For companies that are active in the B2C sector and have a long-term relationship with their customers, the churn rate is a decisive factor for business success. If more customers leave than new ones join, this will sooner or later mean the end of the business. Likewise, it is critical if the acquisition of new customers is very cost-intensive and customers are only financially profitable after a longer joining period (e.g. insurance membership, subscription, etc.).
If a company determines that action is needed in the area of churn prevention, a thorough analysis of the problem is essential before taking any action. There are many companies for which the use of machine learning models to identify customers at risk of churn (churn prediction) and determine optimal countermeasures has proven successful. In such a project, data- and business experts jointly develop machine learning models that predict the risk of churn based on customer data.
There are a few best practices that have been key to the success of our projects. We would like to share these best practices with you, so that you are ideally prepared for your project.
1) The right data for churn analytics and predictions
For successful churn predictions, it is crucial to check whether you are collecting sufficient and also appropriate data along the customer journey in order to understand and comprehend it.
Be sure to check the foundation as early as possible. If corrections or changes to the data model are necessary, it is better to identify them at the beginning.
Relevant data here includes:
- The purchase/order/contract completion process (frequency, point of sale, user flow, times, contract or subscription changes, etc.)
- Order cancellations (recording in general, the time in the user flow, ...)
- Recording of products in shopping carts or on lists, but which were never purchased
- Service and support contacts (time, reason, form, follow-up), especially complaints (number, reason, frequency)
- Returns and cancellations (number, reason, frequency, time)
- Order values and shopping cart volumes, or selected rates and payments
- Website interactions (clicks, dwell time, feedback)
Static data, usually customer master data, is also relevant, but is very rarely sufficient to accurately predict cancellation behavior. If little or no such interaction data is collected, we recommend taking a step back. Surveys of churning customers can provide insights into the reasons for leaving in a short time. In the long run, it pays to work with your data and IT teams to set up the appropriate data structures. This includes defining relevant data, planning data collection and storage, technical implementation, and iterative improvements.
2) Assemble the right project team
In addition to the data itself, the second key factor is to have the right mix of people involved. This is true for the vast majority of data science projects. A successful data science project requires data scientists (or data analysts), business experts, and members of the IT team. While the first are responsible for the data and data science technologies, the second are essential to ensure that the business world is mapped correctly. Vital input from the business team includes, for example:
- Developing hypotheses for churn reasons;
- Understanding and mapping the customer journey in the data (What is the flow? Are there processes that become relevant only under special circumstances, for example? What is the procedure for e.g. complaints?)
- Contributing restrictions/rules/restrictions: Should certain customer groups be excluded from the analysis? Are there different cancellation rules for different customer groups?
- Defining the required output format: Should highly personalized measures be taken for selected customers? Should more large-scale actions be taken?
On all three sides – i.e. business team, data team, IT team – project planning must ensure that people have enough capacity to devote to the project.
Last but not least: Since customer data is sensitive data, you should definitely involve data protection officers in good time.
3) Important considerations when planning churn projects
As indicated above, data science projects in general require commitment and resources. While the effort required will not be the same for all roles, advance planning should ensure that necessary resources are available.
For planning the effort itself, a small, yet relevant, experience from our projects is: Customer data is very often fragmented. This means that the above-mentioned data are located in different systems. This not only increases the effort to collect the data, but can require additional effort in harmonization. There is little that can be done about the fragmentation itself within the scope of the project itself, but plan some buffer for data sourcing and harmonization.
4) Focus the project on the goal
Particularly in churn analyses, our experience has been that many interesting aspects and correlations come to light. This quickly leads to the project getting "sidetracked": one obtained insight might raise multiple questions which are also worthy to investigate. Our advice is to start the project with a specific idea or goal and pursue it to the end. If the initial question is: "Which customers are likely to quit in the next three months?", stick to answering it. The question "What actions are the best fit for this customer?" is also exciting and very relevant, but take it one step at a time. The advantage is that, with combined forces, your team will quickly get better results and their team will be ready to benefit from lessons learned in the follow-up project.
This advice is valid both from a global perspective (the project as a whole) and a local perspective (for individual data science steps).
It is advisable to start with a core data set of the most relevant information during data modeling and feature engineering for the machine learning model. Subsequently, one can then focus on further, more detailed data if it has proven to be of interest.
The end goal of such a project should always focus on the eventual user of the solution. Only prediction models that bring real added value to users (in terms of results and quality) have a chance of becoming established. For an optimal result, data teams and business teams need to discuss together what is possible and what is needed. For example, if a phone campaign is planned, a results list should not include customers who objected to being contacted by phone. In the worst case, an improper fit between result and measure can cause the project to fail
We also observed that, in projects without a real use case, the commitment is often lower and these projects unfortunately tend to fail.
Conclusion: Reducing churn - child's play or mammoth task?
This not entirely serious question can of course be answered with a "neither-nor". With a good planning, a machine learning project can be a very good way to reduce costs and increase profits in the long run by using the gained churn predictions. At their core, these projects are similar to other business or data projects, but differ in the flow of certain phases and their activities. There are some best practices that go along with this, which we have summarized for you in this article. If you have any questions about which aspects your team should consider in your individual situation, please feel free to contact us at any time!