Essays in Textual Analysis Using Modern Machine Learning Methods in Financial Economics: From Traditional Deep Learning to Large Language Models

Abstract

The integration of textual analysis and machine learning allows financial economists toassess how sentiment and information from sources like social media and earnings callsinfluence market dynamics. Recent advances in machine learning have greatly improvedthe ability to detect and interpret textual signals that affect financial outcomes.These essays explore how textual signals shape the behaviour of cryptocurrencyand stock markets. Chapter 2 examines the relationship between Twitter investorbehaviour and cryptocurrency performance. Applying our constructed CryptoBERT,we construct a sentiment index and regress it against crypto returns using GARCHmodels. The analysis indicates that changes in sentiment are associated with higherreturns and lower volatility, thereby illustrating the hold-more effect and underscoringthe importance of investor sentiment in cryptocurrency markets.Chapter 3 analyses the association between CSR disclosures and stock returns. Thischapter introduces our constructed CSR-BERT to generate CSR disclosure scores.These scores are employed to construct portfolios and create dummy variables forpredicting future excess returns. The empirical evidence shows that SRI portfolioswhich comprise stocks with non-zero or high CSR scores achieve abnormal returns.In addition, firms that disclose CSR information and those with high CSR disclosurescores exhibit a positive relationship with future stock returns. This highlights thefinancial value of ethical conduct and transparent CSR communication.Chapter 4 explores how corporate communication corresponds to firm value andfirm-specific risk. Using DeepSeek-generated communication scores and PCA, weextract two communication factors: credibility and opacity. The results show thatcredible disclosures align with higher firm value and lower idiosyncratic risk, whereasopaque disclosures exhibit the opposite pattern. These findings demonstrate thepotential of open-source LLMs to extract meaningful stock market insights fromcorporate narratives.Overall, this thesis contributes to the growing literature on the application of machinelearning in textual analysis within financial economics. It demonstrates how differentmachine learning architectures, from a traditional deep learning model to large languagemodels, can be applied effectively to financial texts, offering new insights into investorbehaviour, corporate responsibility, and corporate communication.

Date of Award	2026
Original language	English
Awarding Institution	University of Bath
Supervisor	Nikolaos Sakkas (Supervisor) & Ron Lavi (Supervisor)