Detecting anomalous payments networks: A dimensionality-reduction approach

Abstract

Abstract Anomaly-detection methods are aimed at identifying observations that deviate manifestly from what is expected. Such methods are usually run on low-dimensional data, such as time series data. However, the increasing importance of high-dimensional payments and exposure data for financial oversight requires methods for detecting anomalous networks. To detect an anomalous network, dimensionality reduction allows measuring of the extent to which the network's main connective features (i.e. the structure) deviate from those regarded as typical. The key to dimensionality-reduction methods is the ability to reconstruct data with an error; this reconstruction error serves as a yardstick for deviation from what is typical. Principal component analysis (PCA) is used as a dimensionality-reduction method, and a clustering algorithm is used to classify reconstruction errors as normal or anomalous. Based on data from Colombia's large-value payments system and a set of synthetic anomalous networks created through simulations of intraday payments, detecting anomalous payments networks is feasible and promising for financial-oversight purposes.