The potential for predicting churn in B2B contexts remains untapped despite the growing availability of digital customer traces, particularly usage data. Nevertheless, this source provides valuable insights into customer behavior and product interaction, serving as a powerful tool for understanding customer needs and improving retention efforts. This paper aims to explore the value of usage data for B2B customer churn prediction using a real-world dataset with 3959 subscriptions from a European software provider. The contribution to literature is two-fold. First, we define a framework to structure usage data into cross-sectional features that could be included in predictive models. We study the impact of usage information on churn prediction performance along three primary components– timing, granularity, and expertise. The findings confirm the value of this data source and two of the three examined components, particularly in terms of AUC and TDL. Second, we provide insights into the carbon footprint of the entire modeling process required to integrate usage data into different machine-learning algorithms. Furthermore, to strengthen the robustness of our findings, we compare five common machine-learning algorithms in CCP. Consequently, this study uncovered unexplored aspects of usage data and practical implications for enhancing churn prediction in a B2B setting.
Read full abstract