Development of an Effective Credit scoring Model Using Artificial Intelligence for Financial Sectors

Abstract

Abstract newlineCredit scoring plays a pivotal role in the financial industry as it enables the management newlineof credit risk and promotes sustainability. Lending institutions, such as newlinebanks, rely on credit scoring models to evaluate the creditworthiness of potential newlinecustomers before approving loan applications. By scrutinizing factors like repayment newlinehistory and financial standing, the bank determines the customer s financial newlinestability through their credit score. Therefore, an accurate and reliable credit scoring newlinemodel is essential in assisting lenders to assess the correlation between customer newlinedata and their capacity to repay loans. newlineThis thesis aims to improve the effectiveness of credit-scoring models by understanding newlinetheir underlying mechanisms. It strive to identify and address the key challenges newlinethat arise while developing these models using conventional machine learning newline(ML) techniques. In this thesis, it focus on three major challenges that researchers newlineface when building ML-based credit scoring models, namely (1) the presence of noisy newlineor outliers, (2) imbalanced class distribution, and (3) high-dimensional credit scoring newlinedatasets. Through the use of proposed ML-based models, a set of methods has newlinebeen designed to overcome each of these challenges and improve the decision-making newlineprocesses for lenders. newlineThe first challenge is addressed by two proposed models, i.e. Adaptive Representative newlineand Density-Based Oversampling Technique (A-RDBOTE) and Multi- newlineObjective Conditional Generative Adversarial Network (MOCGAN). A data point newlineis considered noisy if all its nearest neighbors belong to other classes. These noisy newlinepoints not only degrade the classification performance but also significantly improve newlinethe computational cost of the models. In A-RDBOTE, the reverse k-nearest neighbors newline(RKNN) method is employed to remove these points from the dataset, where a newlinedata point is considered as noisy if its RKNN set is empty. However, in the MOCGAN newlinemethod, a few noise samples may be generated by the trained generator during newlinethe oversampling process. An instance may be suspected as noise if its distribution newlinedeviates from that of the real data. Therefore, the trained discriminator is used newlineii Abstract newlineto identify the dissimilarity among the generated samples. As a result, the trained newlinediscriminator can be used to eliminate the noisy samples from the generated dataset. newlineThe second major challenge in the credit scoring domain is imbalanced class distribution. newlineThis means that the number of applicants with bad credit is significantly newlineless than those with good credit. This inconsistency in sample size can make the newlineclassification models more biased towards the good credit samples, leading to incorrect newlinecreditworthiness evaluation. This thesis proposes three innovative techniques, newlineA-RDBOTE, MOCGAN, and Cluster-Based Oversampling Technique (CUTE), to newlineaddress this imbalanced problem. CUTE uses a cluster-based undersampling approach, newlinewhile A-RDBOTE and MOCGAN use oversampling approaches. newlineThe third challenge in the credit scoring domain is the high-dimensionality, which newlineis commonly known as the curse of dimensionality. The presence of irrelevant and newlineredundant features in the dataset increases the complexity and computational costs newlineof credit scoring models. To tackle this issue, a new approach called the Multiple- newlineOptimized Ensemble Learning (MOEL) method is proposed in this thesis. This newlinemethod eliminates the irrelevant and redundant features from the dataset to overcome newlinethe curse of dimensionality. The thesis demonstrates how innovative artificial newlineintelligence techniques can be used to develop effective credit scoring models for newlinereal-world applications. newlineKeywords: Credit scoring, Imbalanced learning, Ensemble learning, Feature selection, newlineNoise elimination. newline