Forecasting stock returns: A comparison of five models

Ramuada, Vhahangwele Cedrick (2018-12)

Thesis (MSc)--Stellenbosch University, 2018.

Thesis

ENGLISH ABSTRACT : Forecasting the movement of stock returns prices has been of interest to researches for many decades. Due to the complex and chaotic nature of the stock market, it has been difficult for researches to find a model which can be used to accurately predict the movement of stock returns prices. Many statistical models have been proposed for forecasting the direction of movement of stock returns prices. The objective of this study was to use ARMA type models and an Artificial Intelligence Neural Network model to predict the direction of movement of stock returns prices of four JSE listed companies, namely, Netcare Group Ltd, Santam Ltd, Sanlam Group Ltd, and Nedbank Group. The models were assessed in terms of their ability to predict whether the next day’s returns price will go down or up. Four ARMA-type models, namely, ARMA-Maximum Likelihood, ARMAState Space, ARMA-Metropolis Hastings, AR(3)-AVGARCH(1,1)-Student-t model and an Artificial Neural Network (ANN) model were implemented to try to predict the direction of movement of stock returns prices. Historical (past) stock returns prices were used to make inference about future directional movement of stock returns prices. Empirical results show that the ARMA-Maximum Likelihood, ARMA-State Space, AR(3)-AVGARCH(1,1)- Student-t model, and Artificial Neural Network (ANN) models have a strong ability to predict whether the next day’s returns price will go down or up with acceptable accuracy. However, the ARMA-Metropolis Hastings model performed very poorly, its highest accuracy was a mere 68%. Overall, empirical results show that the Artificial Neural Network model was superior or outperformed all the ARMA-type models, the highest accuracy achieved by the model was 89%. The results of the Superior Ability Test also showed that the ANN model was indeed superior to the Box-Jenkins ARMA type models in at least 5 cases.

AFRIKAANSE OPSOMMING : Die voorspelling van die beweging van voorraad opbrengs pryse, is van groot belang vir navorsing vir dekades. As gevolg van die komplekse en chaotiese natuur van die aandele mark, dit mooilik vir navorsers om ´ n model te vind wat gebruik kan word om akkurate voorspelling van die beweging van die voorraad opbrengs pryse te maak. Verskeie statistiese modelle is voorgestel om rigting van beweging te voorspel van die aandele opbrengs prys. Die doel van hierdie studie was om die ARMA- tipe model en ´ n “kunsmatige intelligensie neurale netwerk" (Artificial Intelligence Neural Network) model te gebruik om die rigting van beweging van aandele obrengs prys van vier JSE genoteerde maatskappye te voorspel; naamlik, Netcare Group Ltd, Santam Ltd, Sanlam Group Ltd, and Nedbank Group. Die modelle is beoordeel in terme van hul vermoë om te voorspel of die volgende dag se pryse sal op of afwaarts gaan. Vier ARMA-tipe modelle, naamlik ARMA-Maksimum Waarskynlikheid, ARMAStaat Ruimte, ARMA- Metropolis Hastings, AR(3)-AVGARCH(1,1)-Studentt modelle en ´ n Kunsmatige Neurale Network (Artificial Neural Network : ANN) model is geimplementeer om die bewegingsrigting van aandele opbrengs pryse te voorspel. Historiese aandele pryse is gebruik om afleidings te maak oor toekomstige rigtingbewegings van aandele pryse. Gebaseer op ondervinding die resulte bewys dat die ARMA-Maksimum Waarskynlikheid, ARMA-Staat Rruimte, AR(3)-AVGARCH(1,1)-Student-t Modelle en Kunsmatige Neutral Netwerk (ANN) modelle ´ n sterk vermöe het, om die volgende dag se obrengs pryse af of hoër te voorspel met aanvaarbare akkuraatheid. Nietemin, die ARMA-Metropolis Hastings modelle het baie swak gevaar , die hoogste akkuraatheid was ´ n blote 68%. In die algemeen, gebaseer op ondervinding die resultate wys dat die ANN model beter was en die ARMA-tipe modelle geklop het, die hoogste akkuraatheid behaal van die model was 89%. Die resultate van die Superior Ability Test het aangetoon dat die ANN model beter was as die Box-Jenkins ARMAtipe modelle in ten minste 5 gevalle.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/104923
This item appears in the following collections: