Forecasting crude oil real prices with averaging time-varying VAR models

The aim of this research is to discuss the ability to forecast real crude oil price by the use of Time-Varying Vector Autoregression (TVP-VAR) models. In particular, model averaging and model selection schemes over several TVP-VAR models are performed. These methods address the problem of variable uncertainty. Indeed, several previous studies indicate that explanatory variables for crude oil prices can be different in different periods of time. Further, the strength of the relationship between crude oil price and its determinants can vary in time. The applied model combination schemes are an extension of Dynamic Model Averaging, which has already been found viable. Moreover, geopolitical risk is included in each model as an endogenous variable, and the model combination scheme is constructed in a way to tackle the joint forecasting ability (with respect to crude oil real price and geopolitical risk). It is found that, indeed, the Vector Autoregression approach results in more accurate forecasts than single equation a Time-Varying Regression or the standard Dynamic Model Averaging. Also, the model combination scheme of several Vector Autoregression models outperforms a single Vector Autoregression model approach. However, forecast accuracy is tested with some novel tools, such as Giacomini and Rossi fluctuation test and Murphy diagrams, which are able to capture time-varying predictive ability significances and several scoring functions.