(Seasonal) Unit Roots and Forecasting

Abstract We compare three models for forecasting 387 quarterly time series and 393 monthly time series, all with trends. The models are for (i) first differenced data, for (ii) seasonally differenced data and for (iii) periodically differenced data. These models allow for different forms of potentially changing seasonality, depending on the associated unit root restrictions. We document that the seasonal difference filter leads to the worst performing models in terms of forecast accuracy. At the same time, an equally weighted combination of the three model-based forecasts provides the most accurate forecasts.