Classifying credit card customer attrition using explainable artificial intelligence

Type of the article: Research ArticleAbstractManaging customer retention across retail credit-card portfolios remains imperative for containing revenue attrition and safeguarding banking profitability. This empirical study benchmarks five supervised classification architectures to identify observed credit-card customer attrition and investigates the internal attribution mechanisms of the top-performing model using explainable artificial intelligence. Utilizing the BankChurners benchmark comprising 10,127 records across 37 encoded inputs under an unresampled 75:25 stratified split alongside training-only repeated five-fold cross-validation, we evaluated model discriminatory efficacy across threshold-dependent and ranking-based metrics. Gradient Boosting demonstrated the strongest empirical performance on the held-out test split (accuracy = 0.960; weighted F1 = 0.959; ROC–AUC = 0.986; average precision = 0.949), detecting 332 of the 407 observed churners (recall = 0.816; precision = 0.925; F1 = 0.867) alongside 75 false negatives and 27 false alarms, whilst Random Forest yielded comparable ranking capacity (ROC–AUC = 0.984; average precision = 0.934) but lower minority-class sensitivity. Across fifteen repeated training folds, Gradient Boosting achieved mean ROC–AUC of 0.989 (SD = 0.002) and an attrition-specific F1-score of 0.886 (SD = 0.009), while SHAP and ELI5 global attributions on held-out records consistently pinpointed total transaction amount, transaction frequency, and revolving balance as the most decisive predictive drivers. These insights indicate that diagnostic customer reviews can be substantively enhanced by monitoring longitudinal utilization patterns, though retrospective associations must not be equated with prospective warning or intervention efficacy.