TITLE:
Optimizing Bank Loan Approval with Binary Classification Method and Deep Learning Model
AUTHORS:
Abdalla Mahgoub
KEYWORDS:
Bank Loan Prediction Model, Financial Prediction, Machine Learning Algorithms, Logistic Regression Algorithm, Deep learning Algorithms, Advanced Data Science Techniques and Analysis, Deep Learning Activation Functions
JOURNAL NAME:
Open Journal of Business and Management,
Vol.12 No.3,
May
31,
2024
ABSTRACT: For any bank or financial institution, managing loans and controlling leverage is one of the most important tasks they have to undertake. A bank cannot function efficiently without a well-designed loan-to-deposit business model. As technology continues to evolve, the mechanism of handling and granting loans underwent a significant change with the introduction of use cases concerning machine learning and data science. Hence, this data-driven research utilized advanced machine learning techniques to analyze and manipulate the data, aiming to predict the best possible way to recommend a loan to a client. These predictions are based on modified yet unique features created from the data obtained from the client. The dataset was tested using two different methodologies: a logistic regression model and a Neural Network algorithm. Both of these methodologies produced high-level accuracy rates. However, the latter outperformed the currently used methodologies by over 20%, resulting in an accuracy of 90%. The model achieved its optimal accuracy level through the application of advanced deep learning techniques, optimization, and precise selection of the number of hidden layers, as well as the definition of input and output layers. The optimization process for the deep learning model prioritizes three major activation functions: RELU, Softmax, and Sigmoid. In addition to the previous statement, epochs and batch size were also considered based on the dataset size used in this research paper. The successful research results were obtained due to the use of a perfectly balanced, unbiased, and cleaned dataset, as well as the well-executed combination of activation functions for the Neural Network model. A performance assessment was conducted based on a confusion matrix evaluation to demonstrate its feasibility and performance.