Improve Media and Telecom Growth with Churn Prediction Model

Due to the unprecedented growth of the digital economy and the fast-changing technology landscape, the global economy is encouraging the development of digital ventures. For instance, the world’s most valuable brands, such as Apple, Google, Facebook, Amazon, and Microsoft, started as digital ventures and transformed the way we see the world. Just as joining distant continents together by the trade route was an achievement of humankind around 1800, the success of robust digital startups and their unprecedented scalability is an achievement of the 21st century. 

However, one of the biggest challenges for digital ventures is to retain customers on their platform, captivate them to increase usage over the platform, and enhance consumer experience to eliminate the gap between the product’s online and offline user experience. Low customer retention threatens every industry, compels businesses to roll out loyalty offers, cut revenues, increase offering size, reduce order-to-delivery cycle time and burn more cash. The telecom, gaming, media, and entertainment sectors are not untouched by this dire problem. In such scenarios, the churn prediction model helps businesses achieve unparalleled growth by capitalizing on customer segmentation, special offers, loyalty programs, and content recommendations.

Customer Churn

Customer churn is defined as the rate at which customers discontinue their engagement with the company’s products or services. In other words, if a customer stops using a company’s services for a long time, such a customer is considered churned. Customer churn is also called customer attrition. Retaining existing and newly acquired customers positively impacts a business. This not only helps eliminate competition but also helps reduce promotional expenses, boosts referrals, increases customer loyalty, and helps companies gain valuable feedback from long-term customers. 

One of the approaches to combat customer churn is to retarget consumers based on their propensity to churn or to return to the platform with retargeting campaigns and retention offers. 

In addition, it is imperative for businesses to retain customers who possess a higher propensity to spend and promote the platform. The ultimate goal of every digital platform is to convert visitors to engagers, engagers to subscribers (paid customers), and acquire subscribers with a high customer lifetime value. Additionally, businesses of all shapes and sizes must also invest in creating a hyper-personalized platform with customized UX, accessible content, and most-watched content recommendations to increase the visitor-to-subscriber conversion.

Objectives of Churn Prediction Model

A churn prediction model identifies the customers who are most likely to cancel a service or product subscription. The churn model’s objective is to identify the customers with a high propensity to churn in the next seven days, fifteen days, or a month based on the customer lifecycle. A churn prediction model analyzes the feature set, and its correlation to the dependent variable (propensity to churn output) makes it possible to understand why a customer churns.

The churn model can predict the customer’s propensity to disengage and help digital media ventures gain insights on users that possess a high propensity to churn away. It is manageable to retain these customers with various marketing, publishing, content, product design, and customer relationship management strategies, provided we understand who is churning and why. The telecommunication companies and gaming ventures fight the same challenge in everyday operations with competition from low-cost telecom offerings and more publicly appealing games.

Feature Engineering For Customer Churn

Primarily, features are measurable characteristics that a machine learning model considers to predict future outcomes related to churn probability. During the feature engineering process, data scientists create a set of attributes (feature sets) that represent behavior patterns associated with engagement level with a service or product. When trying to predict churn, data scientists consider features in the following broad categories.

Role of Customer Segmentation

Segmenting the customers enables efficient allocation of marketing resources and maximizing potential cross-selling and up-selling opportunities. Customer segmentation streamlines the marketing efforts in sending bulk customized emails with personalized offers. The process of segmenting the customers helps businesses to focus on segments that are most likely to spend on their services, retain on the platform as loyal customers, and improve product quality as per the most preferred segment of loyal customers. A business can segment their customers into the below categories, depending on their recency, frequency, minutes of usage, and money spent on the platform: 

  • Samplers: The bottom 30 percentile of the engagers on the platform
  • Engagers: The middle 50 percentile of the engagers
  • Super engagers: The top 20 percentile of the engagers 
  • Recent engagers: Customers who have started engaging on the platform a few days ago (threshold can be determined with regards to the subscription cycle)
  • Seldom engagers: Customers who haven’t shown up for the past month or quarter, depending on the subscription cycle of the business.

Moreover, it is an added advantage to analyze the churn concerning customer segmentation and include segmentation as a factor of the churn rate. If the churn rate among the sampler segment is 60% compared to 5% in the engagers, there are high chances that the trend in one cohort is very different from that in an adjacent cohort. 

For customers in one segment to migrate to a separate target segment (power engagers in this case), it is paramount to tweak their behavior similar to the target segment. For instance, if power engagers spend three hours a day watching what they love, it delineates their interest in the app and habit formation of watching the content for long duration. Similar habits can be built in samplers by analyzing the week’s busy time and active day, seasonal changes with engagement, propensity to read notifications and respond to it, content affinity, character fandom, and alternatives to similar content.

