Understand the signals. Anticipate the exit.
Comparing classification models to understand telecom churn risk and inform retention decisions.
Can the warning signs become useful?
A cancellation is a late signal. Historical account and service data can help identify patterns associated with customers leaving before a retention conversation becomes impossible.
Estimate churn risk, compare model approaches, and explain the customer characteristics associated with churn in language a business team can use.
Prepare customer-level features
Cleaned billing fields, encoded categorical variables, standardized numeric features, and developed tenure-related features from more than 7,000 telecom customer records.
Compare three approaches
Used a stratified 80/20 train-test split and compared Logistic Regression, Random Forest, and Gradient Boosting through Scikit-learn pipelines. Stratification preserved the churn distribution between the training and test groups.
Translate risk into action
The reported strongest overall model was Gradient Boosting, at approximately 0.84 ROC-AUC. Contract type and tenure were primary indicators, informing recommendations for early engagement and longer-term contract incentives.
A ranking measure, not “84% accurate.”
ROC-AUC describes how well a model distinguishes churners from non-churners across thresholds. A value of 0.5 is chance-level ranking; 1.0 is perfect separation on the evaluated data.
A useful decision still needs a threshold, the cost of outreach, and the cost of missing an at-risk customer.
What the work delivers.
A documented model comparison with approximately 0.84 ROC-AUC and business recommendations focused on onboarding, contracts, billing, and service value.
Model quality and business usefulness are different questions. Risk ranking needs an interpretable explanation, and a retention team still needs to choose a threshold that fits its capacity and costs.
Project scopeThe reported result is from a held-out project test set. It does not establish production performance or a measured reduction in churn. No live prediction service is presented.