Predictive Modeling of Customer Lifetime Value in E-Commerce Using Deep Learning and Causal Inference

Abstract

Customer Lifetime Value (CLV) is a central metric in e-commerce for acquisition budgeting, personalization, retention, and strategic planning. Classical statistical and probabilistic models (e.g., Pareto/NBD, BG/NBD) provide principled baselines but struggle with high-dimensional covariates, nonstationary behavior, and counterfactual questions required for causal decisioning. Deep learning has recently delivered substantial gains in CLV point and distributional prediction at industrial scale by integrating representation learning, sequence models, and distributional heads; however, purely predictive models risk conflating correlation with causation when used to inform interventions (promotion allocation, pricing, retention offers). In this article we propose a rigorous, production-ready framework that fuses state-of-the-art deep learning architectures for CLV (sequential encoders, attention/Transformers, mixture/distributional output layers) with modern causal inference techniques (potential outcomes, double/debiased machine learning, causal forests, representation learning for counterfactuals) to deliver accurate, robust, and actionable CLV estimates for e-commerce. We provide formal problem statements, architecture blueprints, objective functions, validation protocols (temporal cross-validation, backtest, uplift evaluation), and risk/ethical governance guidance. We demonstrate how probabilistic deep CLV models (e.g., zero-inflated / mixture output, heteroskedastic heads) can be combined with causal estimators (double ML, causal forests, learned balanced representations) to produce both predictive scores and valid estimates of causal effects of marketing actions on CLV   enabling prescriptive decisioning with sound uncertainty quantification. We ground our design choices in the literature and present a reproducible experimental protocol and evaluation suite for industry benchmarking.