No catch-all methodology for Churn modeling
Marketing professionals constantly struggle to leverage advanced analytics to improve upon existing KPIs. One common business issue is reducing customer churn – that is, minimizing the number of customers who stop buying your products or using your services. Predictive analytics for churn modeling is the buzzword that everyone, from marketing to customer service and other non-technical departments is now using. The web offers a lot of academic literature on churn, and the number of proposed analytical approaches is overwhelming. Why is that? Because most of the examples and the techniques presented can only be applied in a very particular context, and there is no one method that works well in every situation.
The real obstacle to effective churn modeling is poor problem definition; that is, marketing professionals and data scientists alike are often guilty of trying to predict who will leave without first properly defining when a customer should be considered “churned”. No technique will make up for a lack of a proper definition of churn, and results will be unreliable or just plain wrong.
Contractual vs. Non-Contractual Churn
The first step in defining who is at risk of leaving is defining who has already left, and for that we need to distinguish between contractual or non-contractual churn.
Contractual churn occurs within subscription-based businesses, such as within the telecommunication or entertainment industries. In contractual churn, the churn date is the date a service or membership is canceled or not renewed. Because the date is clearly defined and universally understood, machine learning techniques that incorporate time effects such as customer tenure or recency of purchase are ideal.
On the other hand, non-contractual churn has no clear churn date as customers simply stop purchasing without having to cancel. The question then becomes: how long should I wait before I consider a customer “churned”? Waiting too long means that it will be more difficult to win them back, while waiting too little could result in wasted resources. Machine learning algorithms should be leveraged to first define the exact window of time that should elapse before a customer is considered “churned”, and then predict who, among the remaining customers is also at risk of leaving. When dealing with this type of churn, customer tenure and recency of purchase usually result in model overfit and poor prediction power.
So, what is the next step?
There isn't a silver bullet technique that will always work, and a correct definition of churn will pave the way to successful machine learning implementation. Correctly identifying "churn" is a hard-learned skill and scaling this knowledge to all the analysts in your organization is a daunting task. Instead, partner with Dunn Solutions Group, a digital transformation consultancy that has already accumulated this hard to acquire knowledge by working with hundreds of clients across several industries.