Because of the low cost and user-friendliness, telemarketing has become a common way for banks to obtain deposits for a long time. Meanwhile, researchers have been attempting to analyze consumer information in-depth to improve the effectiveness of bank telemarketing and attract deposits through telephone communication. In this paper, we construct bank telemarketing prediction models by three machine learning (ML) methods, i.e., Random Subspace (RS), Multi-Boosting (MB) and Random Subspace-Multi-Boosting (RS-MB), and find the best performing model. Also, we make the interpretability analysis to provide banks with decision information to develop and implement an effective marketing plan. We rank the importance of the original independent variables by the ML method and select those variables whose influence on the prediction results was significant. And we reconstruct the bank telemarketing prediction models based on the selected independent variables. Furthermore, we illustrate the marginal effects of the selected independent variables on the consumers’ subscription of deposits by the Partial Dependence Plots (PDP) to analyze how these selected independent variables affect the success of bank telemarketing campaigns. The empirical results indicate that the RS-MB using selected independent variables achieves the best performance for prediction. It is worth noting that banks would rather contact uninterested customers than miss potential deposit customers. Therefore, when predicting the success of telemarketing campaigns, it is more critical to reduce false negative rate than false positive rate. Moreover, banks using telemarketing should pay more attention to type of job that the customer does, the month that the customer was connected, and contact day of week.
Read full abstract