Exploring the Impact of Regularization to Improve Bankruptcy Prediction for Corporations

Abstract

Bankruptcy prediction for corporations is highly essential in today’s fast growing global economy for various reasons, including risk management and financial sustainability. For several years, credit agencies have used statistical methods like regression and discriminant analysis to assess the probability of bankruptcy. However, as Deep Learning and Neural Networks are gaining more momentum to solve more challenging problems, we are turning our attention towards them to address our immediate problems. In this paper, we attempt to explore and apply the working of various neural network methodologies, including the basic architecture, application of regularization techniques, such as L1, L2, Dropout and Early Stopping, to observe the difference in performance for predicting bankruptcy. Other machine learning algorithms such as SVM, Random Forest and XGBoost have also been implemented to compare their performance with neural networks. The results achieved in terms of accuracy were as follows; 82%, 49%, 89%, 90% and 94% for ordinary neural network model, L1, L2, Dropout and Early Stopping methods respectively. Other models, SVM, RF and XGBoost showed an accuracy of 87%, 86% and 85% respectively.