Predictability & explainability of survival analysis in churn prediction

Abstract This study addresses customer churn prediction in contractual utility services by applying survival analysis models, which provide time-to-event insights beyond traditional machine learning approaches. Churn is defined as subscription cancellation within a given period. Unlike classification models that only indicate whether churn will occur, survival models estimate hazard rates, capturing how churn risk evolves over time. Our objectives are twofold: (1) to compare two survival models—the Cox Proportional Hazards (CPH) model and the Aalen Additive (AA) model—in identifying key drivers of churn, and (2) to demonstrate their interpretability in predicting churn timing for more effective customer intervention strategies. Experiments with data from a gas utility company show that survival models can successfully predict customer churn across products and contract types. By estimating individual-level risk profiles, these models highlight customers most likely to leave, enabling segmentation based on churn likelihood and timing. This provides actionable insights for designing targeted retention efforts. Overall, the study demonstrates the added value of survival analysis in churn prediction: it not only forecasts whether customers are at risk but also when churn is likely, supporting timely, tailored strategies that reduce attrition and strengthen customer retention.