High-frequency returns, jumps and the mixture of normals hypothesis

Previous empirical studies find both evidence of jumps in asset prices and that returns standardized by 'realized volatility' are approximately standard normal. These findings appear to be contradictory. Using a sample of high-frequency returns for 20 heavily traded US stocks, we show how microstructure noise distorts the standard deviation and kurtosis of returns normalized using realized variance. When returns are standardized using a recently developed realized kernel estimator, the resulting series is clearly platykurtotic and the standard normal distribution is soundly rejected. Moreover, daily returns standardized using realized bipower variation, an estimator for integrated variance that is robust to the presence of jumps, are more consistent with the standard normal distribution. These results suggest that there is no empirical contradiction: jumps should be included in stock price models.