Bankruptcy prediction models: A comparative analysis

Bankruptcy prediction models are one of the most reliable and useful quantitative tool in financial decision-making performed by financial institutions. This dissertation will review, firstly, the vast literature in this field, distinguishing and analyzing the main theoretical characteristic of statistical and machine learning models. Secondly, by retrieving European Union data through the Orbis database, this research implements and compares eight model performances at one, two and three years prior the bankruptcy by coding in the Python language through the Jupyter Notebook application. The empirical part tries to be as realistic as possible: i) without any ex-post matching procedure that could fictitiously increase the performance of the models and ii) using vectors of only three explanatory variables taken for three consecutive years to constraint and prevent overfitting of the models. The main innovations that this research brings in the bankruptcy prediction field are the SMOTE and Tomek Link approach used to solve the imbalanced dataset problem and the hyperparameter tuning made through a randomized and grid search for each model to find the best parameters for the data sample.