We propose a reliable new measure for realized skewness, which is robust to microstructure noise at an ultra-high frequency level. Asymptotic theory for the new estimator has been derived and simulation studies verify its superior performance. We apply the new estimator to tick data of the S&P 500 index to forecast the equity premium in the U.S. market from 1990 to 2011 and find that it has significant forecastability, both in sample and out-of-sample. We also show that the new skewness measure plus the variance risk premium provides the right decomposition for the skewness risk. We thus provide evidence that the realized skewness is linked to the conditional market premium.