Agricultural credit estimation with agricultural and socio-economic indicators: a comparison of machine learning and statistical models

Agricultural production is strategically important in terms of economic growth and rural development. In this study, the performances of machine learning and statistical models were compared to estimate agricultural credits with data covering the years 2000-2022 in Turkey. Variables such as agricultural lands, irrigation rates, freshwater withdrawal, urban population rate and agricultural added value were used as independent variables in the analysis. Gaussian Process Regression, Support Vector Regression, Ensemble Learning, Artificial Neural Networks, Decision Trees and Multiple Linear Regression methods were evaluated in the analysis process. The findings showed that the GPR model provided the highest estimation accuracy in the test phase, and its capacity to model nonlinear complex relationships was superior. While the GPR model was followed by the Ensemble Learning method, the SVR model also exhibited a remarkable performance. However, it was observed that the generalization capacities of ANN and Decision Tree models were limited. In the study, Principal Component Analysis (PCA) was applied to solve the multicollinearity problem. It was determined that the model performances were increased by reducing the data size with the help of principal components. As a result, this thesis has defined machine learning methods as providing higher accuracy in estimating agricultural credits than statistical methods. The findings show that machine learning models can be an effective tool in developing agricultural credit policies and supporting rural development. Using more extensive data sets and hybrid approaches is recommended in the future.