Clustering and Anomaly Detection for Churn Prediction in POS Payment Systems

This work addresses the early identification of businesses at risk of ceasing or limiting their use of point-of-sale (POS) services, resulting in decreased transaction volumes - a key indicator of customer churn. In the highly competitive POS market, customer retention is essential, as acquiring new clients is typically more costly than retaining existing ones. Early detection of potential churners enables targeted retention strategies, enhancing user experience, increasing loyalty, and mitigating revenue loss. To tackle this problem, businesses were first s egmented u sing u nsupervised l earning t echniques, specifically Gaussian Mixture Models (GMM) and K-Means, based on transaction volume and average amounts. GMM outperformed K-Means and was selected for clustering. Following segmentation, transactional data were analyzed across card types and network levels. Due to the dataset's class imbalance - with significantly fewer churners - unsupervised anomaly detection methods were used: One-Class support-vector machine, Isolation Forest, and Local Outlier Factor (LOF). LOF showed the best results, achieving PR-AUC scores of 0.99 and 0.96 for the Micro and Pequeño segments, with recall values of 0.99 and 0.95. For Masivo, Grande, and Mediano segments, lower PR-AUCs (0.71, 0.52, and 0.53) were obtained due to higher false positive rates. Still, their absolute number was small, and they can serve as early warning indicators for proactive retention efforts.