Evaluating the impact of machine learning models in SME credit risk assessment

  • Eugeniu Litvinenco (Student)

Student thesis: Master's Thesis

Abstract

In recent years, there has been a significant increase in research efforts to incorporate machine learning (ML) models into credit risk assessment. This study focuses on the credit risk assessment of small and medium-sized enterprises (SMEs), which represent a significant source of employment in most economies. According to the regulator, while ML methods can provide added value, through a more accurate assessment of capital requirements, thus facilitating the access of this segment to financial services, there is a gap in the implementation of these methodologies in real-world context. This can be due to the overall complexity of explainability and interpretability, which forces financial institutions to use simpler models. Another reason is the lack of clarity on the benefits that these methodologies can provide. In this study, a hybrid model combining a decision tree and a logistic regression is proposed to address the complexity problem. This model shows comparable performance to the Random Forest and XGBoost while providing interpretability complexity equivalent to a logistic regression. In addition, this study introduces two innovative metrics, Exposure Weighted Distance to Default (EWDD) and Exposure Weighted Rating (EWR), which aim to find a way to distinguish the misclassifications made by a model according to their capital significance and to provide a sense of the total capital requirements that a model can generate. These metrics, along with the commonly used, were employed to compare the models, enabling the financial institutions to make a more informed decision in selecting the model that best meets their objectives.
Date of Award23 Jan 2024
Original languageEnglish
Awarding Institution
  • Universidade Católica Portuguesa
SupervisorNicolò Bertani (Supervisor)

Keywords

  • SME
  • Machine learning
  • Credit risk
  • Hybrid model
  • Exposure weighted metric

Designation

  • Mestrado em Análise de Dados para Gestão

Cite this

'