This study compares different methods of forecasting price volatility in the crude oil futures market using daily data for the period November 1986 through March 1997. It compares the forward-looking implied volatility measure with two backward-looking time-series measures based on past returns - a simple historical volatility estimator and a set of estimators based on the Generalized Autoregressive Conditional Heteroscedasticity (GARCH) class of models.
Tests for the relative information content of implied volatilities vis-à-vis GARCH time series models are conducted within-sample by estimating nested conditional variance equations with returns information and implied volatilities as explanatory variables. Likelihood ratio tests indicate that both implied volatilities and past returns contribute volatility information. The study also checks for and confirms that the conditional Generalized Error Distribution (GED) better describes fat-tailed returns in the crude oil market as compared to the conditional normal distribution.
Out-of-sample forecasts of volatility using the GARCH GED model, implied volatility, and historical volatility are compared with realized volatility over two-week and four-week horizons to determine forecast accuracy. Forecasts are also evaluated for predictive power by regressing realized volatility on the forecasts. GARCH forecasts, though superior to historical volatility, do not perform as well as implied volatility over the two-week horizon. In the four-week case, historical volatility outperforms both of the other measures. Tests of relative information content show that for both forecast horizons, a combination of implied volatility and historical volatility leaves little information to be added by the GARCH model.
