Detecting Anomalies in Foreign Exchange Markets: A Comparison between Structural and Generative Approaches

This paper investigates and compares two distinct approaches to anomaly detection in foreign exchange markets: a structural approach based on correlation networks (Machine Learning approach) and a generative reconstruction-based approach (Generative AI approach), using a twenty-year sample of returns for the EUR-RON, EUR-HUF, EUR-PLN, and EUR-USD exchange rate pairs. The analysis examines whether the two approaches identify the same anomalous periods and what each approach captures. The results show that graph-based methods mainly detect anomalies linked to changes in the structure of currency relationships, even when price movements are not extreme. In contrast, generative models are more sensitive to local shocks in returns but less effective at capturing changes in cross-currency relationships. The paper offers a direct comparison of the two approaches and shows that relational information plays an important role in understanding systemic behavior in foreign exchange markets.