Client Name

Swich

Faculty Advisor

Mr. Ammar Habib

SBS Thought Leadership Areas

Investment Decision Making

SBS Thought Leadership Area Justification

We chose Investment Decision Making because both projects focus on how Swich decides who to trust with capital and how to keep the records behind those decisions accurate. The first project is a credit risk engine that turns merchant transaction behaviour into a risk tier and a default probability, leading to a clear lending verdict: approve, conditional, or decline. Using SHAP, we found that the model was penalizing small merchants for their size rather than their behaviour. These micro-businesses are at the heart of Swich's market. We traced the problem, redefined risk based on actual failure, and confirmed the bias was removed. Now, the engine can support a small but reliable merchant and decline a large but risky one. This approach captures the essence of good investment decision-making: predicting risk while making sure each decision is fair and explainable. The second project is an automated reconciliation system that protects the data behind these decisions. It matches encrypted settlement records against Swich's ledger every day, uses a self-clearing queue for late settlements, and keeps a full audit trail. Lending decisions are only as strong as the records behind them, so making sure the books balance supports every decision the platform makes. Together, these projects show how applied AI can improve both the quality of financial decisions and the reliability of the data they rely on.

Aligned SDGs

GOAL 9: Industry, Innovation and Infrastructure

Aligned SDGs Justification

This project supports SDG 9 (Industry, Innovation and Infrastructure) by building essential digital financial infrastructure for a fintech company in an emerging market. The credit risk engine applies machine learning to convert transaction data into a transparent scoring system. The automated reconciliation system uses an encrypted matching engine to maintain accurate records as the platform scales. Together, these systems enhance Swich’s technology, enabling reliable operations, accurate settlements, and growth. This contributes to the resilient, innovative financial infrastructure promoted by SDG 9 and broadens access to financial services in developing markets.

NDA

Yes

Abstract

This Experiential Learning Project produced two data-science systems for Swich, a Pakistani digital payments platform that lets users send money to each other and pay merchants. The first system scores the credit risk of merchants on the platform, and the second automates the daily reconciliation of partner settlements against Swich’s own ledger. Both deal with the same underlying problem: how a fast-growing payments company can decide which merchants to trust, and keep its books accurate, while handling millions of transactions a month. We worked from an anonymized record of 4.7 million transactions made by 527 merchants between July and December 2025. For the credit-risk system, we summarized each merchant’s history into a small set of behavioural measures, combined them into a transparent 0–100 score, and then tested a range of models against that score, from logistic regression and K-means clustering to the gradient-boosting algorithms LightGBM and XGBoost. The models reported accuracy close to 99 percent, which seemed too good to trust. Using SHAP (SHapley Additive exPlanations) to examine how they reached their decisions, we found they were quietly penalising small merchants: low transaction volume was being read as high risk on its own, whatever a merchant’s actual reliability. The cause was the risk label itself, which had been defined partly by transaction frequency and so leaked the answer into the model. We redefined risk purely in terms of transaction reliability and added a volume weighted reliability measure, which removed the size penalty. Judged honestly, without the leak, the model scored a ROC-AUC of about 0.69 to 0.70. The score feeds a decision engine that returns an approve, conditional, or decline verdict with a suggested limit. The second system matches encrypted settlement records against the core ledger every day and uses a carry-forward queue to handle the timing gaps that are normal in settlement, clearing them automatically instead of raising false errors. The practical takeaway is that the three-tier score is ready to guide onboarding, payout limits, and credit eligibility now, while the machine-learning probability is best treated as a directional signal until richer data is available. We recommend adding chargeback, ticket-size, category, and account-age data, rolling out decisions by tier, and making SHAP fairness checks a routine part of monitoring. By letting small but reliable merchants qualify for credit they would otherwise be refused, the work supports financial inclusion and a more transparent approach to lending in Pakistan’s fintech sector.

Keywords: credit-risk modelling, machine learning, explainable AI, algorithmic fairness, financial inclusion, digital payments, reconciliation, fintech, Pakistan.

Document Type

Restricted Access

Document Name for Citation

Experiential Learning Project

Download

Available for download on Saturday, June 07, 2031

Share

COinS