There's a better way than saying "Don't you churn on me!"
What is churn?
Churned customers are customers who stopped being your customers. Besides raising questions like “Why don’t they like me?”, churn is also very bad for your business. Moving to more structured definitions, for subscription companies it's pretty straightforward. Churned customers are those who cancel/fail to renew their contract. Customers change banks, mobile providers, gyms and we know the exact moment when the customer churned. Non-contractual companies have to customize the definition a little bit. If a customer does not repeat purchase within a predefined time window, he is considered as churned. This time window has to be adjusted to average purchase frequency (or distribution of frequencies) and we really don’t know when the customer churn. Although it sounds as a simple distinction, it’s really important to classify your business accurately because models used to calculate churn are very different depending on the type of problem.
Non-contractual churn
In the environment where a customer can make a purchase at any time, of any value, the most relevant information about the customer are 3 key numbers, recency, frequency and monetary. Nice thing is that all of those can be extracted from transactional data and they are sufficient for a wide family of RFM models. RFM models can be as simple as low frequency and lower recency = high churn risk. They can also be very complex when we add social-demographics and different clustering methodologies into the mixture. For example, we can start with unsupervised clustering to identify subgroups with different expected behavior and then create a supervised classifier for each segment, calculating the probability of churning on an individual level. As it often is in ML world, Occam's Razor is a useful guideline and the model should be chosen having in mind the available data (size and quality) and available options for actions you can take in order to reduce churn.
Contractual (Subscription) churn
Going back to yearly/monthly/daily contracts, it becomes clear why models addressing this type of problem are very different from the non-contractual. Recency and frequency are predefined, there is nothing we can learn about the customer from it (unless he was given the choice for the period length). We have to look for other sources of what might be relevant information about the customer. Common candidates are: type of product, variety of products, customer interactions (such as customer service complaints) and of course social-demographics, if available. The goal is to find risky audience segments but also to gain an insight into why they are risky. Common approaches are different types of supervised clustering, looking for common features of people who ended their contracts sooner than average. A very basic example would be a decision tree trained for binary churned/non-churned target. Every leaf of the tree becomes 'loyal' or 'risky' segment, where the tree path can give insights on what defines those groups and how they can be approached.
Why should we worry about churn?
Simply said, because it’s extremely important that you do it. Customer acquisition is important but customer retention is more important. It’s so important that the state-of-the-art investment funds take it as one of the most important KPIs when evaluating companies, especially in saturated markets. Also acquiring new customers is usually 5-25 times more expensive than retaining one.
“Average life Expectancy of Fortune 500 Companies is less than 15 years now compared to 75 years in 1955. In the past 15 years 52% of the Fortune 500 Companies have vanished.”
Remedies
While giving discounts and calling every customer that you have will reduce churn, killing your profits is also not the solution. That’s where we get to the beauty of personalized predictions. With personalized predictions, you can call only customers from high-risk segments, offer discounts only to customers who wouldn’t purchase the product otherwise, tailor that discount level to the bare minimum required to become appealing to each customer. Long-term actionables go even further, you can understand who are your churning customers and define new products that will meet their expectations, you can adjust your communication channels and tailor your approach to each risky segment.
Final word
Churn management is essential for any business and tools for high-quality, controlled-risk customer-centric decisions are there. Don’t say you weren’t warned.
Comments