A Composable Additive-Tree Hybrid with a Stabilised Decision Threshold for Explainable Churn Decision Support in Telecommunications Information Systems

Customer churn prediction is a core predictive analytics function in telecommunications information systems, where model outputs feed retention campaigns and resource allocation decisions. Although supervised learning is widely used to rank subscribers by attrition risk, published comparisons rarely hold the tuning protocol, resampling design, and decision rule constant, so it remains unclear which part of the pipeline actually generates business value. This study addresses that gap on a widely used public benchmark of 7043 subscriber records described by 19 native attributes, with a churn prevalence of 26.54 per cent. Five learner families and a proposed Composable Additive-Tree Convex Hybrid (CATCH) were evaluated under a single nested protocol: repeated stratified five-fold cross-validation with three repeats, an 18-configuration random search scored by inner three-fold cross-validation, and a decision threshold estimated on the training partition alone. CATCH aggregates four learners whose outputs admit an exact additive decomposition on the logit scale, and couples them with a plateau-stabilised, metric-aware threshold estimator that translates risk scores into operational retention decisions. The hybrid obtained the highest mean area under the receiver operating characteristic curve (0.8499), F1 score (0.6380), and Matthews correlation coefficient (0.4926), exceeding the class-prior benchmark by a wide margin on every threshold metric, although its margins over the tuned single learners were small. Signed-rank tests indicated a ranking advantage over four of the five baselines, of which only that over the random forest persisted under a conservative corrected test, while the explainable boosting machine remained statistically indistinguishable; the hybrid is therefore presented as an exactly attributable aggregate. Decomposing the sources of improvement showed that component tuning and the decision layer dominate: replacing the conventional fixed threshold of 0.5 raised the Matthews correlation coefficient by 0.0217 for the hybrid and by as much as 0.0374 for the random forest, whereas aggregation contributed only 0.0005 in ranking quality. Composed Shapley attribution proved more stable than permutation importance and identified contract type, tenure, and internet service as the dominant drivers, giving retention managers interpretable evidence to act on. For organisations deploying churn models in enterprise systems, the results obtained on this benchmark indicate that well-structured tuning and threshold selection deliver more decision value than ensemble complexity.