(Georgia Institute of Technology, 2024-04-25)
Du, Wendi
The dissertation consists of five essays on FinTech, AI and innovation in finance. These essays center around how innovation and capital market influence each other, and how to use cutting-edge technologies like machine learning and AI to address the economic questions that would otherwise remain unanswered. In the first essay, I investigate the redeployability channel of trademarks' collateral value. Using a novel court decision that exogenously weakens trademark redeployability, I find a 3.4 percentage point reduction in affected firms’ book leverage, equivalent to a 16.9\% decrease in their average book leverage. By using firm-level trademark portfolio data and employing natural language processing (NLP) techniques, including ChatGPT, I show that firms with more licensed trademarks (i.e., those more exposed to the court ruling), experience a stronger negative impact. Additionally, affected firms are less likely to pledge their registered trademarks as collateral afterward. When they do pledge, they pledge a greater number of trademarks, as well as more valuable ones. Affected firms also register fewer new trademarks in the future. In sum, my results highlight the value of trademark collateral in enhancing firms' debt capacity through its redeployability channel.
In the second essay, we develop a text-based measure of firm-level inflation exposure from earnings calls. Our deep learning model identifies sentences discussing price changes, while distinguishing price increases from decreases and inputs from outputs. Our aggregate inflation exposure measure strongly correlates with official inflation measures. Firms with higher inflation exposure experience negative stock price reactions to earnings calls. The price reaction is attenuated when a firm has pricing power. Further, firms with higher inflation exposure have higher future costs of goods sold and lower operating cash flows. They perform worse on Consumer Price Index (CPI) release days when CPI exceeds the consensus forecast.
In the third essay, we explore the area of emerging technologies which can potentially transform business and society but are difficult to identify and prone to hype and uncertainty. We construct a dictionary of emerging technology phrases from earnings calls using deep learning techniques and document an immediate positive stock market reaction to firms’ discussions of emerging technologies. We find that the positive reaction is more pronounced when firms discuss emerging technologies early in their life cycle. Firms with lower ex-ante credibility, such as a prior history of earnings management, innovate less ex-post and experience poorer long-term returns. Overall, our results highlight when firms' discussions of emerging technologies convey credible information to investors.
The fourth essay examines whether managers walk the talk on the environmental and social discussion. We train a deep-learning model on various corporate sustainability frameworks to construct a comprehensive Environmental and Social (E\&S) dictionary. Using this dictionary, we find that the discussion of environmental topics in the earnings conference calls of U.S. public firms is associated with higher pollution abatement and more future green patents. Similarly, the discussion of social topics is positively associated with improved employee ratings. The association with E\&S performance is weaker for firms that give more non-answers and when the topic is immaterial to the industry. Overall, our results provide some evidence that firms do walk their talk on E\&S issues.
In the final essay, we address the limitations of generic training schemes in the realm of financial language models. We propose a novel domain specific Financial LANGuage model (FLANG) which uses financial keywords and phrases for better masking. We further extend it to include span boundary objective and in-filing objective, utilizing the fact that many financial terminologies are phrases. Additionally, the evaluation benchmarks in the field have been limited. To this end, we contribute the Financial Language Understanding Evaluation (FLUE), an open-source comprehensive suite of benchmarks for the financial domain. These include new benchmarks across 5 NLP tasks in financial domain as well as common benchmarks used in the previous research. Experiments on these benchmarks suggest that our model outperforms those in prior literature on a variety of NLP tasks.