Machine-learning about ESG preferences: Evidence from fund flows

This is the authors' abstract. We don't add key points for this paper.

We construct Environmental, Social, and Governance (ESG) scores for U.S. active equity mutual funds using roughly 500 metrics across ESG issues and rating agencies. Under a revealed-preference framework, we use machine learning to identify issues driving fund flows and generate ESG-driven flows as proxies for ESG performance. We estimate expected benchmark-adjusted returns using ex-post realized returns and ex-ante measures. Better-ESG funds attract larger subsequent flows and face fewer redemptions after poor performance, yet have lower expected returns, indicating flows reflect investor ESG preferences rather than high-return beliefs. Investors effectively pay $48 million more annually for a top-ESG fund.