Determining the effectiveness of microfinance in rural areas of Bangladesh using machine learning

Abstract

Financial exclusion hampers economic growth in rural Bangladesh, as the conventional credit scoring model fails to identify the creditworthiness of the millions of unbanked citizens. Microfinance institutions aim to fill this gap; however, their impact assessments are unable to predict long-term borrower sustainability due to subjective judgment criteria. This thesis proposes a machine learning approach to evaluate the effectiveness of microfinance institutions based on non-traditional socioeconomic and behavioral indicators. A primary dataset of 1,001 rural borrowers with 85 engineered features was created, including digital literacy, household infrastructure, lifestyle factors, and psychometric indicators. Four supervised algorithms, LightGBM, XGBoost, Random Forest, and Support Vector Machine, were applied to predict five indicators of a successful loan: Income Generation, Standard of Living, Business Expansion, Ability to Save, and Asset Acquisition. Results show that the gradient boosting algorithms consistently outperform conventional baselines. Light- GBM and XGBoost go hand in hand in successfully predicting the outcomes of microcredit. The findings demonstrate that alternative behavioral data can proxy formal credit histories. The framework provides MFIs as well as rural borrowers with insights on the possible outcome of their loans.