Enterprise Subscription Churn Prediction

In this paper we describe the process of building a churn prediction platform for large-scale subscription based businesses and products. These businesses are constantly innovating on customer retention methods, one of which is to predict which customer will churn in near future. We describe the decision tree model applied, going into, feature engineering, data sampling, model cost sensitization techniques and treatment of false alarms, which is very common in class imbalanced problems such as this. We describe the novel technique of using data segmentation and past prediction of the customer to further increase the precision and recall of the model. Running such a model at large-scale possesses several challenges which we cover in our description of extract, transform, and load and architecture diagram of the platform. We include an empirical proof of our model based on a novel ranking algorithm that proves that the false alarms generated by the model are logical churners. We developed novel tools of model tuning to generate three types of list of "potential churn customers" categorized into high risk, medium risk and low risk. Such a classification enables the business units to tailor customized retention strategies, since each strategy has an associated marketing cost. Further, we are able to use the decision tree to form business rules which help in better targeting the retention schemes. The Churn prediction is a continuous process and it becomes imperative to track customers. We describe the novel implementation of an index/score which we use to track and monitor customer receptiveness to retention schemes and performance over a period of time. Our platform is deployed on several eBay sites and has resulted in the increase of key business metrics. We end the paper by providing new avenues for handling the churn prediction problem better.