Ensemble classifiers for bankruptcy prediction using SMOTE and RFECV

This research investigates the impact of preprocessing strategies, namely feature selection (utilising correlation and recursive feature elimination with cross-validation) and class imbalance handling (employing synthetic minority oversampling technique), on the performance of prediction models using ensemble-learning techniques (random forest, AdaBoost, gradient boosting decision tree, extreme gradient boosting, bagging, LightGBM and extra tree classifier). The study focuses on the Polish bankruptcy dataset to assess the effectiveness of these preprocessing approaches. Experimental results demonstrate that adopting class imbalance handling significantly influences classifier performance compared to feature selection alone. Interestingly, hyperparameter tuning and feature selection exhibit limited impact on classifier performance. Among the ensemble-learning techniques tested, the adaptive boosting classifier shows consistently poor performance throughout the study period, followed by the bagging classifier with statistical significance. These findings shed light on the importance of selecting appropriate preprocessing strategies to improve the performance of ensemble-based prediction models in bankruptcy prediction tasks.