An experimental survey of imbalanced learning algorithms for bankruptcy prediction

Abstract

Information about imminent bankruptcy is crucial for financial institutions, decision-making managers, and state agencies. Since bankruptcy prediction is a prevalent research topic, many new methods have been continuously proposed. Bankruptcy prediction is frequently approached as a binary classification task. Since bankruptcy datasets are inherently imbalanced, bankruptcy classification is usually performed using class imbalance learning methods. The nature of these methods is very diverse, but they can usually be categorized as ensemble, cost-sensitive, sampling, and hybrid methods. In this paper, we provide a comprehensive experimental comparison of 45 methods. These methods were selected because they cover the approaches and algorithms frequently employed for bankruptcy prediction and imbalanced learning. Extensive experiments on 15 publicly available datasets with different imbalance ratios showed that the methods based on a combination of ensemble learning and undersampling are able to handle data imbalance and achieve the best results for bankruptcy classification.