From Crisis to Algorithm: Credit Delinquency Prediction in Peru Under Critical External Factors Using Machine Learning

The integration of external risk factors (EF) has become increasingly vital for robust credit risk assessment, particularly in emerging economies subject to compounded disruptions. This paper presents a comprehensive, scenario-based methodology for incorporating EF—including pandemic variables (COVID-19 mortality and positive cases), climate anomalies (temperature, road blockages), and social unrest—into machine learning (ML) models. We employ an adapted CRISP-DM workflow, testing stationarity (via Dickey-Fuller), causality (via Granger), and model performance (via AUC and ACC) under scenarios with and without government-backed guarantees. Post-hoc explainability techniques (SHAP, LIME) further reveal which features drive delinquency predictions in EF-enriched contexts. Empirical results, derived from over 8.2 million Peruvian credit records (January 2020–September 2023), confirm that EF integration significantly improves predictive accuracy. Time-lagged mortality (COVID MOV) emerges as the strongest single predictor, while combined crises (pandemic, climate, social unrest) amplify delinquency risk. Models such as XGB, CNN, and XNN exhibit the highest adaptability, consistently attaining statistical significance across different partitions of economic activities. Beyond enhancing model accuracy, EF inclusion highlights key drivers—borrower income stability, debt management, EF severity—thereby supporting more transparent, proactive strategies for financial institutions aiming to mitigate credit defaults in volatile, multi-factor environments.