The U.S. Energy Information Administration regularly publishes short-term forecasts of the price of crude oil. Traditionally, such out-of-sample forecasts have been largely judgmental, making them difficult to replicate and justify, and not particularly successful when compared with naive no-change forecasts, as documented in Alquist, Kilian and Vigfusson (2013). Recently, a number of alternative econometric oil price forecasting models have been introduced in the literature and shown to be more accurate than the no-change forecast of the real price of oil. We investigate the merits of constructing real-time forecast combinations of six such models with weights that reflect the recent forecasting success of each model. Forecast combinations are promising for four reasons. First, even the most accurate forecasting models do not work equally well at all times. Second, some forecasting models work better at short horizons and others at longer horizons. Third, even the forecasting model with the lowest mean-squared prediction error (MSPE) may potentially be improved by incorporating information from other models with higher MSPEs. Fourth, one can think of forecast combinations as providing insurance against possible model misspecification and smooth structural change. We demonstrate that over the past 20 years suitably constructed real-time forecast combinations would have been more accurate than the no-change forecast at every horizon up to two years. Relative to the no-change forecast, forecast combinations reduce the MSPE by up to 18 per cent. They also have statistically significant directional accuracy as high as 77 per cent. We conclude that suitably constructed forecast combinations should replace traditional judgmental forecasts of the price of oil.
Econometric and statistical methods International topics