Forecasting Industrial Production Index by its aggregated or disaggregated data? Evidence from one important emerging market

Abstract

Our work aims to address if the use of disaggregate data helps to forecasting industrial production index. We use Brazilian industrial production data and we investigate if disaggregate information improves the accuracy of the forecasts. We use a a number of recent econometric techniques such as the weighted lag adaptative least absolute shrinkage and selection operator (WLadaLASSO) methodology, the exponential smoothing (selecting the most appropriate model) and Autometrics algorithm to model both aggregates and disagregates. As far as we known this is the novelty of the work. We run a a forecasting exercise from one up to 12 months ahead for Brazilian industrial production. Our full sample covers the period from January of 2002 to August of 2017. Our results suggest that modeling disaggregate data better using exponential smoothing model provides the best performance for 1 up to 7 months ahead for Brazilian industrial production using mean square error as a metric and Autometrics algorithm provides better forecast for 8 up to 12 months but it is not clear whether aggregate or disagregate data is the best choice given that they are both part of the final set of good predictions.