The aim of this talk is to present in short some motivation for studying spot oil price. This price is found to be affected by many factors, so it leads to the problem of model uncertainty in case of forecasting it. Secondly, the potentially important oil price drivers might vary in time.
One of the possible solution to this problem is to use the method of Dynamic Model Averaging (DMA). With the use of 10 explanatory variables some interesting results were obtained.
Moreover, some study with application of Dynamic Model Averaging, Dynamic Model Selection and Median Probability Model was done for 69 commodities. It was found that these Bayesian model combination schemes usually produce more accurate one-step ahead forecasts than some “conventional” forecasting models (like, for example, ARIMA). The benefits of model combination over considering just one model with all potentially important explanatory variables was also found. Moreover, in reasonably high number of cases these results are statistically significance (according to the Diebold-Mariano test).
However, these methods of model combination are based on an assumption that there is one, “true” model in the whole analysed period, but it is “discovered” in a recursive way. On the other hand, the idea of “mixture models” assumes that the “true” model can “really” vary in time (be a “time-varying mixture”). The proposed idea relies on joining the method of “mixture models” with “model averaging” and to apply it to spot prices of the selected energy commodities.