Machine Learning Approaches for Detecting and Preventing Financial Fraud: Synthesizing the Evidence

Background: Financial fraud adversely affects the financial capability of especially the most vulnerable. Conventional methods for detecting and preventing fraud often prove insufficient when confronted with intricate fraudulent schemes. In recent years, progress in machine learning (ML) techniques, such as deep learning, big data analytics, data mining, text mining, and artificial intelligence (AI), has offered promising prospects for enhancing the capabilities of fraud detection and prevention. This study provides a review of the academic literature on the application of ML in combating financial fraud. Drawing on a diverse range of scholarly sources, we delve into the evolution of ML techniques in the context of fraud detection, discuss essential concepts and methodologies, and assess their performance in practical settings.

Methods: Undertaking the PRISMA methodology, this research reviewed electronic bibliometric databases, including Web of Science, Scopus, OpenAlex, dimensions, and Lens. The search strategy incorporated a combination of keywords such as "Artificial Intelligence," "big data," and "Machine Learning" alongside "financial fraud," "Ponzi Scheme," or "Wire or money transfer fraud." The inclusion criteria specified studies written in English and published between 2014 and 2024. The OpenAlex database was utilized to generate a dataset of 23,873 articles on financial fraud, which was tokenized for related keywords using text mining techniques. The analysis of the tokenized dataset consisting of 2,639 articles revealed trends and frequencies of the keywords.

Results: We narrowed down our dataset to 2,639 articles after filtering through the 23,873 articles on financial fraud and removing all non-English words. Of these 985 articles, predominantly focus on supervised learning classifiers for credit card fraud detection using real time transactional data. In exploring identity theft detection, 1785 articles primarily focus on recognizing patterns of unusual behavior, triggers for alerts, and investigation techniques. For insurance fraud detection, 149 articles discuss how ML systems typically incorporate historical claims data and policyholder information. Financial fraud detection methods include behavioral profiles that compare user behavior to established norms, and identifying anomalies such as account takeover attempts. Over 2000 articles exploring transaction monitoring systems that apply machine learning to detect suspicious transactions in real time. ML automates risk assessments to ensure compliance with legal obligations. Insufficient protection mechanisms can amplify disparities and perpetuate socio-economic inequalities, especially among low-income households, the elderly, individuals with limited financial literacy, and marginalized communities who are susceptible to financial exploitation.

Conclusion: The findings of the study highlight how machine learning has emerged as a powerful tool that can be leveraged by financial capability researchers and practitioners to detect and prevent financial fraud. Study findings offer implications for using the development of financial capability interventions that are guided by ML. Therefore, social workers could work with financial institutions and policy makers to develop financial capability interventions that leverage machine learning to address current and emerging threats and enhance financial capability.