%0 Journal Article %A Richard D.F. Harris %A Jian Shen %T Estimation of VaR with Bias-Corrected Forecasts of Conditional Volatility %D 2004 %R 10.3905/jod.2004.412359 %J The Journal of Derivatives %P 10-20 %V 11 %N 4 %X One of the central real world derivatives issues is how to get the best volatility prediction. A lot of research has focused on evaluating different forecasting approaches, especially to compare methods based on historical returns data, such as RiskMetrics’ exponentially weighted moving average (EWMA) technique, against volatilities implied out from option prices in the market. The standard tool in such analyses is the simple regression of realized volatility on forecasted volatility. If the forecast is unbiased, the regression constant and slope coefficient will be 0 and 1.0, respectively. Many investigators have run this regression and found that none of their techniques passes the rationality test: slope coefficients are almost always significantly less than 1.0. Nevertheless, they end up suggesting that the least bad forecast, typically the option-implied volatility, should be used in preference to the others. This is actually rather odd: Having proven that a given forecast is biased, it clearly can not be rational to use it as one’s volatility prediction without trying to correct the bias. In this article Harris and Shen explore this issue for bonds, equities, and exchange rates in the U.S., U.K., and Japan. They find that, in all cases, the EWMA forecast is significantly biased as a volatility prediction. But they then take the next step and try to correct the bias using the fitted regression coefficients. This significantly improves forecasting performance, particularly for the value at risk application of predicting the 1% and 21/2% tails of the returns distribution. %U https://jod.pm-research.com/content/iijderiv/11/4/10.full.pdf