Naïve, ARIMA, nonparametric, transfer function and VAR models: A comparison of forecasting performance

We examine the forecasting performance of a number of parametric and nonparametric models based on a training–validation sample approach and the use of rolling short-term forecasts to compute root mean-squared errors. We find that the performance of these models is better than that of the naïve, no-change model. The use of bivariate models (like VAR and transfer functions) provides additional root mean-squared error reductions. In many cases the nonparametric models forecast as well as or better than the parametric models. Our analysis suggests that (a) nonparametric models are attractive complements to parametric univariate models, and (b) simple VAR models should be considered before attempting to fit transfer function models.