Authors: Kosisochukwu Henry Ukpabi, Abdullahi Mohammed Ibrahim, Zainab Abdulsalam Jamilu
Abstract: Small and medium-sized enterprises (SMEs) operating in perishable food supply chains face intertwined operational and financial risks: inventory can deteriorate before sale, postharvest losses reduce cash available for repayment, and many firms have thin or irregular formal credit histories. This study proposes a leakage-aware, explainable artificial intelligence (AI) framework that integrates food shelf-life, postharvest-loss, commodity-price, and digital banking features to assess credit risk for Nigerian perishable-food SMEs. Because linked Nigerian lender-level repayment data and firm-level shelf-life measurements were not available for this research, the empirical demonstration uses a clearly labelled dataset rather than claiming results from real borrowers. A reproducible simulation of 5,000 SME-loan observations was specified, with 70%/15%/15% chronological train/validation/test partitions, a financial-only baseline, and an integrated model using LightGBM-style gradient boosting. Logistic regression and random forest were included as benchmarks; SHAP-style feature attribution was used for interpretability. The simulated results illustrate the intended evaluation, not observed Nigerian banking performance: the financial-only baseline achieved a test ROC-AUC of 0.742 and PR-AUC of 0.356, while the integrated model achieved ROC-AUC 0.801 and PR-AUC 0.421. At a threshold selected on validation data, recall increased from 0.61 to 0.70, while Brier score improved from 0.176 to 0.158. The largest simulated contributors were recent repayment behaviour, cash-flow volatility, estimated remaining shelf life, postharvest-loss rate, and debt-service coverage. A random split produced more optimistic estimates than chronological evaluation, demonstrating why time-aware validation is essential. The main contributions are (i) a domain-specific framework linking food deterioration to credit risk, (ii) a leakage-control protocol for temporal financial data, (iii) an explainability and calibration evaluation plan, and (iv) a transparent synthetic benchmark that can be replaced by real data. These results are illustrative only; empirical conclusions require validation using consented, anonymised lending outcomes and measured food-quality data.
