Forecasting the real prices of crude oil using forecast combinations over time-varying parameter models

In this paper, we forecast real prices of crude oil using real-time forecast combinations over time-varying parameter (TVP) models with single predictor. We reveal the significant predictability at all horizons up to 24 months. The mean squared predictive error reduction over the benchmark of no-change forecast is as high as 17% and the directional accuracy as high as 0.645. A combination with TVP models is found to generate more accurate forecasts than the same combination with constant coefficient models because the forecast errors of individual TVP models are correlated at a lower degree. We also evaluate the forecasting performance in the framework of density forecasting. Our results indicate that the benchmark model can be significantly outperformed by forecast combination at the horizons longer than 3 months.