SMART RISK, SMARTER DETECTION: A VALUE-AT-RISK APPROACH TO FINANCIAL FRAUD IN IMBALANCED DATASETS

Abstract

As more people utilise online banking services, the large losses that banks and other financial institutions have incurred as a result of new bank account (NBA) fraud are concerning. Machine learning (ML) models have been severely challenged by the intrinsic skewness and rarity of NBA fraud occurrences. This occurs when the number of non-fraud instances is greater than the number of fraud instances, causing the ML models to ignore and mistakenly classify fraud as non-fraud instances. Customers' confidence and trust may be damaged by such mistakes. While addressing the skewness of fraud datasets, previous research has focused on fraud patterns rather than possible losses of NBA fraud risk characteristics. As a risk measure that views fraud cases as the worst-case scenario, the identification of NBA fraud is suggested in this study within the framework of value-at-risk. Value-at-risk models risk characteristics as a skewed tail distribution and estimates possible losses of risk features using historical simulation. ML was used to classify the risk-return characteristics derived from value-at-risk on the bank account fraud (BAF) dataset. In order to provide weight to the skewed NBA fraud cases, the value-at-risk manages the fraud skewness using an adjustable threshold probability range. The effectiveness of the fraud detection algorithm was assessed using a unique detection rate (DT) metric that takes risk fraud characteristics into account. A K-nearest neighbour with a true positive (TP) rate of 0.95 and a DT rate of 0.9406 is used to create an enhanced fraud detection model. Value-at-risk offers a clever way to create data-driven standards for fraud risk management within a banking industry's acceptable loss tolerance.