Reading Between the Lines: A Text-based Deep Learning Approach for Understanding Company Dynamics

The Management Discussion and Analysis (MD&A) section in Form 10-K offers highly valuable insight of a company’s fiscal health, operational performance, and future outlook. In today’s ever-changing business environment, both practitioners and academics have extensively studied this textual data, with particular interest in quantifying the evolving information encoded in MD&A narratives. In this work, we challenge the traditional cosine similarity-based method for measuring differences between a firm’s year-over-year MD&A disclosures. We propose a novel method, D3, short for Deep Learning Method for Disclosure Differences, that incorporates external expert evaluation signals, in the form of analyst recommendation updates, to guide the learning of variant information between consecutive MD&A reports. Instead of producing a dissimilarity score as in traditional cosine distance approaches, D3 learns the changing information directly as a variant vector. The variant vector derived by D3 serves as a multifunctional artifact for various downstream financial applications. In our experiments, we show that the variant vector demonstrates strong economic utility. Specifically, a long-short portfolio strategy that sorts firms based on the magnitude of their variant vectors generates significant excess returns, after controlling for common risk factors. Moreover, the variant vector exhibits predictive power for financial risk-related outcomes. We further introduce an interpretability technique based on D3 to uncover key evolving information disclosed in MD&A reports. Our work has important implications for information systems computational design research and the emerging FinTech literature.