Determinants of Consumer Credit Default in Romania: A Comparison of Machine Learning Algorithms

Prof. Dr. Monica Dudian and Ana-Maria Sandica investigate the separation power of several machine learning techniques and compared them with the benchmark logistic regression using real data from 17520 private individuals of a Romanian commercial bank. The result of their research can be read in the book “Social Credit Rating“.

In order to capture the financial crisis effect they equally divided the data in two samples prior and posterior the crisis and we compared 13 models in terms of misclassification Type I and II Errors. As the models aim to catch best the patterns in the “default” profile of a consumer credit borrower, they split the variables in socio-demographic factors (Social Rating) and financial factors (Financial rating) and conclude that “default” profile prior crisis is captured better by the linear models while the patterns of the financial crisis are captured better by the non-linear models.

Monica Dudian and Ana-Maria Sandica found that the accuracy ratio gives the better results on decision trees and ensembles based on decision trees such as adaptive boosting methods (Financial Rating) and Random Forest (Credit Rating, Social Rating) irrespective of the sample choice.

The power of the model to classify the debtors using Social Rating, Financial Rating and the mix of these, the Credit Rating, depends on the trained data used. The Financial Rating’s champion model’s results are best on posterior crisis data, meaning that financial factors counted the most in detecting the patterns in “default” after the financial crisis. The order is not the same for Social Rating, where the best classification is obtained on prior crisis data meaning that classification considering the individual’s creditworthiness is more difficult on posterior crisis “default” patterns.

Prof. Dr. Monica Dudian is Professor of Economics at The Bucharest University of Economic Studies, where she received her PhD in Economics in 1999. She also held the position of Vice dean of the Faculty of Economics of The Bucharest University of Economic Studies, during 2001 – 2008. Her teaching is focused primarily on microeconomics and industrial organization. She manages research grants and performs research on country risk, credit risk, and industrial economics.

Dr. Ana-Maria Sandica has been developing credit risk models for more than 10 years. She started to study machine learning techniques during her master degree in Financial Econometrics (Dofin) and continued with completing a doctorate in the field of stochastic equilibrium models in Macroeconomy. Her thesis on postdoctoral research links the macroeconomic shock transmission mechanism in estimating the probability of bankruptcy for companies. She held a managerial role in model risk validation at a major German bank.

Create a website or blog at

%d bloggers like this: