Agricultural credit models: identifying high risk applications
Loading...
Date
2017
Authors
Bougard, Dominique Alyssa
Journal Title
Journal ISSN
Volume Title
Publisher
University of the Free State
Abstract
The objective of the research was to explore the performance of various statistical credit-scoring models, in order to identify a model that will minimise the misclassification of high-risk applicants, and identify the characteristics that influence repayment ability.
The study was conducted in South Africa, with the use of a case study of a South African financial organisation serving the agricultural sector. The data gathered for this study was gathered through a formal agreement with a commercial financial organisation. Logistic regression (LR), probit analysis (PA) and neural network (NN) were used to construct the credit-scoring models that can be used to classify credit applications in the agricultural sector.
Results of the LR indicate significance at 10% of the following variables, which may have an impact on classification: medium-term loan, credit history, debt to assets (DTA), net farm ratio, diverse 2, high risk, ownership and experience. The PA results demonstrate the following variables at 10% significance: credit history, DTA, net farm ratio, diverse 2, ownership and experience. The identification of characteristics provides confirmation of characteristics that are of importance to credit research. Financial organisations can use the identification of important characteristics as a method to provide guidance to applicants who apply for loans. Doing so will ensure that the organisation will identify characteristics that ensure that the applicant is accepted by the financial organisation. Applicants for loans can ensure that they possess characteristics that correspond to important characteristics identified by the statistical model. The results from the NN are not easily interpretable; due to “black-box” qualities it was not easy to identify the variables that have an influence on the predicted outcome. The NN did, however, outperform the LR and PA in terms of classification accuracy. Neural networks achieved the highest correctly predicted overall accuracy and a lower percentage of Type II error classifications. Logistic regression and PA have overall classification percentages of 96.06% and 3.94% respectively for classifying Type II errors. The NN had an overall classification accuracy of 98.43% and Type II classification error of 1.54%. The main conclusion from this research is that the statistical methods are able to classify credit applications in the agricultural sector and have the ability to improve accuracy in correctly classifying agricultural applicants.
Further research is need to ensure that the correct variables are included in the classification. The classification results of the models are tested and monitored over a period of time to ensure that the accuracy and prediction are acceptable according to the financial organisations. Further research is needed to select the correct variables to be used when supplying credit to smallholder farmers and financial organisations can use the identified important characteristics to provide recommendations and guidance when evaluating applications for loans. Credit applicants can also use these identified important characteristics as a point of reference before applying for the loan at the financial organisation.
Description
Keywords
Credit, Credit Evaluation, Credit Characteristics, Classification Matrix, Logistic Regression, Probit Analysis, Neural Networks, Agricultural credit, Dissertation (M.Agric. (Agricultural Economics))--University of the Free State, [2017?]