Tracking Topics in Earnings Call Transcripts Using Natural Language Processing

Abstract

This thesis explores the incremental value that Large Language Models (LLMs) can provide compared to existing bag-of-words methodologies to automate the reading of earnings call transcripts and create topic mappings that help executives, investors, and policymakers make more informed decisions. We run several small-scale experiments to assess the abilities of LLMs to classify texts and concluded that using standalone LLMs to classify portions of text is not the most optimal approach. Instead, we propose a hybrid approach that leverages LLMs to generate keyword lists, which are subsequently applied within a bag-of-words framework, enabling us to effectively map 30 distinct topics across both the presentation section and the questions earnings calls. We also explore how structural and contextual knowledge could be applied to enhance both LLM and bag-of-words methodologies for topic mapping. Using a dataset comprised of earnings calls from S&P 500 companies from 2011 to 2025, we analyze trends in specific topics mentioned on calls, particularly focusing on tariffs. Furthermore, we present a detailed case study of Dollar Tree to illustrate how our topic mapping approach can effectively inform decision-makers.