<![CDATA[Customer retention is a strategic concern for enrollment-based educational technology (EdTech) firms, where acquiring a new customer costs far more than retaining an existing one. Churn research in education, however, rests on an unresolved measurement premise: models routinely treat any post-onboarding inactivity as churn, conflating genuine dropouts with fast completers who finish and leave, and they typically require full-course data that arrive too late for intervention. Addressing both gaps, this study develops and validates an early-warning model of customer dropout in a vocational, competency-based EdTech platform in Indonesia aligned with the national work competency standards (SKKNI), using only behavioral data from each customer’s first 14 days. The novelty is at once conceptual and methodological: a completer-aware dropout label requiring both sustained inactivity and the absence of an earned certificate combined with a uniform outcome window that removes censoring bias. Analyzing an operational database snapshot (11 February–28 July 2026) of the SertifiTech.id platform, a cohort of 177 paying customers was studied. Nine behavioral features were engineered under an adapted Recency–Frequency–Intensity framework, and logistic regression, random forest, and gradient boosting were compared through 5-fold stratified cross-validation with odds ratios, SHAP, and permutation importance. Dropout reached 23.2%. Early engagement persistence was the single dominant protective factor, robust across every model family and interpretation method (OR = 0.094), whereas activity breadth proved protective rather than harmful reversing the ‘skimming’ effect obtained under a naive label. The best model attained AUC 0.794 (95% CI 0.718–0.864), and targeting the highest-risk 20% of customers captured 46% of all dropouts (2.3× lift), evidencing a practical two-week early-warning system.]]>
Copyrights © 2026
