Predicting Earnings Management from Qualitative Disclosures

Abstract

While analysts, customers, and lenders rely on financial disclosures to make decisions regarding a company, executives often manage the disclosed earnings. Detecting such practices is thus a concern for company stakeholders and regulators. Qualitative disclosures are an additional source of information about a company's financial situation, but executives likely attempt to hide their earnings management activity in these disclosures, as well. We use supervised machine learning models to predict earnings management by property and casualty insurers from the Management’s Discussion and Analysis filings. For this, we utilize a new algorithm that interprets textual data conditional on the reported financial situation of the company. We show that the qualitative disclosures can predict earnings management, revealing that executives are unable to remove all subliminal messages from them. The results demonstrate that qualitative disclosures can be useful for learning about the accounting choices of companies.

Keywords: Machine learning, Earnings management, Reserve error, Disclosure

JEL Classification: C38, C53, G22, M41