Modeling stock price changes based on microstructural market data

Abstract

In the modern electronic stock exchanges it is possible to analyze event microstructural market data. This data is highly informative and describes physical pricing which allows to find complex patterns in price behaviour. Discovering such patterns manually is hard and time consuming. However, modern machine learning models are capable of solving such issues automatically adapting to changing price behavior. This research presents a trading system based on a machine learning model and market microstructure data. The data for the research has been collected from the Moscow stock exchange MICEX and provides event changes in the exchange book of applications and the tape of all transactions in a liquid stock exchange instrument. Logistic regression model has been used and compared to neural network models with different configuration. According to the study results, the logistic regression model is not inferior in forecasting quality to more complex models and at the same time it has a high forecast formation rate. It is very important when making trading decisions in the modern trading market. The developed trading system has an average frequency of transactions which allows to avoid expensive infrastructure compared to high-frequency exchange trading. However, it also allows to use the full potential of high-quality microstructural market data. The article describes all the stages of a trading system development, including feature engineering, a comparative analysis of price change forecasting models, and a trading algorithm with testing for historical data. It can be used by various investment institutions for the capital effective management in exchange trading. The development of more complex and detailed trading algorithms based on a machine learning model will increase the ultimate efficiency of the entire trading system.