Reality Hits Early Warning System: Based on Unsupervised Isolation Forest Anomaly Detection

Over the last few decades, the supervised machine-learning-based early warning system (EWS) has grown in popularity to signal more accurate bank and corporative vulnerabilities. Most of these models are built based on a significant amount of labelled data with non-bankruptcy and bankruptcy; however, many real-world cases do not conveniently have a significant amount of labelled data and often have an extremely skewed distribution due to the rarity of bankruptcy events. We introduce an isolation forest anomaly-detection model to construct an EWS based on no label with an extremely small portion of bankruptcy data and analyse the effectiveness of the EWS built using an unsupervised machine learning framework. Many early warning studies have not explored cases where limited data are available, and to the best of our knowledge, unsupervised learning and anomaly detection are unexplored fields in the EWS literature. Our empirical study shows the possibility of building a significantly accurate EWS using only unlabelled and skewed data. Moreover, we also discuss the reality and limitations of EWS; that is, it is difficult to distinguish actually bankrupt companies from those predicted as bankrupt using an EWS. In particular, this is the case for corporative EWS, because there are more active companies than bankrupt companies, and it is difficult to build an accurate corporative EWS.