Order flow dynamics for prediction of order cancelation and applications to detect market manipulation

In this work, a methodology is proposed to detect and predict the intention of cancelation of a large order at an optimal event-time horizon by analyzing real order flow market data. We achieve this by reconstructing the full history of the limit order book and formulate the case as a binary classification supervised learning problem. The results presented in this study suggest that using the information at the microstructure level of the order book is highly efficient for predicting and detecting the cancelation of the large order than using information at the macrostructure level, and that predicting the cancelation is marginally outperformed by the detection case. With this, we make a step forward in identifying potential orders related to price manipulation but the results can be used by institutional traders to anticipate adversary market impact produced by large orders.