LASSO and Elastic Net on Bank Account Fraud Detection

Abstract

Bank account fraud remains a pervasive threat in the financial industry, resulting in numerous adverse consequences for society. This underscores the critical necessity for the development and implementation of robust and efficient fraud detection mechanisms. The first objective of this study was to compare the performance between LASSO and Elastic Net regression in detecting bank account fraud by using AICc and BIC. The second objective was to choose the best model for bank account fraud detection based on the lowest value of AICc and BIC. Furthermore, this study aimed to identify the factors that affect bank account fraud by using the best model. The dataset used in this study was synthetic data that was obtained from the Kaggle website. The results indicated that factors affecting bank account fraud included income levels, the age of customers, the duration elapsed since the application was initiated, the initial transferred amount during the application, the credit limit proposed by the applicant, disparities between the origin country of the request and the bank's country, applications submitted through application, the duration of a user's session on the banking website, and the operating system of the device used, encompassing Windows, macOS, X11, and other devices that initiated the request. Besides, logistic Elastic Net regression yielded better results than logistic LASSO regression in bank account fraud detection due to its advanced nature, which combined L1 and L2 regularization techniques. This study suggested that logistic Elastic Net regression is useful for enhancing fraud detection in financial institutions and increasing public awareness by identifying the influential factors that affect fraud.