Comparative Analysis of Machine Learning Approaches for Predicting Loan Default Risk
Abstract
Banks and other lenders need to know, before giving out a loan, how likely it is that the
borrower will fail to pay it back. This is called credit risk. In the past, banks used simple
statistical methods to check this risk. These methods were easy to understand but not always
very accurate. Newer machine learning models, such as Random Forest and gradient boosting
methods, can often predict default more accurately. However, these models are harder to
explain. A bank cannot easily tell a customer why they were rejected if the model works like a
black box. This creates a problem: should a lender choose a model that is more accurate, or one
that is easier to explain? Most past studies only test one model on one dataset, so it is hard to
know if the results would stay the same on different data.
This thesis tries to solve both problems together. Five machine learning models are used:
Logistic Regression, Support Vector Machine (with an RBF kernel), Random Forest, XGBoost,
and CatBoost. These models are tested on two different loan datasets. The first dataset has
6,843 loan records, with about 35% marked as default. The second dataset has 614 loan
applications, with about 31% marked as not approved. All five models are trained and tested
using the same method, called nested cross-validation, so the comparison between models is
fair. After training, a tool called SHAP is used to explain which features (such as income or
credit score) affect each model's decision the most.
On the first dataset, all five models perform in a similar range, with ROC-AUC scores between
about 0.73 and 0.74. CatBoost scores the highest (0.743), but the difference between models is
small. Across all models, Income, Credit Score, and Loan Amount are found to be the most
important features. On the second dataset, results are noticeably better overall. Random Forest
gets the best ROC-AUC score (0.754), and SVM-RBF gets the highest accuracy (80.8%). In
this dataset, one feature stands out far more than the rest: Credit History. This matches what is
already known in banking a person's past record of paying credit is one of the strongest signs
of future default risk.
These results suggest that once the data is properly cleaned and prepared, the choice of features
matters more than the choice of algorithm. A dataset with a strong, meaningful feature like
Credit History leads to better results, no matter which model is used. The thesis also shows that
SHAP can be used with both simple models (like Logistic Regression) and complex models
(like CatBoost) to explain their decisions in a similar way. This means banks do not have to
choose a harder-to-explain model just to get better accuracy — explainability tools can work
well even with strong-performing models. The thesis ends with a discussion of the trade-off
between accuracy and explainability, and suggests future work such as testing more datasets,
checking for bias, and trying deep learning models.
Collections
- 2026-2030 [51]