Modelització dels risc de crèdit a través del machine learning

Abstract

[en] EBA (European Banking Authority) is requiring the banking institutions to show the procedures they are carrying out, alongside the models that they are using, to estimate risk metrics, particularly, the credit risk. Against this background, the financial institutions are finding it difficult to justify how the Machine Learning Methods work. This gives rise to the following question: Is it possible to justify the Machine Learning Methods as applied to the credit risk estimation through Mathematics? Aiming at answering this question, we study several Machine Learning methods, both from a mathematical perspective and their reals applications. The structure of the work is divided into four parts. In the first part, we describe in detail the main issue to be studied, that is, in the event of a customer applying for a loan, the model should be able to predict whether the loan should be granted or not. We illustrate our models with a real database from a financial institution. The second part of this research is devoted to the theory of the following Machine Learning Methods: Logistic regression, Classification Trees (including Bagging and Random Forest), and Boosting (Adaboost). The third part is a practical application of the above models, using the statistical software “R”. We train models on a subset form our database, and assess the discriminatory capacity against the new observations. Finally, we analyze the results obtained, propose future research areas and draw final conclusions.