On log-transformations, vector autoregressions and empirical evidence

Abstract

This paper re-analyses a bivariate US macroeconomic time series, previously used to demonstrate the need for bias corrections in the forecasts of levels of variables, modelled and originally forecast using a V AR after log-transformation. It is demonstrated that claims previously made for these data, concerning improvements in forecast accuracy following bias correction, were not well founded. Simple univariate forecasting procedures are shown to be more successful for these data than a cointegrated V AR, with or without bias correction. In the light of previous empirical work, such findings could have been expected. This further reinforces the call for an answer to why well-motivated theoretical advances in time series analysis often do not lead to noticeable improvements in out-of-sample forecast accuracy.