Trending Data and VAR Forecasting

This paper compares the forecast performance of small-scale Bayesian VAR models under various data transformations including level and difference (both with and without structural breaks), the Hodrick-Prescott filter, and linear detrending. The results show that there is no unique data transformation yielding the best forecast in every case, that is, for all variables and at all forecast horizons. Instead, there are rather substantial differences in forecast results across data transformation methods. Some models in detrended data perform reasonably well in several cases. We illustrate that in VAR forecasting, it is a critical consideration for one to use appropriately transformed, or detrended if necessary, data, along with careful model specification. In particular, it is shown that the popular VAR specifications in level or differenced data may be augmented or complemented with alternative VARs in detrended data to improve forecasting