Bankruptcy Prediction Using a GAN-based Data Augmentation Hybrid Model

Abstract

In order to lower the danger of a company failing, research in the area of bankruptcy prediction is still being done. New effective models are being developed employing a variety of cutting-edge methodologies. However, the majorities of bankruptcy databases are unbalanced and may include unnecessary data. So, creating a powerful, trustworthy model to improve prediction is always a difficult undertaking. We made the forecast in this paper in three stages. In the first stage, we concentrated on balancing the datasets using two well-known methods, SMOTETomek and GAN (Generative Adversarial Network), which generate synthesized data. Then, in the second phase, a selection of pertinent features was extracted using three wrapper-based feature selection methods: step forward feature selection, backward elimination, and recursive feature elimination, as well as five filter methods: dropping constant features, feature selection based on correlation, information gain, Chi-square test, feature importance. These three ANN, CNN, and LSTM models have been used for the third step of actual prediction. After obtaining pertinent information by feature selection from both sampling approaches, the results show that the ANN model has a better capacity for prediction than the other two predictive models. It has been demonstrated that the GAN technique outperforms the SMOTETomek with respect to all three predictive models.