The use of log-transformed data has become standard in macroeconomic forecasting with VAR models. However, its appropriateness in the context of out-of-sample forecasts has not yet been exposed to a thorough empirical investigation. With the aim of filling this void, a broad sample of VAR models is employed in a multi-country set up and approximately 42 Mio. pseudo-out-of-sample forecasts of GDP are evaluated. The results show that, on average, the knee-jerk transformation of the data is at best harmless.