Propensity to churn modeling does not help reduce churn

I hear a lot of buzz around advanced methods, like predictive analytics, machine learning, data science, etc. Everyone says that’s what you need if you want to make a difference in your business. For example, executives want to see predictive models that tell them who the most “at risk” customers are so they can be targeted for retention. However, when applied to a real life situation this approach often fails to deliver the results.

This is how propensity to churn modeling is done. We run this model on our existing customers, and thus we know a lot about them, from their name and address, which we can link to demographics to what products they use, to what their payment history has been. While churn forecasting can be challenging at times, generally speaking, with the wealth of information we have on them, we should expect to create a pretty good model. For example, it does not take advanced modeling to determine that subscribers who are late on their bill are at high risk of disconnecting.

But how this model is going to be used once it is complete? The most common use is to find the best target for churn reduction.

How is our predictive model going to help us here? It does not. While it does tell us a lot about natural propensity to churn without any intervention, it tells us nothing about our ability to change customer behavior and stay longer. Turns out, this is not a trivial problem. In fact, many researchers have found that our ability to influence customers may be unrelated or inversely related to their propensity to leave. In other words, just having a customer segment being marked for high churn risk does not mean that convincing them to stay is the best bang for our buck.

The problem with using propensity to churn model is that it optimizes for the wrong thing.

It predicts what the customer is going to do. It does not tell us which customers we can effectively persuade to change their minds. The model that does this is sometimes called an “uplift model”. This type of a model can only be run in conjunction with an in-market test, where we look at different factors that help us find customers that are persuadable. Building an uplift model is usually quite challenging, and my general experience suggests that creating a model is usually an overkill. Many times we can get away with measuring your program(s) against a control group, and then exploring our results by most important segment breakdowns. This is a very efficient, yet simple way to resolve this conundrum.

Positive lift