A Comparative Analysis of Machine Learning and Fuzzy Logic Models for Credit Risk Prediction

Abstract

Credit risk prediction is a crucial task for financial institutions seeking to minimize defaults and maintain financial stability. This study investigates the most relevant predictive variables and compares the performance of three classification model’s fuzzy logic, logistic regression, and random forest, using features selected through variable importance analysis with the random forest method. Based on a dataset of 1,000 Tunisian firms (500 creditworthy and 500 non-creditworthy) described by fourteen financial and non-financial variables, the selected predictors were used as inputs for all three models. Performance was evaluated through accuracy, Type I and II errors, AUC (Area Under the Curve), Cohen’s Kappa, and Fuzzy Average Absolute Error (FAE). Results indicate that fuzzy logic achieved the highest accuracy (98.67%), zero Type I error, a low Type II error rate (2.7%), and excellent AUC (0.99) and Kappa (0.97) scores. However, it recorded a higher FAE (1.2645), suggesting a trade-off between classification accuracy and absolute prediction error. Logistic regression and random forest yielded lower FAEs (0.1997 and 0.2266, respectively) but slightly lower accuracies (87% and 90%). Overall, combining variable selection with multiple classifiers enhances both interpretability and performance assessment, and while fuzzy logic delivers superior classification results, metrics like FAE offer valuable insights into model behavior.