Smooth transition autoregressions, neural networks, and linear models in forecasting macroeconomic time series: A re-examination

In this paper we examine the forecast accuracy of four univariate time series models for 47 macroeconomic variables of the G7 economies. The models considered are the linear autoregressive model, the smooth transition autoregressive model, and two neural network models. The two neural network models are different because they are specified using two different techniques. Forecast accuracy is assessed in a number of ways, comprising evaluation of point, interval and density forecasts. The results indicate that the linear autoregressive and the smooth transition autoregressive model have the best overall performance. Positive results for the nonlinear smooth transition autoregressive model may be largely due to the fact that linearity is tested before building any nonlinear model. This implies that a nonlinear model is employed only when there is a need for it, which makes the risk of fitting unidentified models to the data relatively low.