Highly scalable and user-friendly Credit Decisioning Modelling with focus on retail credit card portfolio for a mid-sized US based Retail Bank
Client : Mid-Sized US Retail Bank
Objective
To Create a Highly Scalable and User-Friendly Credit Decisioning Model for a US Retail Bank’s Credit Card Portfolio
CRISIL's Solution
- Data structuring and converting multi-class labels in target variable to binary completed in a single click
- Pre-processing used mean, median, and mode, along with imputing with KNN algorithm
- Exhaustive EDA carried out quickly by employing easy-to-use interface
- Data imbalance issues addressed with advanced SMOTE analysis and undersampling/oversampling techniques
- Feature selection and performance measured with algorithms (such as genetic algorithms), principal component analysis, weight of evidence, etc.
- Advanced machine learning and deep learning algorithms used for comprehensive model building exercise
- Deep learning, coupled with multiple optimization techniques, used to improve model performance
- All advanced ML techniques integrated with customized model tuning parameter
- Advanced feature selection, along with automated fine and course classing
- Key concerns such as model interpretability of machine learning algorithms addressed with the help of algorithms such as LIME, ICE, and SHAP
- Customizable scorecard generated
- Exhaustive performance analysis performed using tests such as ROC, KS, GINI, Precision, and Recall
- GridSeachCV gave the best performing algorithm
Client Impact
- New model enhanced performance efficiency, reducing analysis cycle time and minimizing error rate
- Results accessible to client through customizable scorecard