Research Background: Enhancing customer churn prediction systems in the telecommunications industry is of strategic importance. Customer retention programs achieve their highest level of success when subscribers at risk of churn can be identified with sufficient accuracy before churn actually occurs.
Research Objective: The main objective of this study is to improve the accuracy, efficiency, and practical effectiveness of prediction models for prepaid customer churn, in such a way that the model output demonstrates strong predictive power.
Research Design and Methodology: This study was conducted using field data from 100,000 prepaid subscribers over a one-year period. To preserve practical validity, the data were split in a forward-looking manner. After data cleaning, feature engineering, and controlling for information leakage, a feedforward neural network was designed as the baseline model. Subsequently, key parameters were optimized using a multi-objective genetic algorithm.
Research Findings: The experimental results indicate that the hybrid neural network–genetic algorithm model achieved an improvement of approximately 0.2% on the test set at a practical scale. In addition, the model led to a more balanced distribution of errors between churn and non-churn classes and noticeably improved balanced accuracy compared with the initial model.
Conclusion: The findings show that even relative improvements in indicators such as recall, when achieved through targeted methods aligned with economic logic, can have a significant impact on the efficiency of retention campaigns and company profitability.
Research Innovation: The main innovation of this study lies in presenting a framework in which, unlike other frameworks, a feedforward neural network is employed to learn nonlinear patterns in subscribers’ behavior, while a multi-objective genetic algorithm plays a role not only in tuning the model parameters but also in optimizing the decision threshold, selecting effective features, and improving the network structure. By increasing the identification of actual churn cases and reducing false alarms, this approach enables more precise customer targeting and optimal allocation of marketing resources. Moreover, the model’s performance is evaluated not only through statistical metrics but also through practical indicators such as Profit@k and Lift, thereby helping to reduce the gap between statistical prediction and economic decision-making.