Prof. Kaustav Sen is a Professor at Lubin School of Business Pace University, New York. He has presented at the Public Company Accounting Oversight Board, been awarded funding by the National Stock Exchange of India, has received best paper awards from the American Accounting Association (Ohio) and the Securities Exchange Board of India, and has been cited by the Economist. In the past, he served as the Accounting Department’s Ernst & Young scholar. Read More
Lubin School of Business Pace University, New York.
November 21, 2023
What matters in predicting earnings changes – financial statement numbers or textual disclosures?
Chen et al (2022) develop models using machine learning methods and high-dimensional detailed financial data (XBRL and Compustat) to predict the direction of one-year-ahead earnings changes. These models outperform two conventional models that use logistic regressions and small sets of accounting variables, as well as professional analysts’ forecasts. We explore the role of text, using various NLP metrics of the 10K filings corpus and Finbert sentiment of the MD&A section and earnings call transcripts, for the same prediction exercise. What we find is quite compelling: a simple text model using eleven variables performs as good as kitchen-sink models using large sets of financial data or analysts’ forecasts. We follow the suggestions in Krupka et al (2022) and use the insights from our predictive model building exercise to delve into explanatory modeling. Using stepwise logistic regression and single variable Granger-causality design, we find that text metrics are more significant determinants than financial ratios. Interestingly, while sentiment does play an important role, other attributes of text, such as similarity and corpus size are found to be very important. These findings offer pause to think about information contained in text vis a vis financial statements for predicting binary outcomes in the financial domain.
