ChurnPulse

Problem

Many churn models return a score with no reasoning, so the team does not know what to do next. Customer data was also messy: missing fields, duplicates, and imbalanced labels.

Approach

I cleaned the data with a reproducible pipeline, compared logistic regression, random forest, and gradient boosting, then chose the model with the best recall/precision trade-off. I added permutation importance and a compact SHAP summary so every prediction has an explanation.

Key features

• Deterministic data cleaning pipeline.
• Three-model comparison with cross-validation.
• Batch and single prediction endpoints.
• Risk-factor summary per customer.

Results

1. 01

   ROC-AUC of 0.86 on a stratified hold-out test set.

2. 02

   Churn-class recall rose from 0.61 to 0.78 after threshold tuning.

3. 03

   The retention team gets a daily priority list, not just a model in a notebook.