Client Name

Swich

Faculty Advisor

Dr. Mohsin Zahid Khawaja

SBS Thought Leadership Areas

Other

SBS Thought Leadership Area Justification

  • Credit Risk Assessment
  • It involves an assessment of customer creditworthiness based on transactions. The credit scoring model is used to identify potential lending risks and to aid in making improved credit decisions. The project utilises data analytics and machine learning to offer a practical approach to credit risk assessment and management.

Aligned SDGs

GOAL 8: Decent Work and Economic Growth

Aligned SDGs Justification

This project is related to SDG 8 (Decent Work and Economic Growth) as credit risk assessment is an enabling tool for making responsible and data-informed lending decisions. The credit scoring model is useful for financial institutions to assess the credit worthiness of the customer and also lets them grant financial services to them better and at the same time lower the risk of giving loans. The project helps to enhance the efficiency and effectiveness of credit assessment, thereby promoting sustainable economic growth and financial inclusion.

NDA

Yes

Abstract

This credit scoring model was developed in collaboration with SWICH, a Pakistan-based fintech company that offers a wide range of digital payment and financial service solutions to businesses. The main aim behind this was to establish an automated credit scoring model for risk assessment purposes that could provide accurate ratings of merchant creditworthiness without relying on human judgment. Such a process would help provide reliable credit scores between 300 to 850 along with appropriate recommendations for each merchant using the SWICH platform. A sample size of about 4.7 million transactions from 527 merchant accounts spanning a duration of six months was obtained from both Pay In and Pay Out channels of SWICH. A machine learning pipeline was formulated, which included data preparation, feature extraction, peer group clustering, modeling, score calibration, and deployment phases. Transaction data were then converted into merchant behavior and risk-related features, producing a total of 56 non-leaky features. As part of the modeling process, temporal splitting technique that involves separation of data into training and testing was employed. Training dataset spans the months of July-October 2025, while the testing dataset covers November-December 2025. An XGBoost binary classification model combined with 5-folds stratified cross-validation and Platt probability calibration will serve as our chosen modeling algorithm. Moreover, a Bayesian Beta-Binomial approach will be formulated to produce scores for under-transaction merchants. The final model proved to be an effective predictor with a ROC-AUC score of 0.8975 and correctly identifying 91% of merchants at high risk. It had a Brier score of 0.0644, which shows a well-calibrated probability model and a valid risk score model. Feature importance analysis showed that the most important features for predicting merchant risk included transaction amount, transaction speed, payout pattern, and peer-normalized metrics. The obtained distribution of credit scores helped differentiate between good and risky merchants, with good merchants having credit scores greater than 670, whereas risky merchants had a credit score below 500. Moreover, the Bayesian method helped to assess the risk of 145 unscorable merchants, increasing portfolio coverage from 61% to 100%. It is evident from the project that in the fintech field, machine learning has the potential to be a major asset in enhancing consistency, transparency and scalability of merchant credit decision making in the field. The framework presented here provides a solution to SWICH to allow the automatic approval, review and rejection of applications along with providing an explanation. Recommendations for improvement are to periodically retrain the model every 3 months, track payout merchants, add more merchants to the training set, and consider longer timeframes to observe seasonal trends. The suggested credit score framework lays the groundwork for further credit risk assessment and financial product innovation at SWICH.

Document Type

Restricted Access

Document Name for Citation

Experiential Learning Project

Download

Available for download on Friday, June 11, 2027

Share

COinS