Bankruptcy Prediction Models: A Machine Learning Approach

In a world where everything happens faster than it used to, where it’s often difficult to grasp the importance of the events that take place around us before their consequences are inevitable, where new technological advancements often make us feel small and inadequate and are perceived as a menace to our workplaces and our abilities as a whole, it’s mandatory that we learn how to exploit the tools that this ever-changing world offers, and that we find out how to spot an opportunity where some may see a threat. Effective bankruptcy prediction models are part of these tools. Bankruptcy prediction has now been studied for almost a century, starting from the first studies conducted by FitzPatrick in 1932 based on simple ratio analysis, until the hundreds of Machine Learning and Deep Learning models that scholars all over the world are still researching and refining. Financial institutions, lenders and governments seek to develop accurate models that correctly identify and quantify the risk of default of their counterparties. Firms themselves may benefit from models that help them pinpoint their weaknesses and take corrective actions before it’s too late to save the damage. The aim of this thesis will be to revise all relevant literature in the field of bankruptcy prediction and to explain in detail the functioning of the most widely used models. Moreover, we will apply some of these models, specifically Logistic Regression, Support Vector Machines, Artificial Neural Networks, Decision Trees, Random Forests, AdaBoost, XGBoost, LightGBM and NGBoost, to a dataset of 4,415 Italian companies to assess which model performs best in predicting the status (active or failed) of the companies and to better observe the advantages, but also the shortcomings, that come from applying the different models.