Uncovering Mutual Fund Private Information with Machine Learning

Abstract

This paper implements natural language processing (NLP) models and neural networks to predict mutual fund performance using the textual information disclosed in mutual fund shareholder letters. Informed funds identified by the prediction model deliver superior abnormal returns and are more likely to receive an upgrade in Morningstar ratings. Informed funds also attract greater flows in three days and up to 24 months after the disclosure of shareholder letters, especially when their disclosure has greater investor attention, suggesting that investors recognize the information from the qualitative disclosure. The machine learning model shows that informed funds tend to discuss sector specializations, portfolio risk taking, big picture of the financial market, and mixed strategies across assets. Collectively, this study shows that mutual fund disclosure contains rich, value-relevant textual information that can be analyzed by state-of-the-art machine learning models and help investors identify informed funds.

Keywords: Machine Learning, Mutual Fund Performance, Disclosure, Textual Analysis, Shareholder Letters, Fund Flows

JEL Classification: C45, G11, G14, G23