BEYOND LINEAR SUBTRACTION: A MULTI-LAYERED META-LEARNING ARCHITECTURE FOR RISK-ADJUSTED CUSTOMER LIFETIME VALUE

ABSTRACT The integration of credit risk into Customer Lifetime Value (CLV) modeling remains a critical challenge in financial analytics, predominantly hindered by the “linear subtraction” paradigm which assumes revenue generation and default probability are independent, orthogonal dimensions. This study dismantles that assumption by identifying the “Risky Whale” phenomenon—where high transaction velocity serves as a leading indicator of latent financial distress rather than genuine engagement. We propose a novel Multi- Layered Stacked Generalization Architecture to estimate Risk-Adjusted CLV (RA-CLV). Decomposing the prediction problem into three stochastic layers, we first isolate risk components (PD and LGD), then estimate revenue using a risk-stacked feature space, and finally employ a meta-learner to empirically discover the non- linear exchange rate between value and risk. Using a longitudinal dataset of 5,000 retail banking customers, our results yield three theoretical contributions. First, we demonstrate that while default probability is highly deterministic (ROC-AUC 0.9529), loss severity (LGD) is inherently stochastic, necessitating robust mean- estimation over granular regression. Second, the meta-learner refutes the symmetric accounting logic of traditional models, revealing that predicted revenue acts as a latent value multiplier (β≈1.77) while expected loss acts as a unitary capital deduction (β≈-0.96). Finally, our risk-value segmentation exposes that traditional models systematically misallocate retention budgets to high-risk/high-revenue customers. This architecture provides a rigorous framework for aligning marketing optimization with Basel-compliant risk management. KEYWORDS: Risk-Adjusted CLV (RA-CLV), Stacked Generalization, Meta-Learning, Credit Risk Modeling, Loss Given Default (LGD), Retail Banking Analytics.