Overview: Churn Modeling Techniques



The reader should be familiar with common machine learning and statistical techniques such as regression, decision trees, random forests, and support vector machines.

Introduction to Churn Modeling Techniques

Ideally, when building a churn model, you want to produce a model that can predict the discrete churn event that best describes the system without overfitting. Your choice of modeling technique will be influenced by the nature of your business and by the final use cases of the model. Usually, there are two broad use cases:

  1. Identifying why customers are churning
    • The goal here is to identify the top drivers of churn
    • Outputs of the model are used to inform retention initiatives, often in the long term
    • Feature importance and effects are of interest, so a choice of interpretable model is essential
  2. Identifying who is going to churn and when
    • The goal is to identify churn risk per customer in the next X days/weeks
    • Outputs of the model will potentially inform short-term retention and win-back campaigns or customer segmentation
    • Accuracy of the model is most important so that retention resources are not wasted on false positives

These two use cases are not necessarily mutually exclusive.

There are three broad approaches to churn modeling: classification methods, survival regression methods, and latent probability models. With the latter method, churn is often modeled as a component of a customer lifetime value model. It’s also important to note that churn is usually more applicable in a contractual (subscription) setting. However, all three techniques can be adjusted to accommodate a non-contractual case after the churn target is clearly defined.
Each of these approaches has pros and cons, and you’ll need to consider the use case when selecting a modeling technique.



Frequent Use Case

Example Models




Which customers have a high risk of churning?

Random forests, SVM, logit, neural networks, ensembles, perceptron

Contractual, Non-Contractual

Need accurate individual churn scores, event-rich data, frequent scoring


What factors are driving churn? What are customer attrition rates? Does customer tenure differ for group A vs. B?

Cox, Nelson Aalen


Interpretable models, cohort comparison, time-specific predictions

Probability Models

Non-contractual, LTV modeling

Pareto-NBD, BB/BG

Non-contractual, contractual

Generative models, natural extension to LTV

Classification Methods

Classification methods will help you identify which customers have a high risk of churning. To get started, you’ll need event-rich data, such as...

Want access to the full notebook? Request a demo of the DataScience Cloud.