Volatility Forecasting in Futures Markets : A Statistical and Value-at-Risk Evaluation

Volatility forecasting has paramount importance in position sizing and risk management of CTAs. In this paper we examine the out-of-sample forecasts of widely used volatility estimators for the S&P 500 and the 10-Year US Note from a statistical and Value-at-Risk perspective. Although we do not find evidence for a volatility estimator that is statistically superior, we show that the volatility process of each asset is different with asymmetric GARCH models generating superior forecasts for the S&P 500, whereas symmetric GARCH, the Yang-Zhang estimator along with the implied volatility forecasting better the 10-Year US Note volatility. We also show that the volatility of the 10-Year US Note is more forecastable than that of the S&P 500 producing smaller errors. More importantly, we find that improving the volatility forecast can generate superior VaR estimates that can be accurate under the normal distribution failing only at the lowest quantiles mainly because the distribution is mispecified and badly approximated by the normal. Semi-parametric QML-GARCH models that use the empirical quantiles of the distribution along with GARCH forecasts address that issue and generate superior VaR estimates outperforming all other methods