Deep Learning for Financial Time Series Forecasting

PhD Thesis Proposal Defence


Title: "Deep Learning for Financial Time Series Forecasting"

by

Mr. Haoren ZHU


Abstract:

Financial time series forecasting plays a crucial role in investment 
decision-making, risk management, and portfolio optimization. In recent years, 
deep learning techniques have garnered significant attention due to their 
ability to extract complex patterns and capture nonlinear dependencies. This 
thesis proposal aims to investigate the application of deep learning models in 
financial time series forecasting, considering the unique properties inherent 
in financial data. Specifically, our work focuses on three key areas. The first 
is on Asset Dependency (AD) forecasting. We introduce an innovative concept 
called the Asset Dependency Matrix (ADM). This matrix organizes pairwise 
dependencies among assets. Our novel approach treats a sequence of ADMs as 
video frames, allowing us to capture both spatial and temporal dependencies. To 
model regime-switching phenomena, we incorporate a Mix-of-Experts (MoE) 
architecture. The second is on volatility modeling. Bridging the gap between 
stochastic and neural network approaches, we establish an equivalence 
relationship between models from the Generalized Autoregressive Conditional 
Heteroskedasticity (GARCH) family and their corresponding neural network 
counterparts. Our methodology seamlessly integrates the GARCH structure as 
components into an established neural network architecture. As a result, we 
effectively incorporate the volatility stylized facts inherent in GARCH models 
into the neural network. The third is on stock movement prediction with 
low-frequency signals. While most trading data are available at fixed 
intervals, macroeconomic data such as CPI and GDP are often released at lower 
frequencies and different time intervals. This data heterogeneity poses 
significant challenges for standard time series forecasting models, which 
typically assume uniform time intervals and sampling frequencies. Our research 
aims to provide insights into the strengths and limitations of deep learning 
models for financial time series forecasting. By analyzing their performance in 
a financial context, we can identify practical applications in real-world 
financial scenarios.


Date:                   Tueday, 4 June 2024

Time:                   3:30pm - 5:30pm

Venue:                  Room 3494
                        Lifts 25/26

Committee Members:      Dr. Wilfred Ng (Supervisor)
                        Prof. Raymond Wong (Chairperson)
                        Prof. Andrew Horner
                        Prof. Ke Yi