The predictability of stock market returns in South Africa: Parametric vs. non-parametric methods

This paper compares the forecasting performance of a sub-class of univariate parametric and non-parametric models in predicting stock market returns in South Africa. To account for conditional heteroskedasticity in stock returns data, the non-parametric model is generated by the conditional heteroskedastic non-linear autoregressive (NAR) model, while the parametric model is produced by the generalised autoregressive conditional heteroskedastic in mean (GARCH-M) model. The results of the paper show that the NAR as a non-parametric model performs better than the GARCH-M model in short-term forecasting horizon, and this indicates the importance of a distribution-free model in predicting stock returns in South Africa.