Description

In this article we analyze bias and autocorrelation properties of inflation forecast errors coming from the surveys carried out by Consensus Economics. We consider monthly forecasts for Chile, México and Brazil as well as quarterly forecasts for the US, Canada, Sweden and Japan. Our sample spans the period from March 2002 to June 2008. We consider forecasts at several predictive horizons. Our results indicate the existence of excessive autocorrelation and bias in forecast errors, which is not consistent with optimality under quadratic loss. We also explore whether these findings of excessive autocorrelation and bias enable us to build new and more accurate inflation forecasts. We evaluate these new forecasts in an out-of sample exercise and find that they are more accurate in the cases of Chile, Mexico, Brazil and the US. For Canada, our results are mixed depending on the forecasting horizon. For Sweden and Japan, however, the new forecasts display lower accuracy. It is worth mentioning that in the best case, our new forecasts display a 45% reduction in the Mean Squared Prediction Error.