All account management, customer success, and support teams have one key goal: to reduce and minimize churn. Churn prediction is the process of identifying segments or specific customers that are at risk of churning, i.e. discontinuing their business, within a short amount of time in order to deal with the customer health issue as much in advance as possible.
A good churn prediction model is important to gauge future expected revenue, and can also help identify customers who would benefit from special promotional messages and deals that could encourage them to delay the discontinuance of their business. Furthermore, accurate churn rate predictions could help identify and improve areas of business where your company is lacking.
The most important data for churn prediction are records of past rates of customers lost in a given period of time compared the amount of customers gained in the same period. Additional data necessary for an in-depth analysis of past churn rates includes records of customers’ experiences and feedback, hopefully indicating patterns leading to churning.
Although churn rates are calculated from internal data, a good churn prediction model will also be based on competitors churn rates, and any data available regarding different demographics and relevant usage patterns with competitors or similar services.
Additional data could include the affect of advertisement through different marketing channels on the churn rate of customers.
The main challenges in this use case are the two following questions:
The complexity of causes and types of churn can pose a difficult challenge in creating an efficient model.
ResearchGate: Customer churn prediction – A case study in retail banking
IEEE Xplore: Customer churn analysis for a software-as-a-service company
XGBoost is the implementation of the gradient boosted tree algorithms that are commonly used for classification and regression problems.
A group of researchers from the University of Virginia studied the time-dependent software feature usage data, such as login numbers and comment numbers, to predict a customer churn within three months. The authors compared model performance across four classification algorithms and concluded that ‘the XGBoost model achieved the best results for identifying the most important software usage features and for classifying customers as either churn type or non-risky type.’
Customer Churn Prediction Using Machine Learning: Main Approaches and Models