Abstract
Forecasting inflation is generally considered a challenging task as forecasters face fundamental uncertainty about the proper selection of variables driving inflation dynamics. In this paper, we investigate the forecasting performance of variables representing economic activity, monetary policy and survey data within VAR and BVAR models.
We propose a scoring algorithm to evaluate their forecasting performance based on various criteria such as the mean square error, the mean absolute error and the Diebold-Mariano test. A one-year horizon is considered for the forecasts and they are constructed by the chain rule using monthly data.
We also determine the forecast accuracy on sub-periods, showing that in a low volatility periods the forecast accuracy can be significantly improved by selecting models using square-root errors. Our results suggest that the survey data have strong predictive power, especially when accompanied by a broad money measure.
The survey data outperform also the indicators of economic activity, probably due to their forward-looking nature. VAR models outperform univariate models in pseudo out-of-sample forecasting, but employing Bayesian restrictions via Minnesota prior did not further improve the forecasting performance.