Class Imbalance Deep Learning for Bankruptcy Prediction

This paper addresses one of the most burning issues among financiers namely bankruptcy prediction. For Bankruptcy prediction many researchers have used various methods ranging from Statistical Modeling to Machine Learning. Training a robust Machine Learning model to accurately predict whether a company goes bankrupt or not is a challenging issue in the sense that real life data is often imbalanced. Hence, we first reduce the imbalance nature of data and then train a Deep Neural Network with the balanced data. The dataset used is that of Polish companies which consists of five years of data corresponding to five different tasks. We reduce class imbalance using an Oversampling method known as Synthetic Minority Oversampling Technique (SMOTE). Our model significantly outperforms the previous neural network models and weak learners trained on this dataset in terms of Area Under Receiver Operator Characteristic Curve (AUC).