Data-Driven Detection of Online Investment Fraud: A Review of Statistical and Machine Learning Approaches

Abstract

Online investment fraud, in which victims are persuaded through websites, messaging groups or trading applications to invest in schemes that do not exist, has become one of the most damaging forms of cyber-enabled financial crime in India. Unlike card fraud, where the bank sees every transaction, investment fraud is usually carried out by the victim’s own authorised payments, which makes it hard to detect with conventional controls. This review examines the statistical and machine learning approaches that have been used to detect financial fraud and considers how they apply to investment fraud. It covers rule-based systems, statistical models, supervised classifiers such as random forests and gradient boosting, unsupervised anomaly detection, and the analysis of text and payment networks. It then discusses the problems that limit these methods in practice: extreme class imbalance, changing fraud patterns, delayed and incomplete labels, and the difficulty of detecting fraud that the customer has been manipulated into authorising. The literature suggests that no single method is sufficient. Effective detection combines supervised models trained on confirmed cases, anomaly detection for new schemes, network analysis to trace mule accounts, and text analysis of advertisements and messages, all supported by human review. The review closes with priorities for research in India, including shared anonymised datasets drawn from cybercrime complaints and closer study of victims’ behaviour.