Credit Card Behavior Score: IIT Bombay Case Study

Introduction

This project aims to develop a predictive model for calculating the "Behavior Score" of credit card customers for Bank A. The score predicts the probability of a customer defaulting on their credit card, enabling robust risk management for the bank's existing portfolio.

Problem Statement

The task is to build a behavior score using historical development data, predicting the probability of credit card defaults. The model will also be used to predict default probabilities on validation data for operational implementation.

Datasets

• Development Data: 96,806 credit card records with bad_flag indicating defaults.
• Validation Data: 41,792 credit card records without bad_flag, for probability prediction.
• Features Include:
  • On-us attributes (credit limit, etc.).
  • Transaction-level attributes (transaction counts, merchant data).
  • Bureau tradeline attributes (historical delinquencies, holdings).
  • Bureau enquiry attributes (e.g., past enquiries).

Approach

1. Data Preprocessing:
   • Handled missing values and standardized features.
   • Addressed data imbalance using SMOTE.
2. Exploratory Data Analysis (EDA):
   • Analyzed feature correlations and distributions.
   • Investigated default patterns across customer segments.
3. Model Development:
   • Used Random Forest, LightGBM, and Gradient Boosting.
   • Conducted hyperparameter tuning using GridSearchCV.
4. Model Evaluation:
   • Evaluated models on metrics like AUC-ROC, PR-AUC, F1-score, G-Mean, and Matthews Correlation Coefficient.
   • Validated on unseen data, ensuring predicted probabilities were between 0 and 1.

Results

• Developed a high-performing predictive model for credit card default risk.
• Provided actionable insights into customer behavior and portfolio risk.

Repository Contents

• Code: Python scripts for data preprocessing, modeling, and evaluation.
• Insights: Key findings from EDA and model interpretations.
• Documentation: Detailed steps, metrics, and observations.

Usage

1. Clone the repository.
2. Install required libraries using requirements.txt.
3. Run the scripts to preprocess data, train the model, and make predictions.

Evaluation Metrics

• Threshold-Independent: AUC-ROC, PR-AUC.
• Threshold-Dependent: F1-Score, Confusion Matrix.
• Business Metrics: Weighted Cost Analysis.
• Class Imbalance: G-Mean, Matthews Correlation Coefficient, Cohen's Kappa.

Conclusion

This project delivers a robust behavior score model, enabling Bank A to manage credit risk effectively. The insights and predictions support strategic decision-making and portfolio optimization.