Detecting Anomalous WM/Reuters Fixes Using Trailing Contextual Anomaly Detection

We propose a Trailing Contextual Anomaly Detection (TCAD) model to detect abnormal movements in the WM/Reuters foreign exchange benchmark setting. By leveraging the high correlation levels among currencey pairs, we demonstrate that the TCAD model outperforms ARIMA, Jump Test, and CAD methods in detecting idiosyncratic cross-sectional anomalies. Additionally, we find that adjusting for intraday seasonality enhances the models' ability to predict on market close manipulation. Furthermore, we quantify and identify abnormal fix movements as high-impact events.