Outlier Detection in TARGET2 Risk Indicators

Abstract

This paper studies the detection of outliers in risk indicators based on large value payment system transaction data. The ten risk indicators are daily time series measuring various risks in the large value payment system, such as operational risk, concentration risk and liquidity flows related to other financial market infrastructures. We use extreme value theory and local outlier factor methods to identify anomalous data points (outliers). In a univariate setup, the extreme value analysis quantifies the unusualness of each outlier. In a multivariate setup, the local outlier factor method identifies outliers by measuring the local deviation of a given data point with respect to its neighbours. We find that most detected outliers are at the beginning and near end of the calendar month when turnover is significantly larger than at other days. Our method can be used e.g. by overseers and financial stability experts who wish to look at many (risk) indicators in relation to each other.