Predicting Credit Card Fraud using Supervised Machine Learning Methods: Comparative Analysis

Currently, with the progress of technology, people’s and institutions’ range of expenditure channels via digital platforms has expanded. In addition, payment methods have become easier with the digital age. An expenditure, made from even a distant corner of the World, takes place instantaneously through the Internet. Although the rapid and global nature of digitisation contains many advantages, ensuring transaction security can be challenging. In this context, banks have undoubtedly become the most crucial institutions that mediate safe transactions between customers and sellers. In an era where credit card transactions are so prevalent, it is seen as a problem that needs to be solved by banks to determine whether these transactions involve fraud or not, both for their profitability and reputation. It takes a serious effort to determine that credit card expenditures, characterised by dynamic nature, are real expenses of the customer. Therefore, the aim of this study is to propose a model based on supervised machine learning with using real and current data with a few key features. The objective is to reduce banks’ operational burden and cost when identifying credit card fraud. In this context, the credit card transactions of a state-owned bank in January 2023 were considered, using a dataset comprising 13,050 observations. Python programming language is used for model building, and classification algorithms with high discriminatory power, such as Random Forest, Logistic Regression, K-Nearest Neighbours, Decision Trees, and Gradient Boosting, are preferred, which are machine learning techniques. The accuracy scores of the algorithms used in the model setup were determined as follows: Logistic Regression, 92.5%; Decision Tree, 93.1%; K-Nearest Neighbour 86.4%; Random Forest 91.8% and Gradient Boosting 86.9% and performance metrics, such as precision, recall, F1 score, and ROC-AUC, were also examined. Based on their performances, five algorithms were recommended for this study.