Exploring markov chain and random forest to predict loan default in rural banks

Abstract

This study investigates loan default risks in rural banks in Ghana's by developing a hybrid predictive model that combines Markov chains and Random Forest machine learning. The Markov chain analysis examines the transition of loans between risk categories (Current, OLEM, Sub-standard, Doubtful, and Loss) based on Bank of Ghana (BoG) standards, revealing the dynamic nature of credit risk over time. The Random Forest model then classifies loans as default or non-default, identifying key features, with Markov chain transition probabilities serving as a crucial input. Using a dataset of 25,981 loan records from a rural bank, the study finds that a significant portion of borrowers operate in non-traditional sectors, with 14% of loans classified as Losses. The Markov analysis indicates stability in the Current and Loss states, while other risk categories exhibit notable transitions. The Random Forest classifier, utilizing features like Days in Arrears and Interest Charge Type, achieves 91% accuracy and strong F1 scores (0.94 for non-defaults and 0.86 for defaults). The model's Receiver Operating Characteristic and Area Under the Curve (ROC- AUC) score of 0.90 demonstrates its excellent ability to distinguish between defaulting and non-defaulting loans. The model identifies loan duration and arrears as primary predictors of default. Scenario simulations confirm that default risk increases as the loan ages and arrears accumulate. Overall, the hybrid model offers superior predictive accuracy and granular risk insights, enhancing credit risk management for rural banks and supporting financial inclusion in underserved communities.