Limit Order Books Anomaly Detection With Transformer-Based Autoencoder And Trade Manipulation Simulator
DOI:
https://doi.org/10.64252/sqd56d81Keywords:
LOB, Anomaly Detection, Transformer Autoencoder, Temporal Representations, Fraud Detection, Dissimilarity Function, Trade-Based Manipulation, Quote Stuffing, Layering, Pump-and-Dump, Simulation Framework, Deep Learning, Out-of-Sample Detection, NASDAQ Stocks, asset-agnostic, Detection Accuracy, Synthetic Data, State of the ArtAbstract
In this study we introduce the first hybrid anomaly detection framework for LOB data which harnesses the recent development in deep learning. At the core of the framework is a modified Transformer-based autoencoder that provides rich temporal representations of LOB subsequences while improving the distinguishing of normal from anomalous trading activity. In the learned representation space, a new dissimilarity function is learnt to capture normal LOB dynamics and permits out-of-sample detection of anomalous activities. To test the hypothesis, we have provided a trade manipulation simulation pipeline which can create synthetic trades like quote stuffing, layering, and pump-and-dump schemes that are based on actual frauds experienced in financial markets. The experimentations done on five NASDAQ stocks LOB datasets show that our technique yields detection accuracy greater than 97% compared to the available state of the art algorithms, without relying on previous manipulation knowledge or specific stock characteristics.
