Numerous questions arise in numerous projects. We have collected our most frequent ones here for you and provided basic answers:
- Is my use case suitable for churn prediction?
- What are success factors for such a project?
- How can I calculate a business case for churn analysis or prediction?
- What are common reasons for churn projects to fail?
- Is an AI always the best solution for churn prediction?
- Which AI algorithms are suitable for churn prediction?
- How much precision can I expect in the predictions?
- What would a pilot project/proof-of-concept for churn prevention look like?
- What technical foundations must exist in order to create churn analyses and churn predictions?
- How do I identify appropriate measures to keep customers from quitting?
1) Is my use case suitable for churn prediction?
How suitable your use case is for churn analysis and prediction depends fundamentally on two criteria: the type of customer relationship and the data collected.
Churn prevention usually adds value when customers have a longer relationship with the company and bring lasting financial value. Otherwise, anti-churn measures may not be worth the effort. Longer customer retention is also key to the second criterion: it is important to have data from the entire customer journey, e.g., cancellation dates, contract end dates, contact history, and all touch points. Prediction is difficult if the customer history cannot be fully reconstructed.
A typical limiting factor happens when data is overwritten during the customer journey. This could happen, for example, when keeping track of contacts: If a contact request via a web form ends in a phone call, both should appear in the data records and not only the last contact, here the phone call. If intermediate steps are erased, it will be impossible to train the machine learning model on accurate historical data and the machine learning model will not be able to make correct predictions.
2) What are success factors for such a project?
Success factors are different for each project. The following have emerged as key factors:
- Get the right team on board. It should include a mix of business experts and data experts. Also involve the IT team, data protection, etc. early on.
- Select a relevant, feasible use case that offers real value to your employees. This will encourage engagement and subsequent adoption. Formulate a concrete goal for the use case. This prevents distraction and "bloating, endless projects".
- Check the underlying data early on, and absolutely before you start data modeling. If data from the customer journey, especially interaction data are not collected so far, this should be changed as a first step.
You can also read more about this here in our best practices for churn prevention projects.
3) How can I calculate a business case for churn analysis or prediction?
For the calculation of business cases, current costs are compared to the sum of future costs and necessary investments. For example, for a hypothetical company, the calculation could look like the following.
In the area of churn, the current cost side may include:
- Lost revenue: "One customer brings 30€ revenue per month. With 1,000 cancellations per month, we lose €360,000 per year."
- Compensating investments to compensate for churn through new acquisition: "To acquire an additional 1,000 new customers each month, we have to invest €60 per customer in sales and marketing. That adds up to €720,000 a year."
The future cost side - if churn prediction and prevention is used - includes future savings and the required investments:
- Reduced lost revenue: "One customer brings in €30 in revenue per month. With only 500 churns per month, we are missing out on €180,000 per year, €180,000 less than previously."
- Reduced compensating investments: "To attract only 500 additional new customers each month, we have to invest 60 € per customer in sales and marketing. That adds up to only €360,000 instead of €720,000 per year."
- Costs of a churn project: These can vary greatly - for example, they depend on the existing infrastructure, the data basis, its quality and the activities that the company itself undertakes. Let's assume 30 000 € in this example. (For individual estimations just write us!)
- Internal costs for a churn project: These also vary greatly and depend largely on the team involved, its size and the existing data infrastructure. In this example, let's assume €5,000.
- Costs for churn measures: "To keep 500 out of 1000 customers at risk of switching, we have to invest 10 € per customer per month as a discount. That adds up to €60,000 a year."
Churn projects are particularly profitable in the long term, so choose a suitable time window that supports your argument but does not obfuscate.
4) What are common reasons for churn projects to fail?
The 3 most typical reasons are:
- Missing, unsuitable data with insufficient quality or overwritten data
- Insufficient data preparation: Data cleaning and transformation are crucial for the validity of the final model and should be performed very thoroughly.
