Explainable Alternative Credit Scoring For Thin-File Msmes Via Multi-Source Behavioral Data Fusion And Xgboost

Abstract

The issue of access to formal credits for Micro, Small, and Medium Enterprises (MSMEs) persists as a significant challenge due to the lack of a credit history for these businesses. The traditional method of evaluating such businesses using credit scoring models fails to do so. In this paper, an alternative credit scoring model using Artificial Intelligence and Machine Learning (AI/ML) techniques is proposed, and its evaluation is done. The alternative model uses the Small Business Administration (SBA) National loan dataset (899,164 records), application records (438,557 records), and credit history data (1,048,756 records) to enrich the borrower profile through feature engineering. The baseline model is a Logistic Regression model, which is compared to an advanced XGBoost model that has been trained on the combined data set. For handling severe class imbalances, Synthetic Minority Over-sampling Technique (SMOTE) is used only on the training data. The performance of the proposed model is validated experimentally, showing that it improves performance considerably. The proposed model has an AUC of 0.9646 and an F1 score of 0.9115. It outperforms traditional statistical approaches. The explainability of the proposed model is ensured by SHapley Additive exPlanations (SHAP) which helps to transparently identify critical factors such as delinquency ratio, credit history length, and business stability factors that impact credit risk. Finally, the proposed model is deployed as a REST-based API to score credit risk. The proposed approach shows that by using alternative behavioral data and explainable machine learning, it is possible to improve credit risk detection while promoting financial inclusion for thin-file MSMEs.