Anomaly Detection in Financial Data: Comparative Analysis of Statistical Methods and Machine Learning in Simulation Studies

Abstract

The increasing prevalence of financial fraud in the digital era demands the development of reliable and efficient detection methods. This study aims to conduct a systematic comparative analysis of the effectiveness and consistency of seven anomaly detection methods on simulated transactional financial data. The methods tested cover a broad spectrum, ranging from Benford's Law, robust statistical methods (based on MAD), supervised machine learning (Logistic Regression and Random Forest), to unsupervised machine learning (K-Means Clustering and Isolation Forest). Using a simulation study based on R software, a dataset with 20,000 transactions was generated, 5% of which were manipulated as fraud with clear scenarios. The analysis results show that almost all methods successfully detected anomalies with a high success rate. Supervised models, such as Binary Logistic Regression, showed near-perfect performance, while unsupervised methods such as K-Means and Robust Distance (MAD) also showed recall rates above 95%. Although all methods performed well, there were differences in the trade-off between recall and false positives, underscoring the importance of choosing a method that suits business objectives. This study concludes that a layered approach that combines statistical screening with advanced machine learning models is the most comprehensive approach to fraud mitigation