A Weighted Voting Ensemble Machine Learning Framework for Customer Lifetime Value Prediction in E-commerce

Abstract

Customer lifetime value (CLTV) prediction supports the allocation of marketing resources and the identification of high-value customers in e-commerce. Traditional approaches may not adequately represent non-linear and multidimensional purchasing behaviour, particularly when customer-value classes are imbalanced. This study developed a weighted soft-voting ensemble for CLTV tier classification using the Dunnhumby Complete Journey dataset. Transactional, product, and household-demographic data were integrated, and customer-level recency, frequency, monetary, behavioural, promotional, temporal, and demographic features were engineered from the training period. A chronological split was applied at day 533 so that the training data preceded the evaluation period, thereby reducing the risk of temporal leakage. The Synthetic Minority Over-sampling Technique was applied only to the training data. The ensemble combined Random Forest, XGBoost, LightGBM, and CatBoost after isotonic probability calibration. Dynamic voting weights were derived from F1-score, receiver operating characteristic area under the curve, and accuracy. The proposed ensemble achieved 77.80% accuracy, 77.25% macro F1-score, 77.70% precision, 77.33% recall, and 90.58% ROC-AUC. It exceeded the individual models in accuracy, macro F1-score, precision, and recall, although CatBoost produced a slightly higher ROC-AUC. The findings indicate that calibrated, performance-weighted probability aggregation can provide a modest improvement in balanced CLTV tier classification. Further validation across additional retail datasets and market settings is required before broader generalisation.