Discussion Comments on: bLinear models, smooth transition autoregressions, and neural networks for forecasting macroeconomic time series: A re-examinationQ Alfonso Novales * Departamento de Economı ´a Cuantitativa, Universidad Complutense-Madrid, Spain 1. Introduction The paper presented by Professors Terasvirta, van- Dijk, and Medeiros (TvDM) is a very thorough and complete discussion on forecast evaluation of smooth transition autoregression (STAR) and neural network (NN) models using monthly macroeconomic time series data. As mentioned in the introduction to the paper, a fair amount of work has already been done suggesting a nonsignificant gain from using NN mod- els for forecasting, while much less work has been done evaluating the forecasting performance of STAR models. The paper contributes by using an interna- tional macroeconomic data set which has not been examined from the perspective of nonlinear forecast- ing. The structure of the data set is interesting, allow- ing for possible regularity patterns to emerge across variables and countries. Such regularities would be very useful as a guide to detect cases in which non- linear forecasting models should be considered. STAR and NN models are both very flexible para- meterizations that should be able to capture many types of nonlinearities in the data. They are not with- out problems, since they need a very careful specifi- cation regarding the transition rule in the case of STAR models and the number of hidden units in the NN model. Except for their simplest versions, both models contain a large number of parameters, leading to flat likelihood surfaces and a consequent poor identification as an indication of overparameteriza- tion. This could be considered the curse of nonlinear- ity: a simple nonlinear specification may be too close to linearity, but a more interesting model may be hard do identify and estimate with precision. The paper makes an interesting reading for fore- casting practitioners, because the authors discuss a number of very relevant issues on forecasting with nonlinear models and describe the main references on each topic: (i) Do linearity tests provide a reliable guide to post-sample forecast accuracy? Should they guide model specification? (ii) Do we need to use different models for different forecast horizons as is usually done? (iii) How should forecast performance be evaluated? Is the RMSE appropriate for nonlinear forecasting models? (iv) In the presence of model uncertainty, how useful are forecast combinations? Other issues, like: (v) What is the more appropriate data transformation for forecasting? (vi) How should we deal with seasonality? (vii) Should we correct for 0169-2070/$ - see front matter D 2005 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved. doi:10.1016/j.ijforecast.2005.04.002 * Tel.: +34 91 3942594; fax: +34 91 3942613. E-mail address: anovales@ccee.ucm.es. International Journal of Forecasting 21 (2005) 775 – 780 www.elsevier.com/locate/ijforecast