A Temporal Feature Engineering Framework for Dual-Target Prediction of Customer Lifetime Value and Churn in E-commerce

Abstract

E-commerce platforms in emerging markets face increasing challenges in predicting customer lifetime value and identifying churn risks for effective resource allocation. Traditional customer analytics rely heavily on static transactional features, which fail to capture the dynamic evolution of customer behavior over time. This study develops a temporal feature engineering framework that extracts 12 time-dependent indicators from customer behavioral data, combined with ensemble learning models for dual-target prediction. Using a dataset of 50,000 customers from the Tokopedia platform in Indonesia, we constructed temporal features across three dimensions: RFM dynamics, activity patterns, and engagement indicators. The research conducted correlation analysis and distribution analysis, followed by ablation experiments to validate feature effectiveness. Random Forest achieved R² of 0.9505 for lifetime value prediction, while Gradient Boosting achieved AUC-ROC of 0.9187 for churn classification. Ablation experiments confirmed that temporal features provide meaningful incremental value, with 0.26 percentage point improvement in R² for lifetime value prediction and 1.66 percentage point improvement in AUC-ROC for churn prediction compared to baseline models. Feature importance analysis revealed that while transactional volume metrics dominate lifetime value prediction, behavioral change indicators such as Activity Deviation and Engagement Deviation are more predictive of churn risk. This framework provides practical support for e-commerce platforms to optimize retention strategies and marketing budget allocation.