Neural Network–Based Customer Retention Forecasting in Mobile Wallet Services Using 200k Historical User Profiles

Abstract

This study developed and evaluated a neural network–based retention forecasting model using 200,000 historical mobile wallet user profiles. Retention was operationalized as continued transactional activity within a defined post-observation horizon, with 68% of users classified as retained and 32% classified as churned. Descriptive results indicated substantial behavioral heterogeneity, with mean transaction frequency of 12.8 transactions (SD = 21.3), median of 6.0, and mean recency gap of 18.6 days (SD = 27.4). Retained users demonstrated higher average merchant diversity (M = 6.8) compared with churned users (M = 2.9), and shorter inter-transaction gaps (M = 11.2 days vs. 34.7 days). Logistic regression results showed significant effects for recency (β = -0.042, p < .001), transaction frequency (β = 0.087, p < .001), and engagement trend (β = 0.054, p < .001). The logistic baseline achieved an AUC of 0.78 and PR-AUC of 0.64, while regularized regression improved performance to an AUC of 0.80 and PR-AUC of 0.67. The neural network model achieved superior discrimination with an AUC of 0.86 and PR-AUC of 0.74, and produced a lift of 3.5 in the top 10% risk segment. Segment-level results remained stable across new users (AUC = 0.84) and mature users (AUC = 0.88). Findings demonstrated that combining intensity, diversity, and stability constructs within a neural architecture significantly enhanced retention forecasting accuracy in large-scale mobile wallet environments.