Deep Learning in Financial Time Series Forecasting

Zhou, Junxian

Deep Learning in Financial Time Series Forecasting

Zhou, Junxian

Permalink

Publication Type:: Thesis
Issue Date:: 2025

Open Access

Copyright Clearance Process

Recently Added
In Progress
Open Access

This item is open access.

Adobe PDF

Download thesisAdobe PDF (3.47 MB)

View statistics

Full metadata record

Field	Value	Language
dc.contributor.author	Zhou, Junxian
dc.date.accessioned	2025-11-10T00:36:42Z
dc.date.available	2025-11-10T00:36:42Z
dc.date.issued	2025
dc.identifier.uri	http://hdl.handle.net/10453/190638
dc.description	University of Technology Sydney. Faculty of Engineering and Information Technology.	en_US.UTF-8
dc.description.abstract	Financial Multivariate Time Series (Fin-MTS) forecasting is a cornerstone of modern financial analytics, underpinning prudent decision-making and risk management in turbulent markets. Compared with other time-series domains, Fin-MTS exhibit five distinctive features: (i) pronounced non-linearity, (ii) regime-switching volatility, (iii) latent periodic structures such as credit or policy cycles, (iv) complex intra- and inter-series dependencies across heterogeneous assets, and (v) continual interaction with unstructured information (e.g., news and analyst sentiment). These intertwined characteristics complicate modelling and demand methods that capture both quantitative dynamics and qualitative context. Despite significant advancements in forecasting methods, existing state-of-the-art models often fall short of addressing the intricacies of Fin-MTS. Specifically, these models face several critical challenges: (C1) How can models uncover hidden periodic structures? (C2) How can intra-series and inter-series dependencies be modelled simultaneously? (C3) How can external, unstructured financial information be effectively integrated? (C4) How can scalability and interpretability be ensured? This thesis introduces two innovative models to address these challenges: the Fourier Graph Convolution Transformer (FreTransformer) and the Fine-Tuned Large Language Model for Financial Time Series Forecasting with FTS-Text Embeddings. The FreTransformer utilises frequency-domain transformations to expose hidden periodicities and employs a novel Fourier Graph Convolution Network to capture intra-series and inter-series dependencies in a unified framework effectively. Complementing this, the Fine-Tuned Large Language Model leverages pre-trained large language models to align structured time-series data with unstructured textual information through the FTS-Text Embedder, while the FinLoss optimisation function enhances core financial metrics. Together, these innovative works improve the accuracy of Fin-MTS forecasting by providing scalable, interpretable, and contextually enriched solutions, establishing a robust foundation for future advancements in financial analytics.	en_US.UTF-8
dc.format	Thesis (MAnalytics)
dc.language.iso	en_US	en_US.UTF-8
dc.relation	https://opus.lib.uts.edu.au/bitstream/10453/190638/1/thesis.pdf
dc.rights	info:eu-repo/semantics/openAccess
dc.rights	The author owns the copyright in this thesis including all reproduction and reuse rights for the work. The work may not be altered without the permission of the copyright owner. Attribution is essential when quoting or paraphrasing from this thesis.
dc.rights	© 2025 Junxian Zhou
dc.rights	au.edu.uts.lib/cph
dc.title	Deep Learning in Financial Time Series Forecasting	en_US.UTF-8
dc.type	Thesis
utslib.copyright.status	open_access	*

Abstract:

Financial Multivariate Time Series (Fin-MTS) forecasting is a cornerstone of modern financial analytics, underpinning prudent decision-making and risk management in turbulent markets. Compared with other time-series domains, Fin-MTS exhibit five distinctive features: (i) pronounced non-linearity, (ii) regime-switching volatility, (iii) latent periodic structures such as credit or policy cycles, (iv) complex intra- and inter-series dependencies across heterogeneous assets, and (v) continual interaction with unstructured information (e.g., news and analyst sentiment). These intertwined characteristics complicate modelling and demand methods that capture both quantitative dynamics and qualitative context. Despite significant advancements in forecasting methods, existing state-of-the-art models often fall short of addressing the intricacies of Fin-MTS. Specifically, these models face several critical challenges: (C1) How can models uncover hidden periodic structures? (C2) How can intra-series and inter-series dependencies be modelled simultaneously? (C3) How can external, unstructured financial information be effectively integrated? (C4) How can scalability and interpretability be ensured? This thesis introduces two innovative models to address these challenges: the Fourier Graph Convolution Transformer (FreTransformer) and the Fine-Tuned Large Language Model for Financial Time Series Forecasting with FTS-Text Embeddings. The FreTransformer utilises frequency-domain transformations to expose hidden periodicities and employs a novel Fourier Graph Convolution Network to capture intra-series and inter-series dependencies in a unified framework effectively. Complementing this, the Fine-Tuned Large Language Model leverages pre-trained large language models to align structured time-series data with unstructured textual information through the FTS-Text Embedder, while the FinLoss optimisation function enhances core financial metrics. Together, these innovative works improve the accuracy of Fin-MTS forecasting by providing scalable, interpretable, and contextually enriched solutions, establishing a robust foundation for future advancements in financial analytics.

Please use this identifier to cite or link to this item:

http://hdl.handle.net/10453/190638