Keeping existing customers is 25 times cheaper than acquiring new ones. Yet subscriber churn remains one of the biggest challenges for telecom companies, which often struggle to act ahead of time. A major telecom operator in Central Asia set out to reduce churn across its corporate and consumer segments. Predictive ML models, developed together with Eastwind experts, made this possible. Read this full case study to learn about the project and its results.

Challenge in B2B: Retaining Corporate Customers with Thousands of Devices

In the B2B segment, the operator serves large enterprises with numerous branches and thousands of employees. Previously, information about contract termination would come in too late, making it impossible to retain corporate customers.

To improve the situation, the operator turned to Eastwind. Our team helped them build a data mart and develop predictive ML models that could accurately identify customers most likely to churn. Based on these insights, the client planned targeted retention efforts for B2B customers.
 

Solution in B2B: Finding the Right Parameter to Predict Churn

On the way to better results, the project team tested three hypotheses using different metrics. Churn prediction was attempted at the following levels:

1. Churn at the subscriber level. Initially, the team focused on the “service termination request” metric. This approach proved ineffective since this stage essentially meant the user had already left. Moreover, in some cases, termination requests were logged in the system months after the actual disconnection.
2. Churn at the corporate client branch level. Here, the focus was on branch-level actions rather than individual subscribers. This approach also failed, as retention efforts came too late—only after an entire branch had already opted out of the operator’s services.
3. Churn at the subscriber level within the framework of a specific branch. To improve accuracy, the team analyzed subscriber behavior inside branches. Instead of “service termination request,” the chosen metric was “no billing activity.” If 50% of subscribers within a branch stopped topping up their accounts, the operator launched B2B retention activities.

This iterative approach to ML model development made it possible to improve prediction accuracy and achieve positive results.

The Evolution of the Predictive Churn Model Target in B2B
 

Challenge in B2C: Finding the Right Approach for Each Customer Across 17 Regions

The telecom operator’s network spans 17 regions of the country, each with its own specifics. The client’s goal was to identify the moment when a subscriber in any given region showed a propensity to switch to a competitor operator. For this, the team first had to determine the data sources, understand which indicators to track, and how to measure them. After that, they began developing the predictive ML model.
 

Solution in B2C: Developing an ML Model for Each Branch

The core of machine learning lies in data preparation—selecting the right data and building a data mart. That’s where the project began. The project progressed through several stages:

1. Populated the data mart with subscriber data from the operator’s internal systems.
2. Integrated external data sources, such as subscriber inquiries from WhatsApp, Telegram, and social media.
3. Developed an ML model to predict subscriber churn. After testing, the results were disappointing: 12% accuracy and 13% coverage. It became clear that regional specifics had to be factored in.
4. Adapted the ML model for each of the 17 branches. The metrics improved: model accuracy rose to 50% and coverage to 30%.
5. Continued training the ML model to further boost the performance indicators.

The Evolution of B2C Churn Prediction ML Models

“For now, we consider the results “average,” but we see clear opportunities to improve them. Our goal is to reach 70% accuracy and 50% coverage. Still, the progress is significant: Whereas before we tried to convince subscribers to stay only after they came to an office to cancel their service, we now proactively retain 30% of them. This greatly improves our chances and reduces the cost of retention.”

Director of the Data Factory Department, Client Representative

 

Business Impact: Reduced Churn and New Opportunities

 

In-depth work with data, machine learning, and an evolutionary approach to developing ML models enabled the project team not only to reduce subscriber churn but also to streamline processes overall:
— Automated the integration of new data sources.
— Made ML models more flexible, allowing regional customization.
— Shifted from post-fact communication to proactive customer retention.
— Opened the door to scaling the approach and testing new hypotheses, such as predicting equipment failures or developing recommendations for retail staff.

“We are constantly working on improving our core use cases that deliver measurable financial results. At the same time, the team generates a wide range of new hypotheses for model development: preventing equipment failures, recommending optimal store locations, and many more. Our main priority now is to keep consolidating data and exploring new opportunities to apply them for the benefit of the business.”

Director of the Data Factory Department, Client Representative

If you’d like to learn more about the practical applications of AI & ML, book a consultation with Eastwind. We’ll help you find solutions to your business challenges and showcase the capabilities of our IT products.