Structured Data Suffices: Machine Learning for SME Credit Risk Assessment in an Emerging Market - Dataset and Analysis Code
Description
Dataset and analysis code for "Structured Data Suffices: Machine Learning for SME Credit Risk Assessment in an Emerging Market", submitted to Financial Innovation (Springer Nature).
The repository contains 1,200 SME loan applications from Bangladesh with structured features (numeric, categorical) and text features (business descriptions, loan narratives). The analysis demonstrates that simple logistic regression on structured features alone achieves 0.900 ROC-AUC in 5-fold cross-validation, with a train–test gap of only 0.017. Text features derived via TF-IDF add negligible predictive value (<1%). SHAP analysis confirms that payment history, loan amount, and cashflow stability are the dominant risk predictors.
Contents:
- bangladesh_sme_1200_records.csv — 1,200 anonymised SME loan records (27 features + binary target)
- bd_sme_1200_complete_analysis.ipynb — Complete reproducible analysis pipeline
- results/ — All figures (ROC curves, SHAP plots, calibration curve) and summary CSVs
Code licence: MIT. Data licence: CC-BY 4.0.
Notes
Files
shoumyac/bd-sme-research-zenodo.zip
Additional details
Related works
- Is supplement to
- Software: https://github.com/shoumyac/bd-sme-research/tree/zenodo (URL)
Software
- Repository URL
- https://github.com/shoumyac/bd-sme-research