Preventing churn in the telecommunications industry
A predictive model identifies telecommunications customers at risk of leaving so retention teams can prioritize action.
A predictive model identifies telecommunications customers at risk of leaving so retention teams can prioritize action.
Challenge
A telecommunications company needed to identify customers likely to leave for a competitor. Churn reduces future revenue, removes opportunities to offer additional services and can affect brand reputation.
Data
The dataset contained fewer customers who left than customers who stayed, creating a class imbalance. This can cause conventional models to favor the majority class and overlook customers at risk of leaving.
Individual variables showed low logistic correlations with churn, below 10%, so predicting churn required a model that considered multiple factors together.
Solution
Artelnics developed a predictive model using an algorithm that minimizes weighted quadratic error to address the class imbalance.
The evaluation considered sensitivity and specificity alongside the other classification measures. For an imbalanced dataset, apparently strong overall measures can conceal poor detection of customers who actually leave.
Predictive modeling with weighted quadratic error.
Identification of customers at high risk of churn.
Results
The reported classification test gave 67.80% precision, 32.20% error, 74.47% sensitivity and 67.59% specificity. The area under the ROC curve was 0.78.
| Classification test | |
|---|---|
| Precision | 67.80% |
| Error | 32.20% |
| Sensitivity | 74.47% |
| Specificity | 67.59% |
The model helps retention teams identify high-risk customers and prioritize opportunities to prevent churn. Some customers flagged as high risk may remain, so predictions should support retention decisions rather than be treated as certain outcomes.
Illustrations

