Citation

Ang, John Syin (2022) ESET-ConvNet: Deep learning hybrid model for stock price prediction. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

Limit order book (LOB) tracks the outstanding limit order on a stock or any security. The LOB data is often used for high-frequency trading as it consists of detailed limit order information which cannot be presented from price data. Deep Learning (DL) is widely employed to train high-frequency trading models with LOB data because of its ability to extract features from unstructured data. There is lack of analysis between the existing models of time series and stock price prediction problems. This study has provided comprehensive analysis for time series modelling and stock price prediction problems. Besides, the existing stock prediction studies do not consider latent space of each stock. The aim is to find if learning the latent space for each stock can improve the model performance. This study has proposed a model architecture to combine and extend Temporal Convolutional Network (TCN), Squeeze-and-Excitation Network (SENet), and an embedding layer which is used for learning the latent space. Furthermore, the existing works have trained 5 stock sequences without splitting them to individual sequence as an input. In larger receptive field, this causes serious data leaking issue and this study have performed stock-splitting process to avoid the issue. The proposed model, ESET-ConvNet, has shown the improvement of F1- score. For 3 different prediction horizons, this model achieve the F1 score of 84.60%, 80.04%, and 83.77% respectively. The improvement of F1 score for each prediction horizons are 1.44%, 4.03%, and 4.26% respectively as compared to DeepLOB, the best performer among the LOB modelling studies. In the error analysis, it is realised that the distribution of classes for each stock and for each prediction horizon can be very different. In the ablation study, it is found that the embedding layer does not significantly improve the model performance. To further investigate the embedding layer, the suggestion of future work is to analyse the LOB models using a dataset with better coverage in terms of time span and number of stocks.

Item Type:	Thesis (Masters)
Additional Information:	Call No.: Q325.73 .A54 2022
Uncontrolled Keywords:	Deep learning (Machine learning)
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics
Divisions:	Faculty of Computing and Informatics (FCI)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	11 Jan 2024 01:28
Last Modified:	11 Jan 2024 01:28
URII:	http://shdl.mmu.edu.my/id/eprint/12038

Downloads

Edit (login required)