An explainable survival aware intelligence framework for customer retention decision support beyond static churn prediction

Abstract

Traditional customer churn prediction treats retention as a static binary classification task. This limits operational value because it fails to address when a customer will leave, why they are leaving, and which intervention is economically viable. No existing framework integrates temporal, explanatory, and prescriptive capabilities under a single audited pipeline. We introduce ESACRIF, an Explainable Survival-Aware Customer Retention Intelligence Framework. It integrates seven predictive model families, Kaplan–Meier and Cox survival analysis, SHAP stability auditing, DiCE counterfactual intervention generation, and a data-learned adaptive expected-profit threshold optimiser. ESACRIF is evaluated on three public telecommunications datasets (IBM Telco, Iranian, Cell2Cell) using robust statistical, fairness, and ablation audits. Simple logistic regression (AUC 0.8397) is statistically non-inferior to complex black-box models. Survival analysis reveals a 2× retention-duration gap between month-to-month and 2-year contracts, and SHAP rankings are highly stable across folds (stability score 0.86–1.00). Under a simulation-based principled expected-profit model, the adaptive threshold optimiser achieves + 48.0% ROI, significantly outperforming random and high-risk-only targeting (paired bootstrap p < 0.001). Fairness audits transparently surface demographic disparities for SeniorCitizen and Partner groups. Rather than offering individual algorithmic novelty, ESACRIF reframes churn analytics into an integrated, auditable decision-support architecture. We show that interpretable, properly audited models can be competitively deployed to support temporal, explanatory, and prescriptive decisions simultaneously.