Improved Customer Churn Estimation Using LSTM Networks

Abstract: In the competitive landscape of financial services, the high cost of client acquisition makes retention a top strategic priority. This study addresses the limitations of conventional credit card churn prediction specifically rigid architectural constraints and inefficient categorical data processing by proposing a dynamic framework based on Long Short-Term Memory (LSTM) networks. By restructuring standard tabular datasets into”pseudo-sequences” and utilizing dense embeddings for high- cardinality features, we enable the model to capture nuanced temporal shifts in customer behavior. Our methodology evaluates four distinct LSTM configurations: Vanilla, Stacked, Bidirectional, and a hybrid Bidirectional-Stacked variant. By integrating these architectures into a unified ensemble, we achieved a peak classification accuracy of 92.35%. Beyond raw accuracy, the ensemble demonstrates exceptional recall performance. For banking institutions, this translates to a more reliable early- warning system that identifies at-risk accounts with precision, effectively reducing the revenue loss associated with undetected churn.

Index Terms: Customer attrition modeling, LSTM architectures, Deep learning frameworks, Ensemble modeling strategies, Credit risk analytics, Class imbalance handling