Identifying Customer Churn Risk in Subscription-Based Digital Services: Manifestations of Customer Engagement as Predictors of Dissolution

Abstract

Background/objective: Customer churn represents a persistent operational challenge for subscription based digital services, where behavioral data is abundant but the relational states driving dissolution remain unobservable. This study develops a theory-informed predictive framework that operationalizes the behavioral dimension of customer engagement (CE) intensity (Brodie et al., 2011) as observable proxies for subscription dissolution risk, responding specifically to Foundational Propositions FP2 and FP5. The framework is evaluated on anonymized customer and session data from Bonnier News, a multi-title digital news subscription provider. Methodology: A predictive research design is employed, triangulating three complementary models: logistic regression, XGBoost with SHAP explainability, and Cox Proportional Hazards. Model performance is assessed primarily through AUC-ROC and concordance index across three analytical frames isolating the predictive contribution of CE-informed behavioral proxies. Results: Behavioral proxies outperform the demographic baseline across all frames, with the largest margin in the engaged-only frame (XGBoost AUC 0.778 vs. 0.708). Engagement persistence signals, specifically frequency, recency, and frequency trend, converge across all three models, supporting H2. Depth (H1) is only partially supported, and breadth (H3) is not supported, with both nulls explained by architectural and content-format constraints consistent with FP5. Structural features, namely trial subscription status (HR = 6.28) and payment method, dominate overall prediction, while 71.4% of observed churners maintained stable engagement until cancellation. Conclusions/implications: Behavioral CE proxies carry genuine but bounded predictive signal: they discriminate within the segment where engagement has formed and is eroding, but are structurally blind to the larger stable-engaged churner segment, which the study interprets as an empirical observation of spurious loyalty. A dual-channel measurement architecture, combining behavioral proxies for vigor with structural commitment signals for absorption and dedication, is proposed as a portable framework for subscription churn prediction where attitudinal measurement is unavailable.