Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
A machine learning hybrid approach to forecasting equity returns volatility: A South African perspective.
For many years, scholars and professionals in the financial markets have been deeply interested in the forecasting of financial market return volatility. There are many methods for predicting the volatility of financial market returns, and various studies have indicated differing degrees of accuracy...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English Eng |
| Published: |
Department of Finance and Tax
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| Summary: | For many years, scholars and professionals in the financial markets have been deeply interested in the forecasting of financial market return volatility. There are many methods for predicting the volatility of financial market returns, and various studies have indicated differing degrees of accuracy in this regard. Research on describing the effectiveness of various approaches under various conditions is still ongoing. This field has moved from simple econometric methodologies like Moving Averages (MA), ARCH-type models and stochastic volatility, to more complex models like LSTM (Long-Short-Term-memory) and SVM (Support Vector Machines) (specifically machine learning algorithms). Machine learning in various forms is currently being explored as an alternative for forecasting the volatility of financial market returns. In this study this exploration is continued by considering a hybrid-based methodology to forecast this volatility, specifically in the South African equities market. There are two guiding principles for this study. The first principle is that specific ARCH-type models that achieve a superior fit to the dataset in question (the JSE All Share Index in this study) can be used in combination with machine learning (ML) models to forecast the volatility in financial market returns. The second principle stems from the work of earlier authors who have demonstrated the suitability and use of LSTM as a ML model that is effective in generating hybrid volatility forecasting models in conjunction with other ARCH-type models. The first guidance is based on the view that accuracy in volatility prediction depends on the ability of a model or group of models to capture volatility stylised facts inherent in a time series dataset used. The approach is based on the idea that various econometric and machine learning forecasting models each have their own advantages and disadvantages, and that combining them results in a stronger forecasting approach. The search for ARCH-type models showing a superior data fit for the Johannesburg Stock Exchange All Share Index (JSE ALSI) revealed the following models as suitable: GARCH(G), EGARCH(E) and TGARCH(T). The study then applies the base econometric models to LSTM and produces seven hybrid models, namely G-LSTM, E-LSTM, T-LSTM, GE-LSTM, GT-LSTM, ET-LSTM and GET-LSTM. Additionally, the averaging of GARCH, EGARCH and TGARCH produces a simple average model. As such, in addition to LSTM and the base econometric models, twelve models are used in this study. This research considered daily prices of the JSE ALSI from January 2004 to December 2022, where 80% of the dataset was used for training purposes and 20% was used for testing purposes. Volatility for this dataset was modeled (both training and testing) using the above models. RMSE, MAE and MAPE were used to evaluate the differential ii out-of-sample performance of the different models. In addition, the Wilcoxon signed-rank test was used to evaluate the significance of the forecasts from the different models generated. The conclusion made is that well-tuned hybrid models outperform all standalone models, including the average model. Furthermore, based on the results of this study it can be argued that GET LSTM, GT-LSTM and ET-LSTM are the most effective financial market returns forecasting models considered, at least as it relates to the South African equities market. Furthermore, more complex hybrid models generally dominate the simpler models as well as the traditional ARCH type models considered. Keywords/ Key phrases: Volatility stylised facts, forecasting volatility, conditional volatility, machine learning, hybrid machine learning, loss functions, LSTM, hybrid volatility forecasting models. |
|---|