Explainable ensemble machine learning for disclosure-informed credit risk assessment in peer-to-business lending

This study develops an explainable ensemble machine learning framework for sustainable credit risk assessment in peer-to-business (P2B) lending, a rapidly expanding FinTech model that enhances access to financing for small and medium-sized enterprises (SMEs). The increasing reliance on algorithmic decision-making underscores the need for transparent and interpretable credit evaluation, particularly in environments characterized by information asymmetry and reliance on borrower-provided disclosures. To address these challenges, a heterogeneous ensemble model is proposed, integrating Random Forest (RF), Light Gradient Boosting Machine (LGBM), and deep learning classifiers within a soft-voting architecture. Feature selection and class balancing are guided by RF importance scores and resampling techniques, resulting in a compact and interpretable 12-feature set comprising pricing, contractual, and transaction-level variables derived from borrower disclosures. Using real-world transaction-level data from a UK platform, the proposed model achieves improved predictive performance (ROC-AUC = 0.767) compared to a neural network baseline (ROC-AUC = 0.717) under severe class imbalance. SHAP-based explainability analysis identifies Maturity Days, Annualised Gross Yield, Advance Rate, and Discount Rate as the most influential predictors, highlighting how disclosed information is translated into pricing and contractual terms in digital lending markets. The findings demonstrate that disclosure-informed features can enhance both predictive accuracy and interpretability, supporting more transparent, robust, and accountable credit risk assessment in FinTech lending environments.