Evaluating AI-driven credit scoring models versus traditional statistical techniques

Abstract This study evaluates AI credit-scoring models against standard statistical models using a real-life data set with 1000 loan applications. The main research question is about whether machine-learning methods are more valuable in terms of predictive accuracy, interpretability and stability to changes in the process of macroeconomic deterioration as compared to traditional methods. The four model variants were developed: the logistic regression and decision tree as examples of the classical ones, and XGBoost gradient-boosting ensemble and multilayer perceptron neural networks as examples of AI-based alternatives. The methodological engine was R (v4.3.1), which established a 70:30 train-test strata union, inner cross-validation and harmony spatial search to tune the hyperparameter. The findings show that XGBoost produces an optimal balanced accuracy cumulative gain curve (cumulative gain, CGC: 0.89; area under the receiver operating characteristic, AUC: 0.89) that is trumped by the neural network (CGC: 0.87; AUC: 0.87) and succeeded by the logistic regression (CGC: 0.76; AUC: 0.76). The SHAP analysis shows that the amount of credit, duration of loan, and age are central predictors. During stress-test simulation, XGBoost is stable in making predictions (AUC = 0.83) compared to a severe decline in logistic-regression results (AUC = 0.68). The results therefore justify the claim that the state-of-the-art AI tools have better predictive potential and robust interpretability, thus rising as viable alternatives to the systems in vogue in modern finance organisations.