Customer Retention Forecasting in Mobile Wallet Services Using Neural Networks: A Comparative Quantitative Study

Abstract

This study addresses the problem that mobile wallet providers often lack reliable, data-driven methods to forecast customer retention early enough to prevent churn, particularly when retention drivers are nonlinear and interact (for example trust and perceived risk). The purpose was to quantify the key determinants of mobile wallet retention and to compare classical regression versus neural networks for retention forecasting in a quantitative, cross-sectional, case-based design using a cloud-delivered, enterprise mobile wallet service context. Survey data from N = 420 active users were analyzed as the bounded “enterprise case,” with constructs measured on 5-point Likert scales: performance expectancy (PE), effort expectancy (EE), social influence (SI), facilitating conditions (FC), habit (HT), price value or incentives (PV), trust or security perception (TR), perceived risk (PR), satisfaction (SAT), and customer retention or continuance intention (CR). Descriptively, retention was moderately high (CR M = 3.84, SD = 0.71), alongside strong PE (M = 4.02) and TR (M = 3.93), while perceived risk was lower (PR M = 2.61, SD = 0.83); internal consistency was acceptable to strong across scales (α = .78 to .91; CR α = .88). The analysis plan applied descriptive statistics, reliability testing, Pearson correlations, multiple regression for hypothesis testing, and a matched-input neural network for forecasting under an 80/20 split. Headline findings showed that regression explained substantial variance (R² = .62; F(9,410) = 74.52; p < .001), with SAT as the strongest predictor (β = .31, p < .001), followed by PE (β = .21, p < .001) and TR (β = .19, p < .001), while PR negatively predicted retention (β = −.16, p < .001) and SI was not significant (p = .081). The neural model outperformed regression in prediction (RMSE 0.49 vs 0.57; MAE 0.38 vs 0.44; test R² 0.69 vs 0.62). These results imply that retention programs should prioritize satisfaction engineering, visible security assurance, and risk reduction, while using neural forecasts for more accurate risk scoring and targeted interventions in enterprise mobile wallet operations.