Modelling churn, part 2.5 – Designing and creating more model predictors

This is the second half of the variable-creation blogpost. If you haven’t read the first one, be sure to check that one out before continuing below:

As before, this is written with logisitic regression in mind, but the general variable concepts are applicable to machine learning churn models as well.

  • Wide and deep usage of a feature (or the lack thereof) might be a signal that someone is going to churn. For example, you can look at how many times a customer has visited every section of your website during a time period (width) and how many times they have visited a particular section or a particular page on the site during that period (depth). The idea is that width and depth both jointly and independently indicate different types of behaviour and behaviour profiles. Or, from a modelling perspective: If a customer is using a feature either widely or in depth (or both), it can be seen as a signal that they won’t churn.
  • Another take on engagement width is to look at combined dummy variables. A normal dummy variable flags if a customer has done something or not, for example if they have made a purchase during the last month, or if they are signed up for your newsletter. However, by themselves many dummy variables have low predictive power, something that can potentially be remedied by combining them. Think for example that we could create a dummy variable that flags ‘1’ if the customer has called in to customer service during the last month or they have chatted with customer service on the website or they have visited any webpages relating to warranty and refunds. It’s even possible to stack these dummies on top of each other to create a linear aggregate variable (that would be a ‘2’ if the customer had visited the warranty webpages and chatted with customer support, or would show a ‘3’ if they had done all three in the example above). Whether you choose to use combined dummies or stacked dummies, the main idea remains the same: We are collecting all behaviour that are representations of the particular mindset under a shared flag, so that we get a more holistic view of the customers behaviour.
  • If you have some form of earning and consumption mechanism, looking at the net values can give you insight as to whether a customer is investing or withdrawing from your brand. For example, if customers can buy credits from you but can also trade in credits to get cash back, they will have a positive net value in a time period if they buy more credits than they trade in, and have a negative net value if they trade in more credits than they buy. A large positive net value is a sign of someone investing (for example a new customer or someone making a yearly large deposit) and a large negative net value is a sign emptying their funds (this could be because they are going to churn, but could also just be because they had an unforeseen expense that they have to cover). And while large negative net values are worth to keep an eye on for obvious reasons, it’s worth watching customers who have small negative net values continuously over time, or who have negative net value because they only withdrew without doing any deposits for the month.

And that’s all for this time. While there are many other types and styles of variables, these six types were the ones we primarily used in the churn project that we did late last year. If you have any favorite type of your own, feel free to share it in the comments below.

Leave a comment

Your email address will not be published. Required fields are marked *