Deep Learning Based Limit Order Book Modelling and Simulation

Abstract

This thesis is dedicated to utilising state-of-the-art deep learning techniques to facilitate the modelling of the limit order book (LOB), with an ultimate goal to achieve realistic LOB simulation. In financial markets, LOB is a queuing system used to store and sort unexecuted orders submitted by investors, revealing fine-grained contemporaneous demand and supply relation of an asset. LOB has been widely used by researchers and practitioners to gain insights into market dynamics, reflecting its cornerstone role in financial microstructure studies.
The emphasis of the thesis is to conduct realistic LOB simulation, of which the needs originate from practical needs. First, the accessibility of LOB data is very limited, as financial data providers charge a substantial fee for streaming, storage, processing cost associated with the data; and in some scenarios the exchanges do not publish the LOB data at all. Second, traditional LOB simulation models either suffer from subjectivity or lack-of-interaction problems, making the simulated data a weak reflection of the real market. A realistic LOB simulation to be attempted in the thesis can appropriately handle both dilemmas.
Based on a comprehensive overview of the LOB literature, the thesis provides a series of solutions to the aforementioned problems. First, the LOB recreation model is proposed to predict deep level LOB information based on top level trades and quotes data, in which a continuous-time recurrent neural network functions as the main module. While the model allows insights into unrevealed deep level information in the LOB through widely available trades and quotes data, the model fails to simulate a realistic LOB owing to the overlooking of the system’s event-based essence. Second, a state-dependent parallel neural Hawkes process is proposed to model event arrivals in financial exchanges as stochastic point processes. The model is able to provide accurate predictions concerning the arrival of next event. Facilitated by stochastic point sampling algorithms, the model can be further utilised to iteratively sample new events, achieving a minimal simulation of the LOB. Finally, the thesis proposes a brand-new perspective in conducting LOB simulation - a hybrid neural stochastic agent-based simulation. The hybrid simulation paradigm embeds a background trader, whose behaviour logic is learnt from real market data using neural point processes and deep generative models, in a well-developed agent-based LOB simulation platform. The hybrid simulation model shares the advantages of both stochastic and agent-based models, leading to an objective and interactive simulation.
Experiments are rigorously carried out to validate the superiority of the proposed models against existing models from the literature. The effectiveness of each proposed deep learning modules, and factors that can potentially influence the performance of the models, are comprehensively discussed. The authenticity of the simulated LOB, and the simulation system’s reaction to experimental agents, are verified against a wide range of empirical findings from the financial literature. In conclusion, this thesis underscores the transformative potential of deep learning in enhancing our understanding of financial markets, paving the way for more accurate, efficient, and realistic LOB simulations.

Date of Award	7 May 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	John Cartlidge (Supervisor) & Dave Cliff (Supervisor)