The impact of market structure on price determination : a simulation approach using multi-agent reinforcement learning in continuous state and action space

Abstract

This thesis proposes a simulation tool to study the question of how market structure and market players' behavior affect price movements. The adaptive market simulation system consists of multiple agents and a centralized exchange. By applying reinforcement learning techniques, agents evolve and become capable of making intelligent trading decisions while adapting to changing market conditions. Trading dynamics in the real world are complex yet compelling. The presence of the human element in trading makes studying it via repeatable scientific models, especially on a large scale, very difficult and almost unfeasible. By making it possible to conduct controlled experiments under various market scenarios, this simulation seeks to help researchers gain a better understanding of how different types of traders affect price formation under distinct market scenarios. The impact of trading frequency on prices is also explored as a test of the simulation tool. Results suggest that the market generates richer information when the frequency of trading is high, and when the market is more frequently accessed, short-term market prices demonstrate higher volatilities and move faster in respond to market sentiments.