Using Transformers for B2B Contractual Churn Prediction Based on Customer Behavior Data

Abstract

In the competitive business-to-business (B2B) landscape, retaining clients is critical to sustaining growth, yet customer churn presents substantial challenges. This paper presents a novel approach to customer churn prediction using a modified Transformer architecture tailored to multivariate time-series data. We suggest that analyzing customer behavior patterns over time can indicate potential churn. Our findings suggest that while uncertainty remains high, the proposed model performs competitively against existing methods. The Transformer architecture achieves a top decile lift of almost 5 and 0.77 AUC. We assess the model’s confidence by employing conformal prediction, providing valuable insights for targeted anti-churn campaigns. This work highlights the potential of Transformers to address churn dynamics, offering a scalable solution to identify at-risk customers and inform strategic retention efforts in B2B contexts.