Geiger Uses Machine Learning to Reduce Customer Churn in the Promotional Products Industry

This paper addresses the problem of customer churn faced by Geiger GmbH, a German midsized promotional products distributor. In its drop-shipment business, Geiger was losing 20%–25% of annual turnover, and replacing these customers through acquisition consumed significant sales and marketing resources. To address this challenge, Geiger launched a project to predict and prevent churn, aiming to turn reactive retention efforts into proactive customer management. Our methodology followed the cross-industry standard process for data mining framework and systematically developed a machine learning churn prediction model based on transactional, demographic, and interaction data from Geiger’s operational systems. We engineered 77 features that reflect customer recency, frequency, monetary value, and engagement patterns. We benchmarked promising off-the-shelf models, including logistic regression, random forest, XGBoost, LightGBM, and CatBoost, using a time-based, nested cross-validation design. Random forest was selected for deployment because of its robust predictive accuracy and track record in comparable problem settings. The model is integrated into a Microsoft Power BI dashboard, in which it provides sales representatives with automated monthly churn risk scores and actionable watch lists. The results show that the deployed model identifies more than twice as many true churners as random selection when targeting the top 50 customers at risk for churning, maintaining a low false-positive rate. Broader targeting scenarios confirm consistent lifts above 1.8, demonstrating significant operational value. Geiger now proactively manages at-risk customers, reducing unnecessary acquisition costs and increasing salesforce efficiency. The case illustrates how structured machine learning integration can transform retention strategies in noncontractual business-to-business contexts, offering managers a replicable approach to balance churn prediction accuracy with practical deployment considerations.

History: This paper was refereed.