Forecasting stock returns: Does switching between models help?

Abstract

Research examining the usefulness of non-linear models for stock market returns has almost reached an impasse. While there is general recognition of the superior ability of non-linear models to describe the data, there is less certainty about their ability to forecast the data. As such simple linear models often dominate in forecasting exercises due to their simplicity and any loss with respect to non-linear models is not economically significant. This paper primarily examines not whether a non-linear model can beat a linear model in a straight horse race but whether allowing for the forecast to be given by either the non-linear or linear model depending upon some in-sample criteria provides for improved forecasts. Using a sample of eight international stock markets over the period 1990-2007 our results suggest that on the basis of a trading rule simulation this model switching approach may provide for forecast improvement although general results based on forecast error metrics are mixed. Nonetheless, the results support the view that this model switching approach is certainly an avenue whereby the value of non-linear forecasts can be realised and is worthy of further exploration.