Churn (cancellations of customers) has become a massive problem for numerous companies in times of online comparisons, constant advertising and especially personalized ad tracking. All activities with which companies can reduce and prevent churn are summarized under the term churn prevention measures. These measures are particularly important for companies because they come before customer recovery in the customer lifecycle and are therefore particularly worthwhile from a financial perspective.
There are several typical phases in the customer lifecycle. Using the example of an insurance company, these are: interest (test phase), onboarding (e.g. online service is introduced and set up), establishment (contract runs), renewal (contract is extended), up-selling or cross-selling (additional insurance is sold), termination, customer recovery (customers are convinced to come back by means of a discount).
The reasons for churn are as varied as they are industry-specific:
- Faulty or inadequate services or products;
- Customers receive better offers from competitors;
- Companies restructure their offerings;
- The needs of customers change.
Churn prevention is particularly necessary in B2C industries with long-term customer relationships. Accordingly, most churn analyses and projects take place there — indeed, retaining customers is relevant wherever one-off services are not the main offer. The goal is therefore to take appropriate measures to prevent cancellations and churn. We will present these churn measures later in the article.
What added value does churn prevention bring?
Keeping a customer not only speaks for the performance of a company, but it is also usually cheaper than acquisition. Especially when sales and marketing are costly and/or customer acquisition takes a long time. This applies particularly if customers are profitable only late in the customer lifecycle (e.g. because of discounts, onboarding bonuses, connection of technical products, or complex internal administration).
In addition, existing customers are sometimes significantly more profitable than new customers; in order to prioritize churn measures correctly, the value of customers for the company should be assessed separately in different phases of the customer lifecycle.
We illustrate this with an example.
Calculation example for profitability of churn prediction and churn prevention
Let’s take as a basis a company with 100,000 customers that loses 2% of its customers per month. Sales & Marketing can acquire 1,000 new customers per month with their current budget. The total customer acquisition cost is 1,000€ per new customer.
The key question is: How much does the company need to invest to compensate for the loss of customers?
To answer this question, let’s first look at the development of the customer base, taking into account current new customer acquisition (with and without churn measures):
Customer base development with and without anti-churn measures.
Now, if the company wants to offset the losses due to churn in the previous month by acquiring more, how much would be invested?
Necessary financial investments to make up for churning customers
Looking at the two tables, it is clear with this example that anti-churn measures contribute to a significant reduction in the amount of investment. This is particularly worthwhile for companies that offer long-term, cost-intensive products and services.
Specifically, anti-churn measures result in the following benefits for your company:
- Cost savings
- Decrease of the administrative burden
- Targeted use of company budgets (a better budget allocation is possible!)
- Higher customer satisfaction and increased customer loyalty
- Measures taken become measurable and comparable
- A well-executed churn project offers a 360° customer view and thus, among other things, an in-depth understanding of churn factors
The results of the churn project are transferable to other areas, such as marketing or customer service (e.g. which products should service employees recommend to customers).
However, in order to be able to take measures against churn, a certain investment is of course necessary:
- Anti-churn measures themselves incur costs (ignored in the example above), e.g. financial incentives (discounts) for existing customers to make them stay or an improvement of the product/service (development costs).
- For the selection of the measures and their applications, the customers at risk of switching must be identified and selected — after all, only they should receive discounts, for example. This in turn requires a solution such as a machine learning model, which identifies customers most at risk of leaving and creates a scoring.
The optimal approach to churn prediction projects
1) Checking the data basis and creating a data model
For churn predictions, it is not only the quality and quantity of the data that is crucial, but also the type of data. All customer history data and especially interaction data are important. Static data, e.g. dates of birth or place of residence, are also relevant, but experience shows that they are not the decisive factor for churn itself.
It is also important to check that there is enough data before starting the project. As a rule of thumb, predictions by machine learning models require at least several thousands of customer data records; for churn analysis, it is more likely that even a smaller data volume will be sufficient.
The data records, which are usually stored in fragmented form, must be harmonized into a holistic data model as the first step of the data science process. This requires bringing together all master data and customer history.
2) Identification of churn factors
The development of a harmonized data model paves the way for more standard data analysis. A good starting point is an exchange between data experts and business experts to develop initial hypotheses on churn reasons. For example, such a hypothesis in the insurance industry might be: “Customers will cancel if the premium is increased by more than 3%.”
Data scientists can rely on various tools or analytics libraries for a wide variety of analyses. Analytics of interest could include cohort analysis, cluster analysis, and visualizations of metrics/statistics. Insights are typically discussed and evaluated with business teams.
Based on these insights gained, actions can already be planned or even already taken. For example, if it has been found that terminations are often caused by e.g. problems with a user interface, the problem can be addressed directly. In many cases, however, the main cause of churn is not serious defects, but very individual events. For such companies, using machine learning models to identify churn-prone customers is often a good solution.
