Going beyond churn prediction to support customer retention

Taking advantage of machine learning methods

Catarina Freitas

Published in

Marionete

5 min readMay 20, 2021

Understanding the problem

One of the most valuable efforts for any company is to invest in keeping their customers happy. Long-term positive relationships between customers and brands increase loyalty and positive customer references. But, it is also cheaper for brands to keep existing customers satisfied than to attract new ones with lower prices and costly marketing campaigns, especially in saturated markets lined with strong competitors.

According to a European Commission report [1], about ten years ago, the average churn rate was 10% in Europe and more than twice as high in the UK. However, the rate and the indicators of churn can depend strongly on the sector, as the statistics below (image 1) show. While the Energy sector is usually more monopolized and less prone to churn, the Telecommunications sector is one of the most competitive. For instance, in Portugal, where the Telco churn rate is 18% [2], the main reasons for customers changing providers are competitive prices and better bundle packages. This is a notable example of how it is much more expensive to acquire new customers than keep the current ones happy because enticing payment plans ultimately lead to less revenue.

Image 1 — Data from the European Commission report, 2008 [1]. Image by author.

In the current digital era, new types of businesses have become more significant. A Deloitte report [3], found that before the pandemic 69% of the US market subscribed to at least one streaming video service, 41% to a music streaming service, and 30% to a gaming service. Since COVID-19, however, streaming video subscriptions jumped up to 80% [4] and the churn rate increased 6%, totaling 41% [5]. Streaming video subscriptions experience twice as much churn as the Telco companies, showing that streaming is a highly competitive business, with less apparent loyalty to a specific company and more likely to experience customer cancellations and switching, based on the entertainment and the pricing options available.

A Comprehensive Strategy

What this tells us is that with an increased rate of churn in different sectors it is important to understand what motivates customers to cut ties with a brand. Companies hold a lot of data that can be used to study the patterns of customer behavior, such as: demographics, longevity, technical support requests, and changes in service usage or payment methods. All of this highly valuable information can be leveraged to build machine learning models to predict customer churn. But can companies reduce customer churn effectively just by predicting which customers are at risk of leaving?

An effective strategy must also give insights into which clients are worth acting on and how. For instance, some customers might be considered a lost cause and will leave irrespective of the deals offered. Other customers might be dormant in which case a retention campaign can actually have a negative effect. Also, given budget limitations for promotions, we should also identify which customers are the most profitable ones to prioritize.

Therefore, to effectively reduce churn, companies should focus on the retention effect of their marketing campaigns. This can be achieved with uplift models instead of simply churn prediction models. Using the outcome of previous campaigns, uplift models can estimate the added value of each promotional action to retain a client. The predicted uplift for each action is then weighted by the Customer Lifetime Value and balanced with its implementation cost. This results in a comprehensive and actionable value that can be used to find the segment of customers that are both valuable and persuadable and approach them in the way that is expected to bring the best reward.

Explaining your AI model

So how do these models help businesses make such critical decisions to help tackle customer churn? By looking at how the most relevant variables contribute to the prediction outcome we can find some important answers.

Here we show an example using a real-life dataset from a music streaming service [7]. The task was to predict whether a user would choose to abandon the service once the subscription expired or whether they would take out another subscription. The data analyzed included specific details relating to: transactional information, the listening habits of the clients, and their demographics.

After training a tree-based model that predicts which customers will churn in the following month with 60% precision and 61% recall, we turn our attention to study the effect of the variables on churn probability.

Image 2 — How the number of active days in the previous month affects the model’s likelihood of churn. Image by author.

In image 2 above we see that the probability of churn is higher when the customer didn’t use the service in the preceding month and the probability decreases with an increase of activity. Additionally, we can see this trend for two separate groups, differentiated by lower and higher values of automatic membership renewals. Since a lower auto-renewal is a risk factor for churn, this group is more affected by the influence of other variables.

Image 3 — Segmentation of the 100 most likely churners against the rest of the customer's group according to the model output probabilities. Data includes 25th, 50th and 75th percentiles for each group. Image by author.

Image 3 shows the well-pronounced differences between the characteristics of the most likely customers to churn and the overall population of customers. Top churners are less prone to memberships that renew automatically, have shorter subscriptions, have significantly less activity in the month prior to the churn event, are more recent users of the service, have higher variation in their subscription plans length, and higher number of alterations or cancelations to their membership.

With this type of analysis, we can enrich our understanding of the model and also assess if its decisions make sense in terms of the business, working as a quality control process.

Final notes

Companies care about preventing churn because it can have a meaningful impact on their revenue.

A strategy that maximizes business value finds clients that are worth targeting with retention campaigns and also estimates what action is best suitable for each considering its cost-benefit.

Data science specialists can leverage companies’ data to build machine learning models that answer these questions with actionable information to the business. The data is at hand and the retention campaigns are already in place. The focus is to optimize the investment in these campaigns and increase their efficacy.

Finally, specialists also add the capability of providing analytics on how the models make their decisions, giving insight on the risk factors of churn and enabling quality assurance.