TEXT

Abstract

In an era with increasing amounts of data and readily available computer power, financial markets have become increasingly competitive for investors seeking to systematically extract actionable information from data. This thesis uses a combination of structured financial data and unstructured text from earnings call transcripts, with the goals of identifying the direction of a firm’s abnormal return surrounding earnings calls and assessing how management and analyst behavior on these calls vary across contexts. Utilizing a corpus of 96,000 quarterly earnings call transcripts representing over 4,000 unique firms, we train classification algorithms to predict immediate abnormal equity market returns. This thesis leverages multiple types of data including features that represent call structure (attendance measures and number of speakers), call tone (relative proportion of tonal language), and financial information including past stock performance and fundamental performance measures. Rather than directly using text, as previously implemented in past research, this thesis attempts to measure how effectively quantitative summaries of calls predict abnormal direction. By implementing a varied set of widely used classification algorithms, including, Naïve Bayes, Logistic Regression, Decision Trees, Ridge Classifier, Support Vector Machines, and K-Nearest Neighbors, we are able to compare classification performance and gain insights into what call features are most informative. In particular, this thesis compares the performance of pure text sentiment analysis to classification using objective call features, linguistic measures of tone and complexity, and financial attributes.