Diving into churn modelling

Many organizations in the profit as well as the nonprofit sector have developed strategies to prevent and minimize the churn (loss) of their customers or donors. These strategies have to go along with the development and application of quantitative metrics that allow measuring churn in order to keep track of how many customers or supporters have gone away over time. How the “gone away” is defined is specific to the respective organizations and the products / services they offer. Churn analyses have a quite long tradition in the telecom sector where different kinds of contracts are offered on the one hand and barriers for customers to leave company A for company B are relatively low on the other hand.

The practise-oriented R case study that is mentioned towards the end of this post was also taken from a telecom example.

At first glance, coming up with a churn rate is fairly easy and simple maths only. Measure the numbers of distinct customers / donors (or contracts or committed gifts) lost in a defined period (say, a month) and divide it by the number of customers / donors / contracts etc. at the beginning of this period. This is straightforward, simple in its application and can be easily
explained – and is therefore a logical first step to tackle churn from an analytical standpoint. An interesting blog post from ecommerce platform provider Shopify offers a discussion on different aspects of churn calculation – and why things are not always as straightforward as they might seem at first glance. 

With regard to an inferential or predictive approach to churn, the blog KD Nuggets summarizes possible approaches and names two major streams:

• Machine learning methods (particularly classifications)
• Survival analyses (e.g. using survival and hazard functions that measure drop-out from a population)

KD Nuggets recommend not to focus too much on one modelling approach but to compare the (particularly predictive) performance of different models. Last, not least they recommend not to forget sound and in-depth explorative analyses as a first step.

Will Kurt wrote a highly recommendable, relatively easy to read (caveat: Still take your time to digest it ...) blog post on how to approach churn using a stochastic model of customer population. Have a particular look at the Takeaway and Conclusion part. A summary attempt in one sentence: Although churn is random by its very nature (Brownian motion), suitable modelling will help you (based upon existing data) tell where your customer population will end up in the best and wort case (see picture below - 4.000 customer just will not be reached if we keep on going as we do now ...)

​Last, not least, if you are interested in a case study from the telecom sector where a Logistic Regression Model was implemented with R, go and have a look at this blog post. The respective R code is also available at Github for the ones interested.

So – enjoy spring if you can and please don´t churn from this blog 😊.