Comparing logistic regression and neural networks for predicting skewed credit score: a LIME-based explainability approach

Over the past years, machine learning emerged as a powerful tool for credit scoring, producing high-quality results compared to traditional statistical methods. However, literature shows that statistical methods are still being used because they still perform and can be interpretable compared to neural network models, considered to be black boxes. This study compares the predictive power of logistic regression and multilayer perceptron algorithms on two credit-risk datasets by applying the Local Interpretable Model-Agnostic Explanations (LIME) explainability technique. Our results show that multilayer perceptron outperforms logistic regression in terms of balanced accuracy, Matthews Correlation Coefficient, and F1 score. Based on our findings from LIME, building models on imbalanced datasets results in biased predictions towards the majority class. Model developers in the field of finance could consider explanation methods such as LIME to extend the use of deep learning models to help them make well-informed decisions.