Credit Card Behaviour Score Prediction Using Classification and Risk-Based Techniques

Overview

This project aims to enhance the credit risk management system of a bank by predicting the likelihood of a credit card customer defaulting on their payment in the following month. Using anonymized historical behavioral data from over 30,000 customers, we build a binary classification model that flags potential defaulters early, enabling proactive risk mitigation strategies.

Objective

• Build a classification model to predict next_month_default (1 = Default, 0 = No Default).
• Perform exploratory and financial analysis to identify behavioral patterns associated with credit default.
• Handle class imbalance using techniques like SMOTE, class weights, or undersampling.
• Engineer meaningful features (e.g., utilization ratio, delinquency streaks).
• Evaluate models using risk-sensitive metrics such as F1, F2, and ROC-AUC.
• Generate production-ready predictions on an unlabeled validation set.

Dataset Description

• Training Data: ~25,000 records with features like:
  • LIMIT_BAL: Credit limit
  • sex, age, education, marriage: Demographic info
  • PAY_0 to PAY_6: Repayment status history
  • BILL_AMT1 to BILL_AMT6: Monthly billed amounts
  • PAY_AMT1 to PAY_AMT6: Monthly payments
  • next_month_default: Target variable
• Validation Data: ~5,000 records with the same features but no labels.

Key Features Engineered

• AVG_Bill_amt: Average of all billed amounts over 6 months
• PAY_TO_BILL_ratio: Total payment divided by total billed amount
• Delinquency Streak: Consecutive months of overdue payments
• Utilization Ratio: Ratio of total bill amount to credit limit

Modeling Techniques

We compared multiple models to identify the best-performing one:

• Logistic Regression
• Decision Tree Classifier
• XGBoost
• LightGBM

Evaluation Strategy

Since real-world credit risk decisions prioritize recall over precision, metrics were chosen accordingly:

• Primary Metrics: F1 Score, F2 Score, ROC-AUC
• Threshold Tuning: The decision threshold was adjusted to balance business implications of false positives (unnecessary alerts) vs. false negatives (missed defaulters).

Business Implications

Accurate prediction of customer defaults helps in:

• Reducing credit losses
• Designing early-warning systems
• Optimizing credit exposure
• Improving risk-based pricing and customer segmentation

What we achieved:

• Jupyter Notebook with:
  • Data processing
  • Exploratory data analysis (EDA)
  • Financial insights
  • Feature engineering
  • Model training and evaluation
  • Handling Class Imbalance
  • Final predictions

Tools and Libraries Used

• Python
• pandas, numpy
• matplotlib, seaborn
• scikit-learn, imbalanced-learn
• xgboost, lightgbm