PERFORMANCE EVALUATION OF MACHINE LEARNING ALGORITHMS IN PREDICTING AGRICULTURAL LOAN DEFAULTERS

Abstract

Financial institutions in Nigeria have continuously extended loan offers to a sector of the economy, say the manufacturing and industrial sector, compared to other sectors, like the agricultural sector, due to the peculiarity of each. To aid such underserved sectors, such as the agriculture sector, the Nigerian government has established risksharing interventions in the agricultural sector such as Nigeria Incentive-Based Risk Sharing System for Agricultural Lending (NIRSAL)to encourage the financial institutions towards the issuance of credits to farmers. Although the risk-sharing incentives has improvedthe issuance of loans to farmers, financial institutions still seek to reduce the leftover risk. Therefore, this research utilized a private agricultural loan dataset collected in Lavun Local Government Area of Niger state, Nigeria to predict the likelihood of agricultural loan default of farmers in Lavun Local Government Area. Dataset dimensionality reduction of Recursive Feature Elimination with Crossvalidation (RFECV) and Principal Component Analysis (PCA) was appliedto the dataset to improve performance metrics. RFECV was used to reduce the features of the dataset from 60 to 44 while PCA extracted the dataset features into 31 principal components. Furthermore, machine learning algorithms of random forest, logistic regression, support vector machine, gradient boosting, and adaptive boosting were applied to the dataset. The results obtained shows that gradient boosting and random forest algorithms were the most effective when the RFECV dimension reduction technique was applied to the dataset in predicting agricultural loan defaults with precision and f1-score of 86.36% with 90.48% and 89.47% with 82.93% respectively. When PCA was applied to the dataset, logistic regression and ada boost achieved results of 78.95% and 74.35% respectively for precision and 76.92% and 74.29% respectively for f1-score. Overall, logistic regression proved to be the most consistent machine learning classifier when either PCA or RFECV is applied to the dataset while gradient boosting proved to be the best algorithm in predicting agricultural loan defaulters. The reduction of accuracy variation observed during cross-validation of the best models is proposed for further study.