I built the Customer Churn Prediction — scikit-learn project as a production‑style ML notebook on my GitHub (see my GitHub profile). The goal is to flag telecom customers who are likely to cancel so the business can intervene early.
The dataset contains 7,043 customers with 20 features and a churn rate of 26.5 %.
I evaluated three models on a stratified 20 % held‑out test set of 1,409 customers.
Logistic Regression achieved ROC‑AUC 0.841, churn recall 79 % (caught 294 of 374 churners), and precision 51 %.
Random Forest (200 trees, balanced class weight) reached ROC‑AUC 0.827.
XGBoost with imbalance weighting obtained ROC‑AUC 0.809.
I chose logistic regression as the final model because it gave the best ROC‑AUC and recall while staying simple and interpretable, and I packed the model, scaler, and feature schema into a pickle for the Streamlit demo.