PGX-ChurnNet a leakage-controlled cross-schema network for telecom churn early warning

Telecom churn prediction is useful for retention planning only when the observation window precedes the outcome and the evaluation reflects the intended deployment setting. We propose the Planning-Gap-aware Cross-Schema Churn Network (PGX-ChurnNet), integrating semantic-group tokenization, set interactions, training-only memory retrieval, reliability-controlled fusion, and a shared-backbone multi-member output layer. The development data comprised 3150 Iranian mobile customers, with predictors documented for months 1–9 and churn recorded at month 12. Five fixed stratified outer folds were evaluated with five model seeds, yielding 25 descriptive fold–seed evaluations. PGX-ChurnNet achieved F1 0.838 ± 0.010, PR-AUC 0.939 ± 0.007, ROC-AUC 0.986 ± 0.004, and Brier 0.057 ± 0.005. Its seed-averaged F1 difference from TabICLv2 was 0.013, but statistical superiority was not established (raw \(p=0.063\) ; Holm-adjusted \(p=0.625\) ). Frozen evaluation on 8454 Bulgarian business accounts yielded F1 0.306 with source-domain preprocessing and 0.333 with label-blind quantile alignment. Because the Bulgarian snapshot lacks an equivalent planning interval, these values assess cross-domain risk transport rather than replication of three-month early warning. Mapping, feature-eligibility, learning-curve, and scaling analyses defined additional sensitivity and implementation boundaries. Prospective operator-specific validation remains necessary before deployment.