Corporate bankruptcy

: can machine learning methods enhance the prediction of failure?

Abstract

This dissertation aims to enhance the performance of traditional corporate bankruptcy prediction models through the application of machine learning techniques and models, and industry effects. The data used includes 3664 companies out of which 144 went bankrupt throughout the period of 2000 until 2019, and it was structured to emulate the design of the variables Campbell et al. (2008) used in their study. Evidence was found that implies the improvement of various metrics’ results from the use of machine learning techniques and models. The model with the highest F1-score, meaning the most balanced, is the Logit with the application of hyperparameter tuning and industry effects. The model with the highest Recall, which means the percentage of bankruptcies correctly predicted, is the Logit with the application of the oversampling technique. Furthermore, both Support Vector Machines (SVM) and Artificial Neural Networks (NN) models delivered balanced and enhanced results compared with the two benchmarks (Altman Z-Score and Simple Logit models). The improvement techniques provided the models with distinct results. Oversampling led mostly to a higher percentage of bankruptcies predicted, while hyperparameter tuning and industry effects provided the models with more precise results. The variable importance in each type of model was also analysed. Overall, the Campbell et al. (2008) market variables (SIGMA, RSIZE, EXRET and PRICE) are highly significant for the positive results of all three types of models studied.

Date of Award	1 Feb 2021
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Dan Tran (Supervisor)