This study evaluates the forecasting performance of extreme-value volatility estimators for the equity-based Nifty Index using two-scale realized volatility. This benchmark mitigates the effect of microstructure noise in the realized volatility. Extreme-value estimates with relatively simple forecasting methods provide substantially better short-term and long-term forecasts, compared to historical volatility. The higher efficiency of extreme-value estimators is primarily responsible for this improvement. The extent of possible improvement in forecasts is likely to be economically significant for applications like options pricing. By including extremevalue estimators, the forecasting performance of generalized autoregressive conditional heteroscedasticity (GARCH) can also be improved. © 2007 Wiley Periodicals, Inc. Jrl Fut Mark 27: 1085–1105, 2007