Critical comparison of statistical and deep learning models applied to the New Zealand Stock Market Index

Abstract

RESEARCH QUESTIONS 1. What are the critical fundamental determinants of the NZX 50 Index movements?

2. How can effective forecasting models based on HWES and ARIMA methodologies be devised and applied with high precision to the NZX 50 Index prediction?

3. How can an efficient univariate LSTM forecasting model and a multivariate LSTM forecasting model be formulated and applied to forecast the NZX 50 Index movement with a high degree of predictive efficacy?

4. Considering all the models tested in different sample periods and scrutiny processes, is it possible to identify a superior forecasting model? Is the recognised model consistently outperforming other tested models in all the testing procedures? Can the redeveloped models efficiently handle the impact of the COVID-19 pandemic?

ABSTRACT Financial markets enable buyers and sellers to trade financial instruments (stocks, bonds, foreign currencies, and derivatives) and improve capital allocation. These markets play a pivotal role in facilitating the interactions between those who seek capital and those who are prepared for capital investments, allowing market participants to transfer risks and stimulate economic growth. Financial time series are inherently dynamic, interdependent, and highly sensitive to many factors. These time series contain deterministic and stochastic characteristics, and many interrelated factors influence them. Accurate predictions of financial time series benefit various market participants to generate wealth through the right trading strategies and other stakeholders to enhance funds. However, due to their inherent complexities, financial time series prediction is considered one of the most challenging problems in data mining.

This thesis employs popular and efficient time series prediction models, reformulates them and implements them to analyse stock market index movements. This scientific exploration uses two widely used classical forecasting techniques [Auto-Regressive Integrated Moving Average (ARIMA) and Holt Winter's Exponential Smoothing (HWES)] and efficient deep learning (DL) [Long Short-Term Memory (LSTM)] network. The predictive precision of the reformulated models will be empirically tested using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE).

Four research questions are meticulously examined to close the identified empirical research gaps in the time series prediction models applied to the New Zealand stock market. Once the redesigned models are sufficiently trained, they are implemented as prediction models on selected stock market indices. Several statistical and econometric tests are executed to substantiate my research findings.