Role of Recommendation Engine

Recommendations are an automated way of telling engagers what to watch during their active period on the platform. The crucial element here is to link the recommendation engine with the churn prediction model, which aids in altering the recommendation output as per the customer’s propensity to churn and explore the content. Furthermore, the success of Netflix’s recommendation system is sufficient to explain the importance of a recommendation engine.

Points to Consider for building the Model

Data understanding, verification, validation, and analysis

To derive meaning from the model’s output, it is crucial to understand every column, its purpose as a feature, and every value within the column with exploratory data analysis (evaluating Pearson’s correlation, Spearman correlation, PhiK correlation). It is equally essential to gauge the accuracy of the data points, their biases, and reasons for those biases.

Identification of multi-collinearity among the feature-set

A correlation map between all variables helps retain a single feature among highly correlated features. This practice is critical to avoid overfitting due to excessive variables in the feature set. A correlation map also helps in gauging the relationship between the dependent and independent variables.

Evaluate confusion matrix with different test-train proportion

It is of utmost importance to test the confusion matrix values and evaluate the changes by changing test and train as 40:60, 30:70, 25:75, 20:80, 15:85, and more. If precision, recall, and accuracy vary a lot, the model has the potential to learn from noise. It is an excellent practice to dedicatedly test the model on the data of two to three months. If the change in the confusion matrix is not too significant, the model becomes eligible to be deployed in the production environment.

Decide on the model execution cycle and frequency of execution

Undoubtedly, executing a churn prediction model only holds significance if appropriate action is taken to retain the churners. It partially depends on the subscription cycle of the customer. If the subscription cycle is annual, it is helpful to learn about the churn probability of customers whose subscription expiration is just 15 to 45 days away so that the growth marketing team has ample time to re-target them and prevent them from churning. 

Another strategy can be to re-target the customers who have already churned in less than 30-day-window for an annual cycle. For the monthly subscription cycle, proactive retargeting could start from 20 to 7 days before the subscription expiration and up to 30 days after the monthly subscription cycle expires. Retargeting customers with a 40% to 80% propensity to churn would be more cost-effective as they have a 60% to 20% probability of retaining.

Reduce the feature-set

Striking the right balance and achieving an optimized state between all variables in the feature-set are vital to achieve an optimum number of features and avoid overfitting and underfitting. There are few ways to reduce the underfitting such as increasing the model’s complexity, removing the noise from the data, and increasing the number of epochs. 

Additionally, techniques such as limiting the training data, data augmentation, data simplification, reducing the model complexity, early stopping during the training phase by mapping loss function for the training period, ridge or lasso regularization, and using dropout for neural networks can help reduce overfitting by quite a large margin.

Considering feature importance and verifying the change in the confusion matrix is the best way to understand the contribution of variables in improving the model’s accuracy, precision, recall, and specificity. Other feature selection methods like selecting the top k variables, selecting K best, selecting the top percentile variables, and selecting the percentile can help eliminate the non-contributing features.

Precision and recall are inversely correlated, which means when precision increases, recall decreases, and vice versa. Depending on the priority of the use case, the data scientist and business analyst can decide on prioritizing precision or recall with the end goal to strike an agreeable balance between the two by using an F1 score.


One of the limitations of the churn prediction model is that it can accurately predict the customer’s propensity to churn but can not act to retain by itself. In other words, the model is an enabler but not an absolute in itself. Additionally, the model can help with the prediction of churners but not with the rationale behind it. Although the reason for a high user churn can be achieved by statistical data analysis, the external factors that impact churn rate as the pricing strategy of competitors, cyclical behavior, product substitute, industry competition, more evolved and innovative content, targeted and acquired unsuitable customer segment, easy entry of small and low-cost players, targeting more diminutive, and more specific customer segments can only be identified by external market analysis. 

The digital revolution has educated global citizens and organizations on new ways to communicate, collaborate, innovate, win customers’ trust, and promote loyalty. It is difficult to imagine the world without groundbreaking solutions from ed-tech, e-commerce, fin-tech, gaming, agri-tech, pharma-tech, retail-tech, and OTT media services. All these ventures strive to retain their customers in a highly competitive market and achieve exponential growth by referrals. It highlights the needs and significance of an intelligent churn prediction solution in digital ventures and data-intensive sectors like telecom, media, and online gaming. 
Quantiphi has been at the forefront of helping customers leverage intelligent churn prediction solutions that enable digital ventures to retain their customers and monitor deployed models, revising and adapting features to maintain the desired level of prediction accuracy. As a result, we thoroughly understand, “When the customer comes first, the customer will last.” as stated by Robert Half.

To learn more about how businesses can harness the power of churn prediction models to proactively reduce reasons for churn, get in touch with our experts.

Written byPrachi Jain

Get your digital transformation started

Let's Talk