Customer Churn Prediction System Using Machine Learning: A Case Study ROSHAN Telecom-Afghanistan

Ilusstration by AI

FORMOSA NEWS

Afghanistan– Machine Learning Model Predicts Telecom Customer Churn with 95.32% Accuracy. A study conducted by Hafeezullah Shoja of Kabul University, Esmatullah Sabet of Parwan University, and Sayed Shafiullah Sadat of Polytechnic University Kabul was published in the International Journal of Integrated Science and Technology (IJIST) Vol. 4 No. 2 (February 2026).

A new study by Hafeezullah Shoja of Kabul University, Esmatullah Sabet of Parwan University, and Sayed Shafiullah Sadat of Polytechnic University Kabul introduces a machine learning–based churn prediction system developed for ROSHAN Telecom in Afghanistan.

Why Customer Churn Matters

In the telecom industry, acquiring new customers is significantly more expensive than retaining existing ones. Previous research cited in the study highlights that increasing customer retention by just 5 percent can boost profits by 25 to 85 percent.

Customer churn prediction allows companies to intervene early—offering targeted promotions, improving service quality, or addressing complaints—before customers switch providers.

Processing 100,000 Customers from a 7-Million Database

The research team collected one year of real-time operational data from ROSHAN Telecom. The data sources included:

• Call Detail Records (CDR) for voice calls and SMS
• Internet usage records
• Daily TOPUP transactions
• Bundle subscription activity
• CRM data such as gender, province, and complaint tickets

As detailed on pages 131–133 of the journal, the final dataset included variables such as:

• MSISDN (customer number)
• Gender
• Age on Network (AON)
• On-net and off-net call usage
• Data consumption
• Bundle activation count and amount
• TOPUP frequency and amount
• Total complaints, inquiries, and requests
• Churn flag (1 = churned, 0 = active)

From a base of approximately 7 million subscribers, 100,000 customers were randomly selected for model development. The dataset was split into 80 percent training data and 20 percent testing data.

Why XGBoost?

The XGBoost algorithm was selected for its ability to:

• Handle large datasets efficiently
• Process both numerical and categorical features
• Manage imbalanced data effectively

As discussed in the literature review (pages 127–130), previous studies consistently show XGBoost outperforming other machine learning models such as Decision Trees, Random Forest, and Support Vector Machines in churn prediction tasks.

Hyperparameter tuning was conducted to optimize performance, and evaluation metrics included accuracy and confusion matrix analysis.

Results: 4,127 High-Risk Customers Identified

The model successfully identified:

• 4,127 customers at high risk of churn
• Overall prediction accuracy: 95.32 percent

According to the confusion matrix presented on page 135:

• 35,604 customers were correctly classified as non-churners
• 12,089 customers were correctly classified as churners
• 1,510 false positives
• 797 false negatives

These results demonstrate that the model effectively handles class imbalance while maintaining strong predictive reliability.

Strategic Insights for Telecom Management

Beyond statistical accuracy, the churn prediction system provides actionable insights for management.

The model segments high-risk customers based on:

• Gender
• SIM profile type
• Province
• Age on Network (AON)

This segmentation enables telecom managers to:

• Design targeted retention campaigns
• Focus promotional offers in high-churn provinces
• Analyze complaint patterns linked to churn
• Improve service strategies for long-term subscribers

The system shifts decision-making from reactive to proactive customer retention.

Limitations and Future Development

While the model shows strong performance, it relies primarily on historical data. Sudden behavioral changes may not be fully captured.

Future research directions include:

• Integrating social media sentiment analysis
• Exploring deep learning and neural network models
• Developing near real-time churn prediction systems

These improvements could enhance model adaptability in rapidly evolving markets.

Broader Industry Implications

Although developed for ROSHAN Telecom in Afghanistan, the findings are relevant for telecom operators globally—especially in emerging markets where competition is intense and profit margins are tight.

By leveraging machine learning:

• Churn rates can be reduced
• Customer loyalty can be strengthened
• Marketing costs can be optimized
• Long-term revenue stability can be improved

The study demonstrates how data-driven digital transformation can directly support business sustainability in the telecom sector.

Author Profiles

• Hafeezullah Shoja-  Kabul University.
• Esmatullah Sabet-  Parwan University, Afghanistan.
• Sayed Shafiullah Sadat-  Polytechnic University Kabul.

Research Source

Shoja, H., Sabet, E., & Sadat, S. S. (2026). Customer Churn Prediction System Using Machine Learning: A Case Study ROSHAN Telecom-Afghanistan. International Journal of Integrated Science and Technology (IJIST), Vol. 4 No. 2, 123–137.

DOI: https://doi.org/10.59890/ijist.v4i2.287

URL: https://ntlmultitechpublisher.my.id/index.php/ijist