Analysis of multi-market high frequency dynamics and trading strategy

Abstract

With the development of electronic financial markets, the portion of high-frequency trading and high-frequency traders have been increasing rapidly. Traditionally, many studies have been conducted to analyze the market movement and trading strategy in the low-frequency domain. This dissertation seeks to explain the dynamics of market microstructure and analyze the high-frequency trading strategy using machine learning techniques with high-frequency market data from Korean stock spot and futures markets.

Firstly, we train the neural network that predicts the short-term movement of asset prices and use the trained model to analyze the lead-lag relationship between the spot and futures market. Also, the influence of each market feature to the price dynamics is visualized using the gradient calculated from the trained neural network. Secondly, we propose a market model and settings for reinforcement learning to train the high-frequency trading strategy. The proposed scheme is verified for a market environment with the known price dynamics and optimal strategy. Then we train the agent in the high-frequency market environment which is simulated using the historical market data, and compare the trained multi-market trading strategy with two benchmark strategies. Finally, we measure the behavioral characteristics of market participants by estimating the hidden reward function of high-frequency traders via inverse reinforcement learning. We analyze how the behavioral characteristics change with assets with different market conditions and test the sensitivity of the proposed model to model parameters.

The main purpose of this dissertation is to investigate the data-driven approaches for high-frequency market dynamics and trading strategies, using machine learning techniques and high-frequency market data.