Structured Paragraph Embeddings of Financial Earnings Calls.

Financial earnings calls contain rich information about the quarterly performance and future projections of public companies. Such information is highly relevant to developing trading strategies and understanding economic trends. However, due to the unstructured nature of call transcripts important signals can be difficult to extract. In this preliminary work, we propose a novel paragraph embedding method that leverages the structure inherent in the Q&A format of earnings calls. We show that the proposed method improves classification performance over more general methods and provides a useful measure of similarity between paragraphs.