Two essays on the application of NLP in financial analysis : firm risk similarity, mutual fund counter-cyclical expectations

Abstract

With the rapid development of machine learning and AI, we are now able to analyze data that was previously difficult to process. In finance, researchers have employed various innovative approaches to proxy important concepts such as risk and belief. Now, we stand in an era that allows us to obtain more direct measures and gain deeper insights into these variables. This enables us to rethink some of the most critical questions in the field of finance. In this thesis, I present two studies that leverage advanced machine learning and AI tools to address important financial questions. In Chapter 1, I use machine learning techniques to extract a comprehensive set of disclosed risk factors from annual reports and construct a risk-based similarity peer network. I then demonstrate that the future returns of firms can be significantly forecasted based on the returns of peer firms with similar risk profiles. A long-short portfolio, sorted by the returns of these risk-similar peer firms, generates a Fama-French three-factor alpha of 84 basis points per month, and this result remains robust across various model specifications. Importantly, this phenomenon differs from conventional industry momentum, as I exclude firms within the same industry when constructing the portfolio. The delayed response of a firm’s returns might be influenced by investor inattentiveness and restrictions on arbitrage opportunities. In Chapter 2, we leverage the unprecedented capabilities of Large Language Models (LLMs) to assess sentiment across different topics within the same text. Using this approach, we extract the structured beliefs of all mutual fund managers in China, based on their reports spanning 2008 to 2023, regarding the economy, government policies, and financial markets. Specifically, we construct a measure for countercyclical policy (CCP) beliefs, where managers anticipate that government policies will counteract economic shocks. We find that funds exhibiting frequent CCP beliefs (CCP funds) significantly outperform other funds. To explain this superior performance, we demonstrate that CCP funds’ market beliefs have stronger predictive power for market returns, their asset allocations show a stronger correlation with their market beliefs, and their performance is directly linked to their time-varying CCP beliefs.

Date of Award	2024
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Yingying LI (Supervisor) & Zhenyu GAO (Supervisor)