Working Paper: Predicting distresses using deep learning of text segments in annual reports

Working Paper - November 2018 - No. 130

Authors Hansen, Casper (Københavns Universitet); Hansen, Christian (Københavns Universitet); Matin, Rastin (Danmarks Nationalbank); Mølgaard, Pia (Danmarks Nationalbank)
Subject Credit risk; Risk management
Type Working paper
Year 2018
Published 15 November 2018
We develop a probability-of-default model for Danish corporate firms based on deep learning that employs the managements' statements and auditors' reports of the annual reports in addition to the numerical financial variables. Our results show that the text segments provide a statistically significant enhancement of the prediction accuracy compared to models that do not employ the text segments, in particular for large firms. Our results furthermore show that the auditors' reports contain more relevant information than the managements' statements.