The Performance Of Linear Versus Non-Linear Models In Forecasting Returns On The Johannesburg Stock Exchange

Abstract

In this paper we provide a comprehensive comparison of the predictive accuracy of linear and non-linear models when forecasting financial returns, using a number of macroeconomic variables, on the Johannesburg Stock Exchange. We implement a range of linear specifications, Markov switching ARMA and Dynamic Regression models, and univariate models in which the conditional heteroskedasticity is captured by GARCH or EGARCH innovations. Our results indicate that Markov switching models provide the most significant in-sample fit. However, results for the stable portion of the out-of-sample period and the recent recovery period are mixed with both EGARCH-based linear models and 2-state Dynamic Regression models outperforming the alternatives. Over the market crisis period we find that the forecast performance of the nonlinear models is worse than that of the linear models, which suggests that the benefit of the nonlinear treatment of conditional volatility diminishes over this period.