MRIDUL SINGH

1. Collection Behavior Scorecard

• Developed a mid-term Collection Behavior Scorecard tailored to credit card segment aimed at proactively identifying and mitigating the risk of customers transitioning into the Non-Performing Asset (NPA) category.
• Segmented overall data into different segments on the basis of tenure of accounts and number of accounts of the customer and trained different Xgboost model at customer level for each segment.
• Model was trained by utilizing credit bureau data (Tradeline/Inquiry/Credit vision features), internal statement data, internal transaction data and acquisition data.
• Model was to give 10-15% better KS than client’s multiple short-term models for each delinquency bucket that predict the roll forward into next bucket.

2. Portfolio Model Scorecard

• Developed a Portfolio Scorecard aimed to identify high-risk customers and predict the long-term risk associated with them on giving credit to each card holder on the basis of their credit history so that appropriate portfolio action like limit increase/decrease, extending/revoking card privileges etc.
• Segmented the portfolio into different segments based on customer's payment and transaction behavior and trained different Xgboost model at customer level for each segment.
• Model was trained using more than 2500 custom variables created by utilizing credit bureau data (Tradeline/Inquiry/Credit vision features), internal statement data, internal transaction data and acquisition data to capture applicant's past onus and offus delinquency, payment, and balance patterns.
• Model gave 15-20% higher KS and better capture rate in top 10% population based on model score on all respective segments in comparison to client’s existing model.

3. Recovery Model Scorecard

• Developed a Recovery Scorecard tailored to the credit card segment aimed at identifying customers who have already transitioned into the NPA category and having higher tendency to repay their dues.
• A Xgboost model was trained at customer level using more than 1000 variables to capture high tendency payers from NPA base.
• Utilized credit bureau data (Tradeline/Inquiry/Credit vision features), internal statement data, internal transaction data, and bureau trigger data to create applicant's offus and onus credit profile.
• The model gave more than 125 times discrimination between least and most-likely to payers.

4. Transaction Fraud Scorecard

• Developed a Transaction Fraud Scorecard tailored to credit card segment aimed at identifying fraudulent transactions in real-time while simultaneously decreasing the number of non-fraudulent declines in order to enhance customer satisfaction and minimize losses attributed to fraud.
• A Xgboost model was trained at transaction level using more than 300 complex interaction variables to capture fraud behavior.
• Created Fraud-To-Gross (FTG), Out-of-Pattern (OOP), Positive history, and velocity features at level of Card, Merchant, MCC, Amount of transaction, Age of card holder, Hour of transaction, months on book and interactive variables between two or more level to discriminate between genuine and suspected fraud transactions.
• The model was able to capture 60% fraud in top 1% transactions where the model predicted highest risk.

5. Acquisition Risk Scorecard

• Developed an Acquisition Risk Scorecard to predict credit risk of a new customer at the time of acquisition which will help to take appropriate approve/decline decision on incoming credit card applications.
• Divided the base into different segments based on business requirements and chosen appropriate target.
• Trained and optimized Xgboost model at application level by selecting optimum number of features.
• Utilized credit bureau data (Tradeline/Inquiry/Credit vision features) to capture applicant's past credit behavior.