Applications of machine learning to estimating the multivariate adaptive price impact functions of the Johannesburg Stock Exchange traded financial instruments

Abstract

The research aims to investigate the accuracy of the functional form of market impact as disclosed in the literature. The concept of market impact is necessary for achieving cost-effective portfolio acquisitions or liquidations. The research replicates the literature. The research extracted numerous market features from the trading data with the purpose of establishing areas of exploitation to which machine learning techniques could be applied. The research implemented machine learning to uncover hidden orders. The investigation concludes with a replication of hidden orders inclusive of research. The research discovers that the market impact function in literature is conditionally true. Generalized Linear Models and Support Vector Machines produced low MSE and high R2 each. The MSE and R2 fits of the literature model have improved subsequent to the inclusion of hidden orders. The market impact of both visible and invisible trades closely resembles the model of the literature. Machine learning techniques, in the context of the research, imply that hidden orders execute within similar market conditions