Anomaly detection methods for bankruptcy prediction

Bankruptcy prediction is remarkable for its importance in modern economics. Many related work have proposed novel supervised learning methods for predicting financial distress. Conventional classification algorithms are likely to mistaken minority class for majority class because their cost function simply optimizes error rate. This paper applies anomaly detection algorithms to the bankruptcy prediction problem in an attempt to suggest a new stable model taking the data distribution into consideration. The efficacy of anomaly detection techniques is tested on bankruptcy prediction datasets viz. Polish banks. Our empirical evaluation shows that Isolation Forest outperforms multivariate Gaussian distribution, one-class SVM, and other classification estimators in terms of the ROC curve.