Uso da Metodologia de Combinação de Previsões para Projeções da Arrecadação de Receitas Brutas Primárias de Tributos Federais (2° Lugar Prêmio do Tesouro /2019)

The objective of this study is to analyze the gains in accuracy (precision) when applying the forecast combination methodology to predict the collection of gross revenues from federal taxes (endogenous series) administered by the Federal Revenue Service of Brazil (RFB). The medium and long-term scenarios containing price and quantity data for macroeconomic variables come from the macroeconomic parameter grid produced by the SPE, these variables being considered exogenous series. The realized data (taxes and macroeconomic variables) comprise the period from January 2002 to June 2019. The forecast horizon (out of the sample) to be used for revenue projections of federal taxes considers the medium-term scenario of the same macroeconomic variables contained in the aforementioned parameter grid, but comprising the period from July 2019 to December 2023. As individual forecasting techniques, the following methodological approaches were used: ARIMAX models, TBATS models, neural networks and STLM method. On the other hand, for the combination of forecasts, two forms of combination were considered: combination with weights (arithmetic mean weighted by the square root of the mean square error - REQM) and combination without weights (simple arithmetic mean). The minimax criterion is used to select the models that have the lowest RMSE. The results obtained show that, among the 22 tributes analyzed, the minimax criterion selected 8 tributes according to the combination of models (with weight and without weight) through REQM. In terms of policy implications, the use of the forecast combination methodology in the case of tax collection, considering the medium and long horizonterm, it is an important empirical strategy for decision-making by public managers and policy makers.