Machine Learning and High-Frequency Algorithms during Batch Auctions

Abstract

We present the first direct evidence of algorithmic imprints during batch auctions. Order anticipation is an integral part of high-frequency traders' strategies. Hence, some participants may have economic incentive to encrypt noise in the data. We use machine learning to identify five types of algorithmic imprints that hinder the processing of auction information and have the encrypted noise characteristics. Our approach rests on the shifted wavelet tree (Yunyue and Shasha (2003)), a burst detection indicator, and the dynamic time warping similarity measure (Skutkova, Vitek, et al. (2013)). We show that market participants can adapt their trading to the presence of encrypted noise by filtering data in real time, thus clarifying the price discovery process. This could reveal the presence of informed traders. The methodology deployed is adaptable to different environments, including continuous trading.