FORECASTING INDUSTRIAL PRODUCTION WITH LINEAR, NON-LINEAR, AND STRUCTURAL BREAKS MODEL

Boriss SILIVERSTOVS
Dick VAN DIJK.

We compare the forecasting performance of a linear autoregressive (AR) model, a model of structural breaks, a self-exciting threshold autoregressive (SETAR) model, and two Markov-switching autoregressive (MS--AR) models (denoted as MSIAH and MSMH) in terms of point-, interval-, and density forecasts. We use the monthly growth rates of the industrial production indices of the G-7 countries available for the period 1960.1 - 2000.12. We find that the linear AR and non-linear MSMH (popularly referred to as the "Hamilton") model are the best performing ones in terms of the point forecast accuracy, whereas the structural change BP model and the non-linear MSIAH model are the worst performing ones. In addition, the results of the forecast encompassing test suggest that the point forecasts of the AR and MSMH models encompass the forecasts of the other models more often than the opposite occurs. In terms of the interval forecasts, we find that the MSMH model offers the superior performance over the rest of the models including the linear AR model. Finally, upon evaluation of the forecast densities we find evidence that the MSIAH model performs worse than the rest of the models. At the same time we do not find any systematic pattern that would allow us to select the best performing model in terms of the density forecasts.