Feature selection for predicting bankruptcy: Comparative analysis

Company Bankruptcy prediction has been a major concern for many company stakeholders for decades. A firm could possibly be in troubles in terms of money due to internal and external situations and it may impact the firm immensely. Due to the large number of factors that contribute to bankruptcy, relying on a single predictive model is insufficient, and the challenge comes in recognizing only the essential attributes that are more liable. And the substantial class imbalance in the data, which impairs the performance of the model, is another significant barrier. Despite the fact that several strategies have already been examined with different pre-processing techniques, there still seems to be room for improved techniques. This paper enhances this research by using a novel combination of a variance threshold, mutual information technique and Pearson correlation along with SMOTE (Synthetic Minority Oversampling Technique) on Taiwanese dataset. On the altered data, four classifiers are implement, Random Forest, Decision Tree, AdaBoost, and XGBoost and assess their performance.