Predicting bankruptcy using machine learning
Date
2025
Authors
Idrees, Muhammad Qasim
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Bankruptcy is a serious issue in the financial sector. ML-based approaches have been widely used to predict bankruptcy, but selecting the right classifier and identifying the key features influencing Model results is essential for achieving good performance. This is especially important in the corporate sector where decision-makers may have a limited understanding of these models and their function. There is also a need for transparency and trust in their outcomes. This thesis utilizes twelve classifiers: CatBoost, RF, SVM, KNN, NB, LR, BDT, Stacking Ensemble, AdaBoost with RF, NN, FT-Transformer, and TabNet. These classifiers are evaluated in various settings using the Polish companies dataset to predict bankruptcy. The results obtained show that CatBoost is the best performing classifier. Further, SP-LIME is used to identify features that impact model decisions. SP-LIME ranks features based on their importance. A comparative analysis of SP-LIME and PCA indicates that SP-LIME identifies more influential features.