Uma análise de nowcasting do crescimento do PIB brasileiro: uma abordagem por meio de métricas de florestas aleatórias

Abstract

The present study aims to conduct nowcasting exercises of Brazilian GDP growth rates. To this end, dimensionality reduction and/or shrinkage techniques are employed in order to select the most relevant variables; once this selection is performed, GDP forecasts are generated using several specifications of a Dynamic Factor Model (DFM), which are compared to an Autoregressive Model (AR). The contribution of this study lies in the use of block permutation importance (BPI) from the Random Forest (RF), implemented with both a Moving Block Bootstrap (MBB) and a Circular Block Bootstrap (CBB), to select the most relevant independent variables. The central hypothesis is that the application of these modern dimensionality reduction techniques may contribute to reducing the Mean Squared Forecast Error (MSFE) and, consequently, improving the predictive accuracy of the DFM. The results suggest that DFMs estimated using variables selected by RF MBB and RF CBB perform well in terms of nowcasting Brazilian GDP growth rates. In contrast, DFMs estimated using all variables (ALL) and those selected by LASSO and ENET exhibit lower predictive accuracy than the AR(1) model in almost all periods. Finally, it is noteworthy that the AR(1) performed considerably well for a simple technique requiring limited data and low computational capacity; in one of the periods, the AR(1) even outperformed the DFM with RF MBB, which was the best-performing specification overall.