This blog post is based on a previous customer case from 2019. The projects were found and structured via a Data Science Workshop and then implemented as Analytics jumptstart. This post aims to inspire and to describe how to work more proactively with a customer base using Data Science and Machine Learning.

Retaining customers is one of the foremost growth pillars for companies, particularly for those with a subscription-based business model. Competition is hard on the market where customers can often choose between several suppliers within the same product category. Repeated poor experiences – or even just one poor experience – can prompt a customer to switch suppliers. It is therefore increasingly important to be relevant in one’s marketing and to use sales resources in an optimal way. Manually creating customised experiences for each customer quickly becomes time-consuming and complex. Instead, this should be done in a more scalable way using smart tools – such as machine learning. 

However, it is not only external communication that can benefit from the application of machine learning. Even internal processes relating to the customer’s journey can be enriched and made more efficient with the help of forecasts. One example of a term that is often used in the drive to understand and monitor the customer journey is CHURN – when the company loses a customer.

CHURN occurs when a customer stops doing business with a company. The percentage of customers that stop using a company’s products or services for a given period is known as the CHURN rate and it is often an indication of the health of a company whose customers are subscribers who pay for services on a recurring basis.

The cost of acquiring a new customer is usually higher than the cost of retaining an existing one. One common method for many companies is to reactively attempt to win back customers who have already terminated their subscription or who have stopped being customers (known as win-backs). By instead predicting which customers are likely to stop making purchases within a given future, companies can proactively try to engage with these customers before they leave.

At the same time, it is important to understand what is meant by CHURN and which insights are desired with this analysis. In short, companies must decide which question they want to ask and after that, what kind of problem they have to solve. The way CHURN is defined and the type of analysis that is suitable differs between companies as a result of factors such as type of industry and business model.


  • Which customers are likely to leave the company?
  • Within what time perspective are different types of customer likely to leave the company?
  • Which customers are not likely to renew their subscriptions?
  • Which customers are likely to downgrade their price plan?

Machine learning identifies and utilises historical data to solve a given problem. It is therefore important to save all interactions with all customers in an accessible format. One major advantage of using machine learning is that the tool can handle a lot of data about a customer, several hundred customer attributes are not a problem. Thanks to this scalability, this kind of algorithm is able to identify more complex patterns in the data than human beings can manage in the same space of time.


  • Demography: Basic information about a customer such as age, gender and residence.
  • Consumption: How and what kind of service or product the customer uses. For example type of contract, number of log-ins, length of sessions, number of purchases per product group and so on.
  • Interaction:How the customer communicates with the company for example through customer support (number of calls or questions), via mobile phone or computer, customer satisfaction and so on.

What remains is to select and adapt a machine learning algorithm to the historic customer images, where those customers that have CHURN-ed out of the company and those that remain are clearly identified. By regularly applying the algorithm to the existing customer base, what you get is a predicted CHURN probability for each individual customer, which can be used to enrich processes for handling the existing customer base.

In one of the projects, machine learning achieved a CHURN precision rate of 70 %. This means that when the model reports a CHURN, it is correct 70 % of the time. At the same time, it had a recall rating of 71 %, which means that of all those customers who left the company, it identified 71 % of them. Both we and the customer felt this was a good result – after all it is people’s behaviour we are trying to predict, something that can impact factors unknown to many of us, factors that are not reflected in the data. If you would like to read more about this particular project, there is a brief description on the website.

The CHURN analysis does not generate any business value if the company does not act on it. This can be done in several different ways and at many different levels. One first step is usually to deal in a separate segment with those that show a high risk of leaving. This segment may for instance automatically receive offers via email by utilising a marker in a marketing automation tool, or it may be flagged for a call by a salesperson. Having said that, prioritisation of which customers should be approached in this way should take into account several factors that reflect a potential gain from each customer, not just the risk that they may leave.

One common proactive key performance indicator here is CLV (Customer Lifetime Value), where the aim is to predict how much the company is expected to make from a given customer during this customer’s expected lifetime (or at least for a longer period in the future). By combining this with a CHURN prediction carried out earlier, companies can gain an understanding of which customer relationships are profitable and are also in the CHURN risk zone. Together these key performance indicators become much more interesting in the process of prioritising customer relationships and any subsequent customer contacts. Because companies do not want to spend time or money on customers from whom they will not make money.

Yet another step can be to add how much it would cost to retain the customer in the form of marketing, special offers and sales resources for each communication channel. The priority list is now an excellent business intelligence basis but it can be too complex to handle manually. Instead, the company can optimise its marketing with the help of decision optimisation. That is to say, which customer should receive which offer in which channel, given the applicable business regulations, budget and other limitations.

Do not hesitate to get in touch if there is anything to want to ask. Just send an email to to discuss the issue further.