In a subscription-based business, even a small rate of monthly/quarterly churn will compound quickly over time. Just 1% monthly churn translates to almost 12% yearly churn. Given that it’s far more expensive to acquire a new customer than to retain an existing one, businesses with high churn rates will quickly find themselves in a financial hole as they have to devote more and more resources to new customer acquisition.
Michael Redbord, General Manager of Service Hub at HubSpot, Customer Churn Prediction Using Machine Learning: Main Approaches and Models, KDnuggets, 2019
Common Pitfalls of Churn Prediction
When building any machine learning-based model, but especially for churn, one has to be careful that the model is actually learning the right thing. For instance, one of the common pitfalls for a churn modeling project is to train the model on both past and future events. To avoid this common mistake, think about what the model will know once deployed into production:
- What data will be available?
- What time frame should be considered for a prediction – probability to churn next week, next month?
Another common pitfall is to produce models that predict people that are obviously going to churn. This is not only a problem with the model, but also with the business implications: the output suggests action when it will be ineffective, since the customers have already decided their intent to churn and thus aren’t sensitive to marketing actions. Instead, models should catch the people who are evaluating leaving and trigger an early warning system.
Finally, the last pitfall is to see a churn model as a one-shot study. Multiple reviews and iterations create a successful strategy to retain and increase customer loyalty. This will likely improve the scalability and reproducibility of the project as well.