Analyzing anomalies for financial fraud detection: A case study of selected insurance companies listed in Borsa Istanbul

Abstract

This study aims to identify anomalies in the financial data of six leading insurance companies listed on Borsa Istanbul, Türkiye. Traditional anomaly detection methods like GARCH, ARIMA and moving averages have inherent limitations, including the requirement of stationarity, strict distributional assumptions and risks of model mis-specification. To address these issues, we employ four alternative risk measures, i.e., Down-to-Up Volatility (DUV), Negative Conditional Skewness (NCS), Relative Frequency (RF) and the Garman-Klass (GK) on daily stock price data, thereby avoiding stationarity and distribution-related constraints. Our findings reveal significant differences in anomaly detection across these measures. While DUV and RF, which are based on second-moment calculations, capture variations in volatility, the GK approach (computed daily) and the NCS, which considers third-moment characteristics, provide complementary insight. To enhance robustness, we apply both Z-score normalization and Mahalanobis distance for joint anomaly detection. The Z-score method treats all risk measures equally and is suitable for normally distributed data but overlooks potential correlations. In contrast, Mahalanobis distance accounts for multivariate anomalies and interdependencies between risk measures, offering a more holistic approach. Our results indicate that Mahalanobis distance outperforms Z-Score normalization in detecting anomalies in five out of six insurance companies, except in the case of RAYSG. This study underscores the importance of alternative risk measures and multivariate anomaly detection techniques in financial fraud analysis, offering valuable insights for risk management and regulatory practices in emerging financial markets.