A Machine Learning Approach to Corporate Bankruptcy Prediction Using BERT-Based Sentiment Analysis

Abstract

The study of bankruptcy prediction has centred on whether firm level information is predictive. Seminal work by Altman (1968) articulates the failure of a business utilising its financial variables that are associated and classified in part to either the liquidity, profitability, solvency, leverage, or activity of a corporation. While this understanding is intuitive, recent studies have broadened the scope of financial ratios used in this prediction as well as incorporated exterior forces affecting the firm, either at an enterprise-wide or an economic-wide level to predict corporate bankruptcy. In the same breath, one cannot ignore the insider knowledge that the leaders and managers of firms would have leading to corporate bankruptcy. Therefore, this provides a curious opportunity in which we can incorporate the sentiment in the analysis provided by the leaders of such firms as an input in predicting the bankruptcy of a given firm. This study applies the Bidirectional Encoder Representations from Transformers (BERT) based sentiment analysis approach to import human sentiment as a variable from corporate disclosure data and apply it to existing corporate bankruptcy models over the period between 1995 to 2022 in South Africa, the United Kingdom and the United States of America