atdokmeci/churn_model

Bank Customer Churn Prediction Model Card

Model Details

• Architecture: Artificial Neural Network (ANN) with 2 hidden layers (12 and 6 units, ReLU activation), output layer (1 unit, sigmoid activation).
• Framework: Keras (TensorFlow backend)
• Input Features: 14 normalized and engineered features:
  • CreditScore, Gender, Age, Tenure, Balance, NumOfProducts, HasCrCard, IsActiveMember, EstimatedSalary, BalanceSalaryRatio, TenureByAge, Geography_France, Geography_Germany, Geography_Spain
• Output: Binary classification (Exited: 0 = retained, 1 = churned)

Intended Use

• Predict whether a bank customer will churn (exit) based on their profile and account activity.
• Useful for financial institutions to identify at-risk customers and take retention actions.

Training Data

• Dataset: Custom bank churn dataset (churn.csv)
• Size: 10,000 samples
• Split: 80% train, 20% test
• Preprocessing: Feature engineering (BalanceSalaryRatio, TenureByAge), categorical encoding, min-max scaling.

Metrics

• Loss: Binary cross-entropy
• Accuracy: ~81% on test set
• Evaluation: Confusion matrix, classification report

Limitations

• Model trained on a specific dataset; may not generalize to other banks or regions.
• Sensitive to feature distribution and preprocessing steps.
• Does not explain feature importance.