A comparative analysis of non-linear techniques in South African stock selection

Forecasting stock performance has long been one of the primary objectives of financial practitioners. Literature has shown that the classical linear approach to modelling the interactions among company-specific factors and its stock market re- turns in time have become less suited for capturing the movements of the stock market. Hence, attempts to predict the performance of a stock have become associated with additional layers of complexity. This has led to the adoption of non-linear approaches to forecast stock performance. This dissertation explores the performance of some non-linear models in the South African market. These were classification and regression trees (CART), logistic regression and a random forest approach com- pared against a linear regression model. Moreover, a hybrid model between CART and logistic regression was considered. The models fell into two categories (i.e., static and dynamic models). Using a set of classification and portfolio performance metrics it was found that that a dynamic modelling approach outperformed a static approach. Overall, the logistic and linear regression models dominated in terms of performance against the tree-based models and hybrid approaches. The results also demonstrated that a hybrid approach offered an improvement over a stand-alone CART.

Grantor dc:publisher.institution

Division of Actuarial Science

Year dc:date.issued

2015

Author dc:creator

• Hutheram, Nikhil Arnaidas

Advisor dc:contributor.advisor

• Bosman, Petrus

Language dc:language.iso

eng

Handle dc:identifier.uri

http://hdl.handle.net/11427/15732

OAI identifier oai:identifier

oai:open.uct.ac.za:11427/15732