Churn amongst existing subscribers is a big issue for most publishers. A lot of marketing resources are allocated in contacting the current customer base to offer them a discount or promotion so they would prolong their subscription. Given the fact that a lot of the people that are contacted and offered a discount never actually intended the churn, this is a huge amount of wasted income — not to mention the big costs associated with contacting all of these people.

This is why de Persgroep developed a churn model that will predict the probability that a particular subscriber will cancel his subscription. Making this model and identifying potential churners is actually the easy part.

A big hurdle in any data science project is convincing the business stakeholders of its worth and start trusting the numbers rather than their gut feelings. It is important to align all people involved in the project and to present results in the appropriate manner. A model cannot be an artefact developed at a data lab. It should be a joint effort so it is trusted, put into production, and relied upon. 

Challengers will be interested in seeing lift curves, accuracy, recall. Proponents will want to understand the algorithm you used. Skeptics want to have insights in the importance of features. Others trust you and just want to hear about accuracy so they can allocate budgets based on these numbers and make decisions.
Challengers will be interested in seeing lift curves, accuracy, recall. Proponents will want to understand the algorithm you used. Skeptics want to have insights in the importance of features. Others trust you and just want to hear about accuracy so they can allocate budgets based on these numbers and make decisions.

In close collaboration with the marketing and operations departments, we set up an automated retention flow. Based on the churn score of a subscriber, one of many flows gets activated. The higher the risk that a subscriber will churn, the more marketing effort we put in.

People with a low risk of churn receive a service call to ask if they are happy with the product, and they receive a standard renewal offer when their subscription comes to expiration. Customers with a high risk of churn are contacted by e-mail, direct mail, and outbound call before their subscription expires.

The new retention flow means only subscribers with a high likelihood of churn would be contacted pro-actively with a special offer before their subscription expires. Those with low churn likelihood are contacted later with standard renewal offers.
The new retention flow means only subscribers with a high likelihood of churn would be contacted pro-actively with a special offer before their subscription expires. Those with low churn likelihood are contacted later with standard renewal offers.

Another important integration was adding these churn scores to the subscription service. In the past, people would call in with a question about their subscription and would be offered a discount if they would renew their subscription for another year. This offer would be presented to most people who called, because the operators in the call centre could not know whether this person was likely to churn. 

So we now offer call centres a “crystal ball” in the form of an indicator (green or red) that appears on their screen, based on which they can decide whether or not to present the customer with a special offer for renewal.

Data experts also compare their churn forecast vs. the actual churn experienced.
Data experts also compare their churn forecast vs. the actual churn experienced.

The final step in the integration of a model as a foundation for our marketing efforts into reducing churn was closing the loop. The performance of the model was presented in a dashboard. Since the start of the automated flow, an average decrease in churn of more than 18% was achieved. This joint effort managed to significantly reduce the number of subscribers that churned and entailed an increase in the yearly ARPU of over 2%.