3) Creating a prediction model
In most cases, the question of manual vs. algorithmic identification of customers at risk of switching does not arise at all. The effort would simply be too high. The use of an algorithm offers additional advantages:
- Efficient, automated identification of customers at risk of switching (as already mentioned)
- Relief of data analysts and administrative employees
- Learning models get better with each cycle run
- Complex, non-obvious relationships can be detected by powerful, state-of-the-art models
- Machine learning models can identify and incorporate new trends
Such prediction models can be based on different algorithms, depending on the company’s requirements, the data, and the question at hand. Identifying a suitable approach is one of the main tasks of data science teams. Both regression and classification models can be suitable for churn prediction. Typically, linear regression models often used in survival analysis or non-linear models such as ensemble (most often tree-based) algorithms are employed.
An almost ubiquitous challenge in churn prediction, regardless of the model chosen, is achieving a useful precision. The model looks at customers who have churned and compares them to those who have not. For the majority of the time, most customers stay and relatively few cancel. The resulting challenge is a so called “class imbalance”. To reuse the example with 2% churn given above: If a model were to predict that every customer will not quit, it would always be 98% correct. However, this does not make the model useful. The model has to learn from 2% of the customers what made them churn. This “class imbalance” can lead to a lot of “noise” and poor predictions — especially when the data base is thin and customer information is incompletely captured.
To prevent poor predictions and ineffective measures, the model should be well tested. For example, backtesting involves looking at the customers for which the model would have predicted a churn event three months ago and how many of them actually churned.
Our tip: When working with external service providers, ask questions — don’t be fooled by buzzwords and vague metrics! ;)
Once a well-performing model has been created, the next step is to deploy it — in other words, to make the application available to users. The data scientists and IT will take care of the deployment. A number of factors have to be taken into account to find the right approach. If the customer lifetime is relatively short (i.e. there is a high turnover in the customer base), the model could run more frequently; if the lifetime is longer, it may make more sense to run batch predictions ahead of planned campaigns.
4) Take action against churn
The actions that companies can take are too individual to be explained in detail here. Basically, a distinction can be made between immediate actions and long-term measures.
Examples of immediate actions:
- Compensations for errors/problems/defects that have occurred.
- Seeking dialog with customers
- Incentives in the form of vouchers, discounts, product recommendations (e.g., in the case of energy suppliers, the suggestion of environmentally friendly tariffs to promising target groups)
Examples of longer-term measures:
- Designing hard and soft barriers (e.g., minimum contract terms or lock-in effects).
- Improving the service or the offer (e.g. automatic tariff recommendations)
The benefits of churn prevention measures can take different forms. Some, such as the sample calculation presented at the beginning of the text, are clearly quantifiable, while others are less so. A package of measures can, of course, have multiple benefits.
Examples of easily quantifiable benefits of churn prevention are:
- Reduced costs
- Increased revenue with constant customer growth and reduced churn
Benefits that are more difficult to quantify could include:
- Increased employee capacity (and better working conditions because of reduced stress),
- Better customer satisfaction,
- All leading to better brand image and reputation
5) Evaluation and iterative improvement of the churn prevention process.
A critical step for any project is the “next” — so what happens once data has been analyzed, churn predicted, and actions have been taken?
This is followed by the critical evaluation of the activities carried out. It is then imperative that the results and experience are fed back to the data team as well, so that they can continue to improve the models, analyses and data. For example, it is conceivable that the first model works well for customer group A, but not for group B. In such cases, separating into two machine learning models could be useful. Of course, some measures may simply not work.
All these experiences and iterative improvements can become part of an anti-churn strategy. This can include various aspects, commonly encountered are:
- Vision and performance targets → To what percentage should the churn rate be reduced? Which churn causes are to be minimized?
- Areas of responsibility → Who drives the projects and who is the responsible contact person in the departments?
- Approach to experimentation and piloting → What has been learned from previous trials? How should data projects be prepared and how is the collaboration with customer service designed?
- Evaluation of measures or solutions → Which requirements must be met? When will measures be replaced by others?
- Targeted data infrastructure → How do we want our employees to work with churn predictions? How often do we need predictions?
Summary — How do we prevent customer churn?
In order to secure both increased revenues (retention of existing customers) and reduced costs (new acquisition) for your company, churn analysis and churn predictions are very effective tools. The big advantage here is also that additional insights are generated during a project and reusable data assets (such as harmonized customer journey datasets) are obtained.
The path to churn prevention measures is a data science project with several steps. Typically, these core steps are followed:
- Data preparation;
- Churn factor analysis;
- Prediction generation;
- Taking measures;
- Iterative improvement towards a long-term strategy.
Of course, each project phase has its own requirements and specifics. Our key take-away for you would be that the close collaboration between business and data teams is the key to success.
If, in addition, important criteria such as a suitable data basis are met, great added value can be achieved for your company on various levels (finance, reputation, growth).
If you have questions, feedback or would like to discuss how you could approach churn prevention, please do not hesitate to contact me: firstname.lastname@example.org or my colleague Mike from Contiamo, email@example.com.