- Lack of user acceptance or poor solution-user fit: the solution is simply not well suited to be used in real operations
Other risks include, for example, problems with data protection or data processing authorization, overly complex data provisioning and poor planning, or insufficient resources. We are happy to support you from start to finish in planning and implementation!
5) Is an AI always the best solution for churn prediction?
No, a machine learning model (which is what is usually meant by AI) is not the best choice for every business and every issue. Perhaps the churn analysis previously performed reveals that there is a very specific reason for termination that can be solved in a different way. Another scenario would be that there are a handful of clearly defined risk factors on which the customer base can be more efficiently filtered directly. In such cases, the effort of maintaining the AI must generally be compared to the expected benefit.
6) Which AI algorithms are suitable for churn prediction?
This question cannot be answered in such a general way. 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.
Depending on the size of the data, some models may not be adequately trained (e.g., gradient boosted trees require thousands of data points for training; neural networks tend to require even more). The type of data you have about your customers is also critical. Some models work better with categorical values (i.e., places of residence, gender, etc.) than others.
The choice of an algorithm can usually be narrowed down to the most promising approaches after the analysis phase. Selected algorithms can then be tested and compared.
7) How much precision can I expect in the predictions?
No one can guarantee precision of the final machine learning model at the beginning of a project. The predictive power depends largely on the data quality and the information mapped in the data. Beware of cheap promises.
The challenge with churn predictions is to distinguish the small percentage of churning customers from the large number of non-churning ones. This imbalance (e.g., 2% quitters versus 98% non-quitters) makes it difficult to identify characteristics of the churn-prone because it is very likely to detect the same characteristics in the much larger group of non-churners.
Precision is nevertheless critical to whether a machine learning model ends up being useful to your team at all. Our recommendation is therefore to discuss at an early stage together with the specialist departments and the data team what the requirements are for a good result. The data team can set priorities for optimization accordingly. If it is foreseeable that the desired performance will not be achieved, the business department may have to rethink their approach. For costly and labor-intensive churn-prevention measures, the accuracy should of course be very high; more general and less expensive measures would be an alternative if lower accuracy is expected.
8) What would a pilot project/proof-of-concept for churn prevention look like?
If you are not sure whether a churn prediction project is feasible with the data you have, or whether churn prevention will help you achieve your business goals, you can first start with a pilot project or a proof-of-concept (PoC). PoCs should be faster, more experimental, and less complicated in terms of setup.
Based on a very specific question, a data extract is shared via file exports (e.g. as CSV). After exploration and modeling in a flexible environment, the results are exported as a file in an equally straightforward manner. The results are then used to test and evaluate performance according to the initial question.
PoCs of this kind are significantly less expensive than holistic implementations and take just a few weeks if executed well.
9) What technical foundations must exist in order to create churn analyses and churn predictions?
The fundamental requirement is that your company captures customer data (master data and interaction data) in a structured way. The software you use matters less, what is important is that you get to the underlying data. If data is exported as a file, a location accessible to the data team is needed.
For the other steps, such as setting up a data science environment, data processing and data science activities, there are excellent open source tools that can be set up quickly. So if your company is not yet on a cloud platform, that is definitely not a deal-breaker for churn prevention projects. If you have questions about how such an environment could be designed at your company, feel free to write to us for a free initial consultation!
10) How do I identify appropriate measures to keep customers from quitting?
This is the mother of all questions of churn prevention: what actions should be taken? Chances are your team already has plenty of ideas. In the data analysis phase of churn projects, insights gained often provide clues for prioritizing possible actions. An example from the energy industry: if many young adults cancel from classic tariffs, but fewer from "green" tariffs, a discount for switching to an eco-tariff can be an effective way to retain this customer group. Clustering analyses are often helpful for developing ideas and prioritization.
We made the best effort to answer the question in general. Nevertheless it is always possible that the questions in your individual case would have to be answered differently or with restrictions! Therefore, please feel free to talk to us directly so we can answer your questions specifically.