Machine Learning for Credit Risk Scoring in Microfinance and Fintech Lending

Background to the Study

Access to affordable credit is widely recognised as a critical enabler of small business growth, household resilience, and broader economic development, particularly in developing economies where a large share of the population and small enterprises operate outside the formal banking system. Microfinance institutions and, more recently, fintech lenders have emerged to serve this underbanked segment, offering small, often short-tenor loans to individuals and micro-enterprises who would not typically qualify for conventional bank credit. A central operational challenge for these lenders is credit risk assessment: determining, at the point of application, the likelihood that a given borrower will repay as agreed, in order to price and approve loans appropriately while managing portfolio-level default risk.

Traditional credit scoring, as practised in established banking systems, relies heavily on extensive formal credit bureau history — a resource that is frequently sparse or entirely absent for microfinance and fintech loan applicants, a condition commonly termed the thin-file or no-file problem. This has motivated growing interest in machine learning approaches capable of incorporating alternative, non-traditional data sources, such as mobile-money transaction patterns, utility bill payment history, and business registration status, alongside whatever conventional financial and demographic information is available, to build a more complete picture of an applicant's creditworthiness than conventional scorecard methods relying solely on sparse formal credit history can achieve. This study contributes to that body of work by designing a machine learning credit risk scoring model that combines conventional loan-application features with alternative behavioural indicators, benchmarking several modelling approaches, incorporating model interpretability through SHAP analysis, and explicitly auditing the resulting model for fairness across applicant subgroups, given the direct financial consequences credit decisions carry for loan applicants.

Statement of the Problem

Many microfinance institutions in developing economies continue to rely on manual, rule-of-thumb credit assessment or simple logistic-regression-based scorecards drawing primarily on conventional financial and demographic features — approaches that may perform poorly for the substantial share of applicants with limited formal financial history. While machine learning approaches, particularly gradient-boosted tree models, have been shown in the literature to improve default prediction accuracy over conventional scorecards, several practical and ethical concerns limit their straightforward adoption in a lending context. More complex models are frequently less interpretable, complicating both regulatory compliance and loan officers' ability to explain a decision to an applicant, and models trained on historical lending data risk perpetuating or amplifying existing biases against particular demographic or business-sector subgroups if not explicitly audited.

This study addresses these concerns directly by pairing a benchmarked comparison of modelling approaches — from interpretable logistic regression through to gradient-boosted trees and a neural network — with an explicit SHAP-based interpretability analysis and a dedicated fairness audit across gender and business-sector subgroups, rather than treating raw predictive accuracy as the sole criterion for model selection. This kind of combined focus on accuracy, interpretability, and fairness together is not consistently represented in the applied credit-scoring literature, which tends to emphasise predictive performance in isolation.

Aim and Objectives

The aim of this study is to design, implement, and evaluate a machine learning model for credit risk scoring suited to microfinance and fintech lending contexts, incorporating both conventional and alternative applicant data, and to assess the resulting model's interpretability and fairness properties. The specific objectives are to:

1. Review existing literature and systems relevant to machine-learning-based credit risk scoring, with particular attention to microfinance and alternative-data contexts.

2. Combine and preprocess a dataset incorporating both conventional loan/demographic features and alternative behavioural features (mobile-money activity, utility payment history, business registration status).

3. Train and compare Logistic Regression, Random Forest, XGBoost, and Artificial Neural Network models for binary default-risk classification.

4. Apply SHAP analysis to identify and interpret the features most influential to individual model predictions.

5. Conduct a fairness audit comparing model error rates across gender and business-sector applicant subgroups.

6. Implement and evaluate a web-based loan-officer decision-support dashboard presenting a risk score, decision recommendation, and explanatory factors for each loan application.

7. Evaluate the overall system through classification performance metrics, interpretability analysis, fairness metrics, response-time testing, and a usability assessment.

Research Questions

1. What machine learning techniques and data sources (conventional and alternative) have been applied to credit risk scoring in existing literature, particularly in microfinance and thin-file contexts?

2. To what extent do alternative behavioural features, such as mobile-money transaction regularity, contribute to default-risk prediction alongside conventional financial features?

3. How do Logistic Regression, Random Forest, XGBoost, and an Artificial Neural Network compare in predicting loan default risk, with particular attention to recall on the default class?

4. What are the features most influential to the trained model's predictions, as identified through SHAP analysis?

5. To what extent does the trained model exhibit disparities in error rates across gender and business-sector applicant subgroups?

Significance of the Study

This study is significant to several stakeholder groups. To microfinance institutions and fintech lenders, it demonstrates a practical approach to incorporating alternative behavioural data alongside conventional features to improve default-risk identification for thin-file applicants, potentially expanding responsible access to credit for underbanked populations. To loan officers, the deployed decision-support dashboard, with its SHAP-derived explanatory factors, offers a more transparent basis for reviewing and communicating a model-assisted lending decision than an opaque risk score alone.

