About this project
Objective, data source, methodology and limitations of the ChurnIQ churn prediction platform.
Objective
Identify which telecom customers are likely to churn, quantify the revenue exposure of that churn, and surface segment-level patterns that can guide retention action. The app combines descriptive analytics with a supervised classification model.
Dataset
- IBM Telco Customer Churn dataset, bundled with the app (no external service calls).
- 21 original columns per customer: demographics, subscribed services, contract, billing and the churn target.
- Row counts appear once the dataset finishes loading.
Cleaning rules
- Schema validation: the required 21 columns must be present or the app reports an error instead of guessing.
- TotalCharges converted to numeric; blank values for zero-tenure customers imputed as tenure × MonthlyCharges.
- Rows with an invalid churn label or a duplicate customer ID are removed and counted in the data-quality report.
Engineered features
tenure_group— 0-12, 13-24, 25-48, 49-72 months.total_services— count of active services per customer.has_security_support— online security or tech support attached.high_monthly_charge— above the dataset median monthly charge.average_monthly_value— total charges divided by tenure.
Model pipeline
- Stratified 80/20 train/test split with a fixed seed for reproducibility.
- Preprocessing fit on training data only: median imputation and standard scaling for numeric features, one-hot encoding with a dropped reference level for categorical features.
- Logistic regression trained by full-batch gradient descent with L2 regularisation and balanced class weights.
- Evaluation on the held-out test set plus 5-fold stratified cross-validation.
Limitations
- The dataset is a static historical snapshot; it contains no live customer behaviour.
- Relationships shown are correlations, not causal effects.
- Logistic regression is a linear model; it will miss complex interaction effects.
- Risk thresholds (0.30 / 0.60) are project conventions, not industry standards.
- Predictions support human decisions — they should not be used to act on a customer automatically.