Aplicação de modelos dinâmicos bayesianos para projeção de arrecadação tributária

In this study, we undertook an econometric analysis aim to forecast the disaggregated series of ICMS administered by Confaz. Three methodologies were applied: i) the dynamic structural model (BSTS); ii) the dynamic linear model (MLD); and iii) the dynamic factorial model (MFD), all of them estimated based on Bayesian practice. The econometric exercises aimed at three types of results: i) the evaluation of the forecast; ii) the elasticity of the tax in relation to the relevant variable; and iii) the projection 60 months ahead outside the sample. Our database is composed of data from January 2006 to December 2019. Considering the difficulty to treat the Confaz series due to the lack of regularity, the exercises done to validate the forecast displayed a quite reasonable performance. From about 20 series for each state, approximately 80% register an MAPE below 15%.