Automated model selection with applications to Brazilian industrial production index

Abstract

Brazilian Industrial Production Index undergoes different methodological updates and periods of high inflation over time, which prompts researchers to avoid using too long industrial production series. We analyze how performance of different models in forecasting the Brazilian IndustrialProduction Index one-step ahead is influenced by the use of samples of different lengths. Relative performance of these models is also assessed. Results show that most models benefitfrom expanding the estimation sample beginning at least up to 1993:12. Autometrics lag selection with impulse dummy saturation forecasting performance is improved almost monotonically with sample size. Forestimation starting inJanuary 1975 and 1985, predictions fromAutometricswith impulse dummy saturation and the average of forecasts are statistically more accurate than those from the benchmark AR model. However, the average of predictions performs better in the first half of the forecast horizon and Autometrics performs better in the second half.