Data scientist with a statistician's discipline.

Reproducible case study

Customer churn prediction & retention review

Which customers should a retention team review when capacity is limited? I compared five models on the 7,043-record Telco sample, selected logistic regression through training cross-validation, and used a separate validation set to choose the review cutoff.

Test ROC-AUC

0.8402

Churn recall

80.7%

Precision

51.3%

On 1,409 test profiles, the model flagged 302 of 374 churners and 287 non-churners. The result makes the review workload visible, without claiming that a prediction prevents churn.

A shared training and scoring pipeline, an executed notebook, and a local dashboard support individual and batch predictions. Permutation importance explains model reliance; 14 automated checks verify the workflow.

Python · scikit-learn · XGBoost · cross-validation · local dashboard