COMMERCIAL BANK OF ETHIOPIA LOAN APPROVAL PREDICTION USING MACHINE LEARNING TECHNIQUES

MAIN ADVISOR: GADDISA OLANI (Ph.D.) One of the issues that affects the operational process at Commercial Bank of Ethiopia is loan approval process, which might be caused by faulty estimations or a lack of customer information. In order to prevent unanticipated problems, the banks evaluate loan applications through a thorough assessment procedure in an effort to reduce credit risks. Predicting loan approval based on provided and gathered data is therefore crucial in this context. Machine learning is a favourable direction to give accurate and on-time decisions to approve/Reject the Customer’s loan applications. This research investigation aims to investigate the loan approval prediction process with several machine-learning techniques. The proposed technique began with pre-processing the data in order to identify the most notable feature by cleaning the data, eliminating outliers, and determining the correlation between the features. Then, seven machine learning algorithms have been trained and tested: Random Forest classifier (RF), Decision Tree Algorithm (DT), Gaussian Naïve Bayes algorithm, Support Vector Classifier (SVM), Gradient Boosting Classifier, K-Nearest Neighbors (KNN) and Logistic Regression algorithms. The experiment used of around 10,256 rows and 35 columns used to train and test the Machine learning models to predict loan approval. The result of the experimental research mean accuracy are 98% in GaussianNB, 95% in Decision Tree Algorithm, 98% Logistic Regression algorithms,98% in K-Nearest Neighbors,55% in Support Vector Classifier,99% in the Random forest algorithm and 99% in the Gradient Boosting classifier algorithm. It has clearly shown that; the Random Forest and Gradient Boosting algorithm has the highest accuracy over all algorithms applied for loan approval prediction in CBE. This research's novelty can be summed up by contrasting seven machine-learning algorithms to determine which one makes the best predictions. According to the testing data, the Random Forest techniques outperformed the other algorithms in terms of f1-score, accuracy, precision, and recall.