Previsão da receita tributária do Brasil com modelo dinâmico de alta dimensão

Abstract

The Executive Branch must prepare the Annual Budget Law in the fiscal year prior to its implementation. Given this particularity, it is necessary to set expenditures and forecast revenues. From this perspective, forecasting becomes essential for the public budget to be executed as planned. The greater the discrepancy between the forecast and the actual revenue collected, the more the public entity must adjust the budget—for example, through spending cuts, changes in tax rates, or an increase in available funds, such as through the contracting of credit operations. The latter directly impacts the country's fiscal situation, as it can lead to public indebtedness and fiscal deficit. Therefore, accurate forecasting can help prevent the failure to implement pre-established actions and programs, as well as a potential increase in public debt. In this context, the largest share of the Federal Executive's revenue comes from tax revenues— the focus of this study—which are managed by the Federal Revenue Service of Brazil. Several authors show that factor models, by incorporating a large amount of information (variables) without compromising the parsimony of the models used, can produce more efficient forecasts. This study compares the forecasts generated by factoraugmented vector autoregressive (FAVAR) models, vector autoregressive (VAR) models, and autoregressive moving average (ARMA) models. The factor models use a dataset with 123 monthly-frequency variables, covering the period from January 2002 to June 2024. The results show that the VAR models produce more accurate forecasts than the ARMA model, and that the FAVAR model can improve upon the forecasts of the VAR model.