differences design, I find that generative AI adoption increases investment in index funds relative to individual stocks by 7.6 percentage points. As a result, AI adoption reduces investors’ return loss, which measures the expected return forgone due to underdiversification at a given level of risk, by 43 basis points on average, with substantial heterogeneity across investors. Less diversified portfolios improve the most, with return loss declining by 101 basis points.

A large literature is concerned with understanding why asset price bubbles form. The theoretical literature has proposed multiple mechanisms, including extrapolative beliefs, heterogeneous expectations, and speculation as possible explanations. I develop a methodology using LLM agents as traders in experimental asset markets to study why asset price bubbles form. The methodology exploits unique properties of LLM subjects to disentangle the contribution of mechanisms proposed in the literature: their memories can be selectively edited without awareness, their reasoning can be probed without contaminating subsequent behavior, and market conditions can be varied at scale. I find that LLM traders generate many of the same market phenomena observed with human subjects: price bubbles emerge and crash, experience reduces mispricing, and short selling attenuates overpricing. These results hold across variations in trading mechanisms and participant characteristics. LLM agents exhibit extrapolative beliefs, expecting rising prices to continue rising even above fundamental value. Their reported reasoning places greater weight on past price trends as those trends emerge, before shifting back toward fundamentals as the terminal period approaches. By homogenizing agents’ forecasts while preserving extrapolation, I isolate the contribution of heterogeneous beliefs to bubble formation.