Bank Customer Churn Analysis

Classification, survival analysis, and a live risk predictor for bank customer retention

Demo

What this shows: a walkthrough of the Bank Customer Churn Risk Predictor — filling in a customer profile (credit score, geography, age, tenure, balance, products held, and activity status), then running the trained XGBoost model to get a live churn probability, risk level, and revenue-at-risk estimate.

Interactive dashboard

Open dashboard in full screen

Project information

Predicting and explaining customer churn

This project combines a classification model with survival analysis to answer two questions a retail bank actually needs answered: which customers are likely to churn, and when in their lifecycle is that risk highest. Three models (Logistic Regression, Random Forest, XGBoost) were compared, then explained with SHAP, and paired with a Cox Proportional Hazards model to translate risk into interpretable hazard ratios — the kind of dual approach an actuarial or risk team would expect.

Project deep dive

Problem statement

Customer acquisition costs far more than retention, but most churn models stop at "will this customer leave?" — which tells a bank who to worry about but not when to act or how much is actually at stake. This project was built to answer all three: who is likely to churn (classification), why — which factors actually drive the decision (SHAP + Cox hazard ratios), and when in the customer relationship the risk peaks (survival analysis), then translates that into a revenue-at-risk ranking a retention team can act on directly.

Dataset

Kaggle "Bank Customer Churn" dataset — 10,000 customers, 2,037 observed churn events (~20% churn rate).

  • Features: CreditScore, Geography, Gender, Age, Tenure, Balance, NumOfProducts, HasCrCard, IsActiveMember, EstimatedSalary
  • Target: Exited (1 = churned, 0 = retained)

Methodology

  1. Exploratory analysis — churn rate broken down by geography, activity status, product count, and age. Germany shows a notably higher churn rate than France/Spain; inactive members and customers holding 3–4 products churn far more than average.
  2. Feature engineering — one-hot encoding for Geography, binary encoding for Gender, plus two engineered signals: BalanceSalaryRatio (disengagement signal) and EngagementScore (activity + card ownership combined).
  3. Classification — Logistic Regression, Random Forest, and XGBoost compared using AUC, precision/recall, and a confusion matrix rather than plain accuracy, since the target is imbalanced.
  4. Explainability — SHAP values on the XGBoost model to rank the true drivers of individual predictions, not just global feature importance.
  5. Survival analysis — Kaplan-Meier curves (overall and stratified by product count) plus a Cox Proportional Hazards model, using Tenure as duration and Exited as the event flag, to model when risk peaks rather than just whether it occurs.
  6. Business layer — churn probability weighted by each customer's balance to estimate revenue at risk, ranked into a prioritized retention list for the top 10% highest-risk customers.

Key findings

ModelAUCRecall (churn class)
Logistic Regressionbaseline—
Random Forest——
XGBoost0.8680.744
  • Cox model (n=10,000, 2,037 events, concordance = 0.70): active members face a 54% lower churn hazard (HR 0.46); each additional year of age raises hazard by ~5%.
  • The linear NumOfProducts term looked statistically insignificant on its own — but SHAP and Kaplan-Meier stratification both revealed a sharp, non-linear risk spike at 3–4 products. Binning it into a categorical flag recovered a significant effect, a methodological finding worth flagging in its own right.

Tech stack

Python pandas scikit-learn XGBoost SHAP lifelines Plotly Streamlit

Full notebook, priority retention list, and dashboard chart exports are in the GitHub repo →