Dead, Alive and In Between: An Investigation of Multinomial Bankruptcy Prediction Models

Abstract

The purpose of this thesis is to investigate whether the estimated accuracy of binomial and multinomial bankruptcy prediction models is comparable. First, we seek to understand how the utilized variables impact corporate bankruptcy. This investigation is accomplished by estimating logit and penalized logit models. Our results show innovation, macroeconomic, and financial variables are significant to describe corporate bankruptcy models. To estimate accuracy, we develop machine learning models, which include logistic regression, decision tree, support vector machines, artificial neural networks, and stacking models. We compute the prediction accuracy of the binomial bankruptcy models to be significantly higher compared to the multinomial bankruptcy models for all machine learning models except the logistic regression. Finally, we estimate there to be no significant difference in prediction accuracy between the stacking model and the highest performing base learner. This thesis contributes to the literature by estimating the implication of expanding the definition of bankruptcy from a binomial to a multinomial interpretation.