Predicting Short Term Stock Prices Based on Company Earnings Calls

Predicting short-term stock price movements is a challenging yet crucial task in the field of finance, particularly given the significant influence of investor sentiment following a company's quarterly earnings calls. This thesis explores the application of various machine learning techniques to predict stock price fluctuations based on the textual analysis of earnings call transcripts. By leveraging advanced natural language processing (NLP) methods, including BERT embeddings and sentiment analysis using the Loughran-McDonald and Afinn dictionaries, this research aims to extract meaningful features from the earnings calls to enhance prediction accuracy. The study involves constructing Document Term Matrices (DTMs) and employing Term Frequency-Inverse Document Frequency (TF-IDF) to represent textual data, followed by the application of multiple machine learning models such as Support Vector Machines (SVM), Random Forest, and Naive Bayes. Additionally, t-Distributed Stochastic Neighbor Embedding (t-SNE) is used for data visualization to uncover underlying patterns and clusters within the transcripts. Results indicate that while traditional models achieve moderate accuracy, the integration of BERT embeddings shows potential for capturing deeper contextual information, although it did not significantly outperform other methods in this study. The findings highlight the importance of both prepared statements and Q&A sections in influencing investor sentiment and subsequent stock price movements.