Forecasting enterprises bankruptcy by extreme gradient boosting

The application of models for forecasting bankruptcy of enterprises for controlling investment is the basis for monitoring activities of financial institutions. A crucial factor in allowing financial institutions to determine the amount of capital to cover credit losses is the accuracy of the forecast. Most studies use traditional statistical methods (for example, linear discriminant analysis and logistic regression) to build models of enterprise bankruptcy forecasting, but the accuracy of these models is usually quite low. The reason for that is the imbalanced nature of training data sets (the share of bankrupt firms is a small percent of the total number of firms). Nowadays, such machine learning methods as the random forest and the gradient boosting are becoming widespread. This study focuses on the use of extreme gradient boosting to predict bankruptcy. Extreme gradient boosting, using LASSO or Ridge regularization, penalizes complex models to avoid overfitting. Also, during training, extreme gradient boosting fills in the missing values of the data set, depending on the value of the loss. In this article, we proposed SMOTE technique to enhance the minority class of the training data sets, which helps to improve the performance of extreme gradient boosting. The experiment results of improved extreme gradient boosting are compared to the outcomes obtained by other methods.