To the academic community, the study contributes a same-protocol benchmark comparison of modelling approaches on a combined conventional-plus-alternative-data credit dataset, together with an explicit fairness audit that is not consistently reported in the applied credit-scoring literature. To regulators and responsible-lending advocates, the fairness audit findings illustrate both the potential and the continued need for vigilance when deploying machine learning models in a domain with direct financial consequences for individuals — a concern shared by financial inclusion research bodies such as CGAP, which studies exactly this intersection of responsible lending and inclusive finance. Researchers building a comparable applied machine learning study may find it worth working through their own methodology with ScholarNest's research coaching support.

Scope of the Study

This study is limited to binary default-risk classification — predicting whether a loan will default or be repaid as agreed — using a combined dataset built from the Statlog (German Credit) benchmark dataset and a Kaggle-sourced microfinance/small-business loan dataset incorporating alternative behavioural features. It covers the complete pipeline from data combination and preprocessing, through model training, benchmarking, interpretability analysis, and fairness auditing, to the implementation of a web-based loan-officer decision-support dashboard and system-level testing. The study does not extend to loan-pricing optimisation, portfolio-level risk management, real-time integration with live mobile-money or credit-bureau data providers, or a full legal or regulatory compliance assessment of the fairness audit findings, though these are identified as directions for further work.

Operational Definition of Terms

Credit Risk Scoring

The process of estimating the likelihood that a loan applicant will default on a loan, typically expressed as a risk score or probability.

Thin-File / No-File Applicant

A loan applicant with limited or no formal credit bureau history, common among microfinance and fintech lending populations.

Alternative Data

Non-traditional data sources used to assess creditworthiness, such as mobile-money transaction history, utility bill payments, and business registration status, as opposed to conventional financial and demographic data.

Default

A loan outcome in which the borrower fails to repay as contractually agreed, operationally defined in this study by the labelling of the source datasets used.

SHAP (SHapley Additive exPlanations)

A model-agnostic technique, grounded in cooperative game theory, for attributing a model's prediction to the contribution of each individual input feature.

Fairness Audit

An evaluation comparing a model's error rates or other performance metrics across defined applicant subgroups, such as gender or business sector, to detect disparate treatment or impact.

Decision Band

A categorisation of a loan application's computed risk score into an actionable recommendation category, such as Approve, Refer for Review, or Decline.

Conclusion

The thin-file problem isn't going away on its own — the borrowers most underserved by conventional credit scoring are exactly the ones microfinance and fintech lenders exist to serve. What this study shows is that gradient-boosted models like XGBoost, fed with both conventional data and alternative signals like mobile-money regularity, can meaningfully improve default detection over the logistic-regression scorecards still common in microfinance practice. But the fairness audit is the part worth sitting with: a 6.1 percentage-point gap in false-positive rate between gender subgroups is not disqualifying, but it is not nothing either — and it's a reminder that better accuracy and fair treatment are two separate things a lending model has to earn, not one thing that follows automatically from the other. Readers researching related applied machine learning or fintech questions can find further comparative material in our computer science project topics library.

Frequently Asked Questions

1. What is the 'thin-file' problem in credit scoring?

It refers to loan applicants who have limited or no formal credit bureau history, a condition common among microfinance and fintech borrowers that makes conventional credit scoring difficult to apply.

2. Which machine learning model performed best for credit risk scoring in this study?

XGBoost achieved the best overall performance, with 88.7% accuracy, 79.4% recall on the default class, and a ROC-AUC of 0.91, outperforming Logistic Regression, Random Forest, and a neural network.

3. Why does recall on the default class matter more than overall accuracy?

Misclassifying a genuinely high-risk borrower as low-risk carries a much higher cost for a lender than the reverse, so recall on the minority default class is weighted more heavily than raw accuracy alone.

4. What alternative data sources were used alongside conventional credit features?

Mobile-money transaction regularity, utility-payment history, and business-registration status were combined with conventional loan and demographic attributes in the dataset.

5. What is SHAP, and why was it used in this study?

SHAP (SHapley Additive exPlanations) is a model-agnostic technique for attributing a model's prediction to each individual input feature, used here to make the model's decisions interpretable to loan officers and researchers alike.

6. Which features were most influential in predicting loan default?

SHAP analysis identified prior repayment delinquency, debt-to-income ratio, and mobile-money transaction regularity as the three most influential predictors of default risk.

7. What did the fairness audit find?

The audit found a modest but non-negligible 6.1 percentage-point disparity in false-positive rate between gender subgroups, flagged as requiring further mitigation rather than treated as a settled result.

8. What is a decision band in a credit scoring dashboard?

It is a categorisation of a computed risk score into an actionable recommendation, such as Approve, Refer for Review, or Decline, presented to loan officers alongside the underlying risk score.

9. Why is model interpretability important in lending decisions?

More complex models are often less interpretable, which complicates both regulatory compliance and a loan officer's ability to explain a decision to an applicant — a gap SHAP analysis is specifically designed to close.

10. Can this kind of model fully replace human loan officers?

No — the study concludes that fairness auditing and human-in-the-loop review remain essential complements to model deployment in a lending context with direct financial consequences for applicants.