Textual Classification of SEC Comment Letters

Abstract

I utilize Naive Bayesian text classification to signal important SEC comment letters, where negative abnormal returns following comment letter disclosure is the measure of importance. In a holdout sample, classification identifies important comment letters between 10 and 40 percent better than chance. The average market response to signaled comment letters is a -5.8 percent abnormal return over the subsequent 90 days, but only when the comment letters were viewed on EDGAR, indicating market underreaction to these disclosures. Signaled comment letters are associated with lower persistence of profits and increased levels of restatements in the year following comment letter disclosure. Together these results suggest that text classification can be used to signal important comment letters and that these letters are associated with with lower future performance and undisclosed financial reporting deficiencies.