Machine Learning and Data Balancing Methods for Bankruptcy Prediction

Abstract

The paper examines the use of various machine learning algorithms for the task of forecasting the company’s bankruptcy based on financial indicators. Different approaches to the formation of the data set on which the models are trained are compared, in particular, data balancing methods. Nine machine learning algorithms are implemented, in addition five data balancing methods (random oversampling, SMOTE, ADASYN, random undersampling, and near miss) were applied to classification tasks. It was found that bagging and random forest together with Near-Miss and Random under-sampling showed the best results in terms of the possibility of identifying bankrupt companies in small samples, while artificial neural networks and decision tree methods, together with SMOTE and random resampling, worked better on large samples. With highly unbalanced data accumulation, both small and large training samples can be used to distinguish between bankrupt companies.