In 2001 when I co-founded Genroe, our focus was almost exclusively helping clients with their customer retention strategy and execution. And a big part of customer retention is customer churn prediction: identifying customers who might churn and intervening to prevent it.

Since then the number of subscription style business models has exploded and the need to predict and act on customer churn has become even more important.

In this post I review why customer churn prediction is important and how to go about it for both B2B and B2C companies.

What is Customer Churn

There are three main types of churn:

Customer Churn

Customer churn is when a customer stops buying all products and services and becomes an ex-customer. Typically this is a negative event for the company as it reduces the businesses revenue.

Importantly this is done on a cohort basis, i.e. new customers added in the period are ignored in the calculation.

Product churn

Product churn is the number of products or services lost in the period. The number of customers may stay the same, go up or down.

This is relevant in organisations where a single customer purchases multiple services, e.g. banks where a customer might buy a credit card, have a loan and savings account. Product churn is when that customer closes out their credit card.

Revenue Churn

Revenue churn is the revenue lost in the period.

Again the number of customers and products might go up or down but revenue churn is how much revenue has been lost in the period.

If you lose 10 small customers one month but one really big customer the next, your Customer Churn might go down but your Revenue Churn has gone up.

Measuring  Customer Churn

Customer churn is typically reported in percent terms over a period of time, e.g. 5% per month or 24% per year.

The percentage indicates the total number of customers lost in the period as a proportion of the starting amount and ignores new customers added during the period.

Why Customer Churn Prediction Is Important

Customer churn prediction allows organisations to proactively approach at-risk clients and try to repair the relationship before the customer actually leaves. In effect it is a pre-warning that lets you act.

It is generally accepted that retaining a customer costs a lot less than acquiring a new one. So being able to predict a customer is going to churn and preventing it, is better than losing them and finding another to replace them.

Find out exactly how profitable retention is by downloading our free Return on Customer Retention Estimator.

Determining when a Customer Has Churned

Unfortunately it’s not alway obvious when a customer has churned. Depending on how the customer buys, continuous/discrete purchases, and the relationship, contractual/non-contractual, you need to approach it in different ways.

For contractual/continuous relationships, the exact point when a customer churns is pretty easy to identify. 

For instance with telecoms, SaaS, utility companies, a customer must actively inform the company they are no longer going to purchase their services. When they cancel their agreement they have churned.

On the other hand for non-contractual relationships, e.g. retail, printing, consulting and graphic design industries, there is often no agreement to reference. 

Customers also make purchases on an irregular basis so it’s very difficult to distinguish between a customer just not having made a purchase in a while and a customer having churned and now using another supplier.

Things get even more complex when customers can only purchase at specific intervals, e.g. annual conference suppliers.

For more information on determine when churn has occurred see these two posts:

• Calculating Retail Sales Forecasts, Customer Life Time Value, and other customer variables
• Forecasting customer value when you don’t have a contract: Discrete transactions

Predicting Customer Churn

While knowing when a customer has churned is important, predicting that a customer is about to churn is much more important.

Customer churn prediction allows the company to identify customers at risk of leaving and take proactive action to prevent them from ever happening. 

Here are the three main approaches to churn prediction:

Statistical & Machine Learning Churn Prediction Models

Very often customer activity changes in subtle, and not so subtle, ways before they churn. 

Helpfully, companies routinely collect and store enormous amounts of deep and wide customer activity data. 

A churn prediction model combines these two ideas to build mathematical, statistical and/or neural network based models that predict the likelihood a specific customer will churn.

Churn Prediction Model Problems

While these models can be very useful and accurate, there several practical issues with their use:

• Black-box: It is difficult for staff to trust them as they are typically “black box” in nature – there is often no way to explain how a particular churn likelihood score has been calculated.
• Skill requirements: Churn prediction models require highly specialised and skilled people to build and maintain the model
• Converting to action: It can be difficult to convert the prediction that someone will leave to an action the company can take to prevent their leaving. You can’t just call a customer and tell them “our software says you’re going to cancel your agreement soon, what can we do to make you want to stay”?
• Difficult for low transaction businesses: It is difficult for low/no transaction volume businesses to implement, e.g. retail gas, retail electricity, because there are very few items of data on which to make a prediction.

Churn Prediction Model Examples

Here are some example churn prediction model approaches (warning, serious statistics ahead!)

• Median absolute deviation or MAD, is a relatively simple statistical technique, akin to standard deviation, can be used to create a predictive test that can be applied to a large variety of SaaS customer attrition prediction tasks.
  See this post for an example: Crazy Simple Anomaly Detection for Customer Success
• Logistic Regression models try to guess the probability of a customer belonging to one group or another, in this case likely to churn, not likely to churn.
• Neural network churn predictors are systems where software “learns” to perform some task, predicting if a customer will defect, by analysing training examples, in this case, information about customers that left, and did not leave.

Improving Interpretability of Churn Prediction Models

Interpretability is an important aspect of churn prediction models, as it helps build trust in the model’s predictions and enables businesses to take appropriate actions to prevent churn.

While some churn prediction models, such as logistic regression, are inherently interpretable, others, such as neural networks, can be more difficult to understand. Fortunately, there are several techniques that can be used to improve the interpretability of these models:

• Feature Importance Analysis: This technique involves analysing the contribution of individual features to the model’s prediction. By ranking the features by their importance, businesses can understand which factors are driving customer churn and take appropriate actions to address them.
• Partial Dependence Plots: Partial dependence plots show the relationship between a feature and the model’s prediction while holding other features constant. By analysing these plots, businesses can understand how changes in individual features affect the model’s prediction.
• LIME (Local Interpretable Model-Agnostic Explanations): LIME is a technique for generating locally interpretable explanations for individual predictions. By identifying the features that are most important for a particular prediction, LIME enables businesses to understand why the model made a specific prediction and take appropriate actions.

By using these techniques to improve the interpretability of churn prediction models, businesses can build trust in the model’s predictions and take appropriate actions to prevent churn. It is important to note, however, that interpretability can sometimes come at the cost of predictive accuracy, so businesses should carefully balance these trade-offs when building churn prediction models.

Deduced Customer Churn Predictors

Often, businesses are able to identify customer pre-churn indicators from business experience and a good understanding of the customer journey map.

Here are some examples of indicators that the customer is at least considering churning:

• Mortgage lending – customers calling and asking for the “payout figure” on their loan
• B2B SaaS vendors – customers asking for a copy of the agreement and/or when they need to renew
• On-line subscription services of all types – clients who have not logged in for some time or whose use has decreased suddenly 

These approaches are relatively “low-tech” and can be implemented by businesses of all sizes and complexities.

Predicting Customer Churn with Net Promoter Score 

For B2B companies, surveying your customers on a regular basis, relationship surveys, has been show to be an effective way to predict customer churn.

In a recent study it was shown that using “would recommend” style survey questions such as Net Promoter Score, is a very effective way of predicting B2B customer churn 3-6 months before the churn event occurrs

By asking customers the NPS question on a regular basis, typically 3 monthly, and looking for declining scores it is possible to identify at risk customers.

The 3-6 month lead time is particularly important in B2B organisations because trying to save a customer at the last minute is almost impossible. Consistency of supply is key in a B2B context and before cancelling with one supplier, most B2B organisations will already have selected and onboarded a new supplier.

Customer Churn Prediction FAQ