We evaluate the economic value of large language models (LLMs) in global fixed-income portfolio management. Using a sample of more than 500 actively managed global bond funds, we construct AI-driven portfolios by leveraging ChatGPT to generate country-allocation weights that mimic each fund’s mandate, allowing us to evaluate their performance relative to the actual funds. We document that, although AI-driven portfolios are more diversified, they deliver lower returns, higher volatility, and lower portfolio turnover relative to actual funds. On a risk-adjusted basis, the performance spread between AI-driven and actual portfolios is largely negative, with fewer than 5% of funds showing significantly positive values. The underperformance persists across credit quality, duration, and currency segments and remains robust to the exclusion of major risk-off periods. These results suggest that LLMs remain limited in replicating the forward-looking judgment of portfolio managers.
Ceballos et al. (Mon,) studied this question.