Machine-Learning about ESG Preferences: Evidence from Fund Flows

Abstract

We construct a broad dataset of Environmental, Social, and Governance (ESG) scores for equity mutual funds based on the stocks they hold and stock-level scores from six prominent ESG data providers. We find that many ESG scores predict fund flows despite substantial disagreement among providers. We then use machine learning to concentrate the information in scores and generate accurate flow forecasts (ESG-driven flows). We use ESG-driven flows to proxy for a fund’s ESG performance and find that better-performing funds realize higher flows, lower returns, and hold stocks with lower returns. Investors also pay $11 million/year more for a top ESG fund.

Keywords: ESG, Fund flow, Value-added, Machine learning

JEL Classification: G10, G23