Optimizing Bankruptcy Prediction Through Filterbased Feature Selection

Predicting corporate bankruptcy is a critical task in the financial world. Artificial intelligence techniques have become a promising solution for predicting a company’s condition. One of the crucial phases in the bankruptcy prediction process is feature selection. It is the procedure for choosing from a data set the variables that will be most useful in a prediction model. This study examines and evaluates the important feature selection for machine learning-based bankruptcy prediction. A Taiwan Stock Exchange public dataset, which includes 95 financial ratio features, was used in the study. Several filtering feature selection methods, such as Analysis of Variance (ANOVA), Kendall Tau correlation coefficient (KT), and Mutual Information (MI), were used to identify relevant features. The methods of bankruptcy classification applied were Logistic Regression and Random Forest. According to the test results, ANOVA produces the highest recall value for the Taiwan Stock Exchange dataset, both with the Logistic Regression (LR) and Random Forest (RF). Using 61% of the available features or 58 features, the bankruptcy prediction model with Random Forest performed well without compromising accuracy, precision, recall, or F1 score.