Anchoring Long-term VAR Forecasts Based On Survey Data and State-space Models

The objective of this paper is to forecast Brazilian inflation using a hybrid approach that combines a standard Vector Autoregression (VAR) model with expectations from surveys of consumers or professional forecasters. We cast a VAR model with parameter restriction into a state-space setup, where the long-run forecast from the model matches the long-run survey prediction. The proposed method also allows for exogenous variables in the system of equations as a way to enlarge the information set, and is designed to quickly adapt the multi-step-ahead forecasts in response to new survey information. An empirical exercise with Brazilian data illustrates the usefulness of the method. The results using a pre-COVID-19 sample indicate forecasts obtained from the proposed model prevail over traditional methods at longer horizons, thus confirming the benefits of using forward-looking information from survey in the forecasting process. The main reason is that the method incorporates relevant transformations observed in the Brazilian economy in recent years, such as monetary policy credibility gains and lower inflation targets. In turn, the results based on the full sample, up to August 2022, show larger forecast errors after the pandemic, which caused huge outliers in macroeconomic variables world-wide. Altogether, these findings offer a valuable contribution to applied macroeconomics, especially with regard to forecasting inflation in Brazil using VARs and survey data.