Conditionally Optimal Weights and Forward-Looking Approaches to Combining Forecasts

As forecasting remains a major topic of interest in economics, combining forecasts has been described extensively in literature. Gibbs and Vasnev (2017) use a forward looking approach to combining forecasts, in which predictions of future forecast errors are made which are used to construct combination weights. In this paper, this method is further explored. Besides unconditional optimal weights and bias-corrected forecasting models, two combination models using conditionally optimal weights are constructed. Specifically, we investigate whether it is beneficial for forecasting to use these conditionally optimal weights. US inflation data is used to examine the forecast performance of the different individual models, bias-corrected models and the forecast combination models. The forecast performances are compared with those of two parsimonious benchmark models, namely the Naive forecasting model as described by Atkeson et al. (2001) and the Equal Weights forecasting model, which both have relatively good forecast performance. In contrast to the promising results in Gibbs and Vasnev (2017), we do not find that the forecast combining models outperform the parsimonious forecasting models.