10/29/2026
Tell a large language model that a portfolio manager protected 77% of investor capital in a market sell-off, and it scores the manager’s drawdown management one way.
Tell it the same manager lost 23% of investor capital, same event, same numbers, and it scores it differently.
This is framing bias: a cognitive bias in which identical information leads to different decisions depending on how it is presented. A new CFA Institute Research and Policy Center report, Managing LLM Bias in Investing: From Detection to Mitigation, found this effect consistently across nine large language models and ten investment scenarios.
The report also compares approaches to mitigating bias.
Read the report: https://cfainst.is/4hQBf6e