Classifying Restatements: An Application of Machine Learning and Textual Analytics

Abstract

Restatements of audited financial statements are used for evaluating reporting quality, audit quality and for other evaluative purposes. Prior research shows that restatements that correct unintentional errors have different implications for statement preparers, users, auditors and regulators than restatements that correct intentional misstatements. However, manually classifying restatements into these categories can be tedious, time-consuming and inconsistently performed. Therefore, we constructed a Naïve Bayes machine learning algorithm to classify restatements by management intent based on the language in restatement announcements. Empirical tests of the algorithmically classified restatements show that this classification is an effective, efficient alternative to manual classification and more reliable than other commonly used automated methods such as classifying based on restatement direction or magnitude. Our method does not require a dictionary of words associated with management intent, is easily replicated and scalable and may be used to classify restatements disclosed at the same time as financial results.

Keywords: restatements, unintentional error, intentional misstatement, textual analysis, machine learning, naïve Bayes, data analytics

JEL Classification: G38, M41, M42, M48