Causal Effects of Communication Channels in Bank Telemarketing Campaigns: Evidence from Interpretable Causal Machine Learning Models

Abstract

This study examines the causal effect of communication channel choice on customer subscription decisions in bank telemarketing campaigns. While existing studies rely heavily on predictive machine learning models to assess campaign success, these approaches are often limited in supporting causal decision-making processes. To address this gap, this study adopts a causal machine learning framework using an open-access bank marketing dataset. The dataset consists of approximately 41,000 observations, with communication channel mobile phone communication channel (cellular) or telephone (landline phone communication channel) treated as a binary strategic intervention variable. Mean Treatment Effect (ATE) is estimated using Dual Machine Learning, while Conditional Mean Treatment Effects (CATE) are analyzed through a Causal Forest model to capture customer-level heterogeneity. The results show that although the mean effect of the communication channel is modest, significant heterogeneity exists among customer segments. To interpret the sources of treatment effect heterogeneity, a SHAP-based explainability analysis is performed on a high-accuracy surrogate model that approximately represents the CATE estimates. The findings reveal that macroeconomic indicators such as employment levels (np.employed) and interest rates, along with campaign-related features, play a critical role in shaping the effectiveness of communication channels. Mobile communication has strong positive effects for certain customer groups, while its impact is weak or even negative for others. Overall, this study contributes to the bank telemarketing literature by moving beyond predictive accuracy and towards a causal and interpretable decision support framework. The results highlight the importance of designing marketing strategies that account for customer-level heterogeneity and contextual factors, rather than relying solely on average effects.