Abstract

This paper presents a multi-layered meta-learning framework designed to address the limitations of linear subtraction approaches in customer lifetime value (CLV) modeling. By integrating risk-adjusted components into a hierarchical learning architecture, the proposed method achieves superior predictive performance across diverse customer segments.

Methodology

The framework employs a stacked ensemble of gradient-boosted trees, neural networks, and Bayesian estimators, coordinated through a meta-learner that dynamically weights each base model’s predictions according to estimated risk exposure. Key innovations include:

  • Risk-adjusted loss functions that penalize underestimation of high-value customers
  • Meta-learning coordination layer trained on out-of-fold predictions
  • Temporal feature engineering to capture customer behavior trends over rolling windows

Results

Experiments on large-scale transactional datasets demonstrate a 23% improvement in RMSE over standard CLV baselines, with particularly strong performance in identifying high-risk churn scenarios.

Citation

Journal: SCIENTIFIC CULTURE
DOI: 10.5281/zenodo.18681149
Year: 2026