Objective: Bank telemarketing campaigns often rely on increasing the number of calls to improve conversion rates. However, repeated contact may eventually lead to call fatigue, perceived intrusiveness, reduced customer willingness to engage, and inefficient use of call-center resources. Unlike most previous studies, which have primarily focused on comparing algorithmic performance and accurately predicting customer responses, this study seeks to transform predictive modeling into a practical managerial decision-support tool. Specifically, it aims to identify the nonlinear relationship between the number of calls and the likelihood of offer acceptance, determine the threshold beyond which additional calls produce diminishing or negative returns, and examine whether sensitivity to repeated contact differs across age groups. From this perspective, the study goes beyond improving customer targeting by proposing a framework for reducing low-yield calls, controlling campaign costs, protecting the customer experience, and allocating marketing resources more efficiently.
Method: The study used the UCI Bank Marketing dataset, which contains 45,211 observations and 17 demographic, financial, and campaign-related variables. The target variable indicated whether a customer subscribed to a term deposit. Three classification algorithms—logistic regression, random forest, and a multilayer perceptron—were employed to predict customer responses. Because the dataset was highly imbalanced, model performance was evaluated using accuracy, precision, recall, F1-score, area under the receiver operating characteristic curve (AUC), and confusion matrices. To move beyond predictive accuracy and clarify the underlying logic of the models, SHAP values, partial dependence plots, and a two-dimensional interaction analysis of customer age and number of calls were also applied.
Findings: The random forest model achieved the most balanced performance, with an AUC of 0.802 and an F1-score of 0.46 for the minority class. Descriptive evidence also revealed a marked decline in conversion rates: from 13.19% among customers receiving one or two calls to 10.34% among those receiving three or four calls, and 6.41% among those contacted five times or more. The SHAP and partial dependence analyses consistently showed that the first call generated the strongest positive contribution, the marginal effectiveness of contact declined during the second call, and the negative effect of repeated calling became more pronounced from the third call onward. Customers under the age of 30 were more responsive to initial contact but also more sensitive to repeated calls, whereas older customers exhibited a more gradual decline in response probability.
Conclusion: The convergence of the descriptive and model-based evidence suggests that two calls represent an appropriate operational limit for most customers. After two unsuccessful attempts, continuing to call with the same message is likely to create greater cost and irritation than potential benefit. Banks should therefore move away from mass-calling practices, prioritize customers according to predicted response scores, and reserve a third call for individuals with a high likelihood of acceptance or clear behavioral evidence of interest. Other customers should be redirected to less intrusive channels, such as text messages, email, or mobile-app notifications. Implementing such a strategy can reduce campaign costs, release call-center capacity for higher-value prospects, strengthen customer trust and satisfaction, and support more sustainable long-term relationships between banks and their customers.