<![CDATA[Among 93,357 customers analyzed from the Olist Brazil e-commerce platform, nearly four in ten were found to be in a state of permanent churn a condition invisible to conventional transaction reporting without data-driven segmentation. This study proposes a two-stage analytical pipeline integrating RFM-based (Recency, Frequency, Monetary) K-Means Clustering with a Random Forest Classifier for churn prediction, structured within the CRISP-DM framework. Data were drawn from the Olist Brazilian E-Commerce Public Dataset covering 115,653 orders between 2016 and 2018. Churn was operationalized as customers with recency exceeding 180 days and a transaction frequency of one, yielding a churn proportion of 56.4% across the sample. Clustering at K=4 (Silhouette Score=0.526) partitioned customers into four behaviorally distinct segments: Active (53%, churn rate 29%), Lost (39%, churn rate 100%), Big Spender (4%, churn rate 60%), and Loyal (3%, churn rate 0%). Cluster labels were subsequently incorporated as input features into the Random Forest model a design decision that proved consequential, as the cluster variable emerged as the single strongest predictor with a feature importance score of 0.826, outweighing all individual behavioral features combined. The model achieved an ROC-AUC of 0.897, accuracy of 82.9%, precision of 97.7%, recall of 71.5%, and F1-Score of 82.6%. These results demonstrate that customer segmentation, when embedded within a predictive pipeline rather than used in isolation, yields substantial gains in churn detection capability.]]>
Copyrights © 2026
