Benchmarking machine learning models for credit risk: performance and interpretability

Summarise with VIRAT
Authors: Karel Štícha
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Accurate credit default prediction is essential for banks as they operate under the strict Basel regulatory framework, where internally estimated probability of default directly affects capital requirements and lending capacity. While traditional statistical models such as Linear Discriminant Analysis (LDA) and Logistic Regression (LR) remain widely used mainly due to their transparency and regulatory acceptance, machine learning models often offer greater modelling flexibility but raise important questions regarding interpretability, which is a key concern in credit risk management. This thesis presents a controlled comparison of Bayesian Additive Regression Trees (BART) as a credit risk model against LDA, LR, and widely applied machine learning models in credit risk literature: Classification and Regression Trees (CART), Random Forest (RF), and Support Vector Machines (SVM). Model performance is assessed using standard classification metrics, measures of discriminative ability, and probabilistic performance. In addition, local Explainable Artificial Intelligence (XAI) techniques, namely LIME and Anchors are examined to address transparency considerations. The empirical results indicate that linear models including linear SVM outperform more complex non-parametric approaches. This finding, together with results of previous research works, and the weaker performance of the non-linear SVM specification suggest that the underlying risk structure of the used dataset is predominantly additive. Although BART does not exceed the performance of linear models, it consistently outperforms RF, which is often regarded in prior literature as a strong benchmarking standard. This suggests that BART’s Bayesian regularization and additive tree structure may provide advantages over traditional tree-based ensembles. Overall, while linear models remain highly effective in additive settings, BART represents a promising alternative to conventional machine learning models, particularly in more complex credit risk environments. Interpretability, however, remains a significant challenge. Although Anchors proved very stable and reliable, neither evaluated XAI methods fully resolved transparency concerns for practical regulatory acceptance. These finding motivate further academic investigation into both BART’s performance in non-linear settings and the development of robust XAI techniques.

Fields of Study & Subjects

Financial Distress and Bankruptcy Prediction Explainable Artificial Intelligence (XAI) Imbalanced Data Classification Techniques

Publication details

Journal / source
University Library Linz repository (Johannes Kepler Universitat Linz)
Publication date
1 January 2026
Language
EN
Document type
Thesis