NBER Study: AI-Managed Portfolios Lack Significant Abnormal Returns
A preliminary report by the National Bureau of Economic Research investigates how artificial intelligence manages household stock portfolios. Authors Bruce I. Carlin, Ryan D. Israelsen, and Christopher F. Wazzan analyzed daily time-series data of stock recommendations generated by several Large Language Models (LLMs). The study reveals that AI tends to recommend undiversified portfolios with positive exposure to momentum, large-cap companies, and firms with low book-to-market ratios. Crucially, the research finds that AI primarily selects stocks based on the volume of media attention they receive rather than fundamental financial metrics. When evaluating performance using the methodology established by Daniel et al. (1997), the authors determined that neither buy-and-hold nor actively managed AI portfolios achieved statistically significant abnormal returns. This suggests that despite the sophistication of modern LLMs, their investment strategies do currently not outperform traditional benchmarks in a meaningful way. The findings highlight potential biases in AI-driven financial advice, particularly its reliance on media sentiment, and raise questions about the efficacy of automated portfolio management for individual households.
Wire timeline
NBER Study: AI-Managed Portfolios Lack Significant Abnormal Returns
A preliminary report by the National Bureau of Economic Research investigates how artificial intelligence manages household stock portfolios. Authors Bruce I. Carlin, Ryan D. Israelsen, and Christopher F. Wazzan analyzed daily time-series data of stock recommendations generated by several Large Language Models (LLMs). The study reveals that AI tends to recommend undiversified portfolios with positive exposure to momentum, large-cap companies, and firms with low book-to-market ratios. Crucially, the research finds that AI primarily selects stocks based on the volume of media attention they receive rather than fundamental financial metrics. When evaluating performance using the methodology established by Daniel et al. (1997), the authors determined that neither buy-and-hold nor actively managed AI portfolios achieved statistically significant abnormal returns. This suggests that despite the sophistication of modern LLMs, their investment strategies do currently not outperform traditional benchmarks in a meaningful way. The findings highlight potential biases in AI-driven financial advice, particularly its reliance on media sentiment, and raise questions about the efficacy of automated portfolio management for individual households.
National Bureau of Economic Research Working Papers