ML-based Credit Scoring Models

Background

Credit risk models are useful tools for modelling and predicting individual firm default. They are greatly used by banks and Fintech credit providers, for the purpose of assessing the risk of default of applicants for loans.

Credit scoring is a classification problem. The approach takes a vector of attributes for a loan applicant, and given these attributes, attempt to discriminate between the classes.
An accurate and well-performing credit scoring models enables banks and Fintech credit providers to control their risk exposure through the selective allocation of credit based on the empirical analysis of the data provided by the applicant.

The literature offers many studies on the application of ML methods for improving the predictive accuracy of credit scoring models. In addition to the higher predictive performance, further arguments in favor of applying ML algorithms in credit rating can be found in the inherent properties of these models. Namely, they: (i) explore the structure of a system without pre-assumption on the data, (ii) allow for non-linearity and high-order interactions between factors, and (iii) are particularly useful for wide data (number of input variables to exceed the number of subjects), etc.

This topic becomes even more relevant in view of the fast emergence of Fintech credit providers that rely extensively on novel methodologies for the purpose of carrying out their credit scoring tasks.

Expected outcome:

• Develop a strategy for dealing with imbalanced classes in classification problem sets;
• Examine the performance of different ML classifiers;
• Quantify the difference in classifiers’ performance
• Identify the classifier that performs best