From Dissatisfaction to Defection: Interpretable Churn Prediction in a Slovak Life Insurance Portfolio

Life insurance lapse rates across the euro area have risen to the point of constituting a distinct supervisory risk category, making the early identification of policyholders at risk of termination a priority for insurers. Under Solvency II and the GDPR, however, automated decisions affecting policyholders must remain auditable, which constrains the models insurers can deploy, and the churn prediction literature offers little insurance-specific guidance on what interpretable models achieve or which predictors they surface. This study evaluates six white-box classifiers, namely CART, C5.0, CHAID, QUEST, logistic regression, and linear discriminant analysis, on the complete advisor portfolios of a Slovak life insurer comprising 229 contracts, comparing them with respect to predictive performance, variable selection, and internal model structure, with the stability of the results examined under ten-fold cross-validation. On a testing partition of 74 contracts, three models achieved an overall accuracy of 94.59%, a sensitivity of 87.10%, and a specificity of 100%, with AUC values between 0.94 and 0.97. Logistic regression matched this accuracy through a higher sensitivity of 90.32% but produced numerically unstable coefficients owing to complete separation in several predictor categories, demonstrating that equivalence in predictive performance does not imply equivalence in interpretability. Under cross-validation, all six models exceeded 88% accuracy, although the differences between them fell within the variation attributable to the partitioning of the data. C5.0 offered the most favourable combination of parsimony and rule transparency at comparable predictive performance. Product type was the most consistent churn determinant across models, while surrender value exhibited a non-monotonic relationship with termination and overpayment emerged as a conditional behavioural marker.