Design, implement, and evaluate a AI-driven credit risk model for SMEs in emerging markets

Thesis Abstract

In the context of rising SME financing gaps in emerging markets and the limitations of traditional credit scoring under constrained data environments, this study addresses the challenge of designing, implementing, and evaluating an AI-driven credit risk model that leverages alternative data sources to enhance predictive accuracy, reduce information asymmetry, and improve lending access for small and medium-sized enterprises. The aim is to develop a robust, scalable credit risk framework that integrates machine learning techniques with established credit risk theory to deliver reliable default predictions and actionable lending decisions. Specific objectives include (1) identifying and curating a multi-source dataset comprising traditional financial metrics, non-traditional indicators (e.g., transaction patterns, social media signals, supplier and payment histories), and macroeconomic factors for SMEs in three representative emerging markets; (2) selecting and comparing supervised learning algorithms (logistic regression, gradient boosting, random forest, and neural networks) with a hybrid ensemble approach to optimize discriminatory power and calibration; (3) incorporating theoretical constructs from the Basel II/III credit risk framework and the Information Asymmetry Theory to define risk drivers and feature engineering strategies; (4) validating model performance using out-of-sample forecasting and conducting stability tests across sectoral subgroups; and (5) assessing operational feasibility, including data governance, fairness, interpretability, and integration pathways for credit decisioning within existing lending platforms. The methodology follows a design, implement, and evaluate paradigm. A mixed-methods research design combines quantitative model development and validation with qualitative insights from lending practitioners. The population comprises SMEs across manufacturing, retail, and services sectors within Nigeria, Kenya, and Indonesia. A stratified random sample of 2,400 SMEs is drawn, with 800 firms per country, and a holdout test set of 20% reserved for out-of-sample evaluation. Data collection employs a triangulated instrument suite (i) financial statement data and credit bureau records; (ii) alternative data streams such as transactional data from payment processors, supply chain information from procurement systems, and digital footprint indicators; (iii) expert interviews with 20 lending officers to capture decisioning practices and risk perceptions. Instrument validity and reliability are established through a pilot study (n=120) and Cronbach’s alpha for composite scales where applicable. Data analysis proceeds in three stages (i) feature engineering and preprocessing including missing-data imputation and normalization; (ii) model development using logistic regression, XGBoost, Random Forest, and Multi-Layer Perceptron, with Bayesian optimization for hyper-parameter tuning, and stacking ensemble to enhance predictive performance; (iii) model evaluation employing AUC-ROC, Brier score for calibration, KS statistic, and Population Stability Index across time and subgroups, complemented by decision-curve analysis to assess net benefit. Interpretability is addressed via SHAP value analysis and partial dependence plots to identify key risk drivers, while fairness checks assess disparate impact across country and sector. The study also performs sensitivity analyses to evaluate robustness to data sparsity and varying payment behavior patterns. Anticipated findings include (a) superior out-of-sample predictive discrimination and calibration for the ensemble AI model relative to traditional logistic models; (b) identification of robust alternative-data features that significantly augment default prediction in data-poor SME contexts; (c) demonstration of model stability across sectors and countries with acceptable fairness profiles; and (d) practical insights into governance, interpretability, and integration requirements for deployment within micro, small, and medium enterprise lending platforms. The study contributes to knowledge by extending credit risk modeling with AI-augmented, multi-source data frameworks tailored to SMEs in emerging markets, providing empirical evidence on the trade-offs between predictive accuracy, interpretability, and fairness, and offering a replicable blueprint for financial institutions to adopt AI-enabled credit scoring in data-constrained environments. The main conclusion posits that an ensemble AI-driven model, grounded in Basel-informed risk principles and bolstered by high-quality alternative data, can materially improve default prediction and lending decisions for SMEs in emerging markets while remaining operationally feasible. Recommendations include establishing data governance and ethical guidelines, investing in data pipelines for continuous feature update, adopting hybrid decision rules that combine model outputs with lender judgment, and conducting ongoing monitoring of model performance to detect drift and ensure equitable access to credit.