Topic Modeling in Financial Documents

This paper describes the application of topic modeling techniques to quarterly earnings call transcripts of publicly traded companies. Earnings call transcripts represent an interesting case for analysis because the document is relatively unstructured and potentially more informative than 10K and 10Q disclosures due to the question and answer session consisting of unprepared statements. This paper addresses the clustering of these documents as well as the segmentation of individual documents into clusters for products and industries the company is active in. The goal is for each transcript to be assigned to some number of topics, and the specific segments of the transcript which address a given topic to be specified as well. Thus, not only will the documents be classified as covering some set of topics, but the documents themselves will be partitioned into different sub-topics. I will discuss progress I made in achieving these goals as well as challenges and issues which remain. This work could prove useful in financial document summarization as well as improving search and display of documents and information relevant to a user’s search and interests. Furtheremore, applying NLP and machine learning concepts to financial document analysis is increasingly being used by trading firms and hedge funds to gain competitive advantage.