Pattern Recognition in Microtrading Behaviors Preceding Stock Price Jumps: A Study Based on Mutual Information for Multivariate Time Series

In this study, we propose a new framework to analyze the stock-specific mictrotrading patterns preceding stock price jumps, which should be useful for financial regulation or investment decisions. Using high-frequency trading data, the key step of our framework is to extract a set of core features to distinguish the prejump trading patterns of various stocks taking into account of the temporal information within the feature trajectories. We adopt 10 liquidity measures and 30 technical indicators to generate a high-dimensional candidate feature trajectory set and use a combination of the time-series-based mutual information and the minimum-Redundancy Maximum-Relevancy technique to perform the feature selection. A clustering analysis is then adopted to identify the outlier stocks with idiosyncratic prejump trading patterns. In the end, an application case is conducted based on the level-2 data of 189 constituent stocks of the China Security Index 300 to illustrate the viability of our proposed methodology. Comparison results show that the features we selected has higher capacity to identify the similarity among trading trajectories than those without considering temporal feature information, which provides more reliable features in detecting the outlier trading patterns.