Oil price volatility forecasting is of major importance due to the financialisation of the oil market and the fact that the oil market participants’ decisions are based on such forecasts (e.g. oil-intensive industries, policy makers, portfolio traders).
Currently, forecasters mainly predict oil price conditional and realized volatility using primarily Generalised Autoregressive Conditional Heteroscedasticity (GARCH) and Heterogeneous Autoregressive (HAR) models and evaluate the forecasts’ performance using statistical loss functions, such as the Mean Absolute Predictive Error (MAPE).
Nevertheless, oil price volatility users are faced with (i) multiple volatility measures apart from conditional and realized (e.g. historical, implied, range-based, bipower, semi-variance, two-scale realized), (ii) multiple forecasting models (HAR, GARCH, ARIMA, Switch-Regimes, Weighted Moving Average (WMA)) and (iii) different applications for which they use oil price volatility forecasts (e.g. policy making, portfolio allocation, risk management).
Hence, the evaluation of the different forecasts using statistical loss functions is not adequate.