Abstract: The analysis of high-frequency data is diffucult, because of the white noise. In this paper, the estimate of RV on financial assets is studied, especially when high frequency asset prices are available. The analysis is based on a pure jump process for prices. The compound Poisson process (CPP) is introduced to describe the dynamics of price.Additionally, a new assumption is made on noise that generalizes the traditional i.i.d.one, which allows the variance of noise varying across the day. With the CPP model, the path of noise is separated from the observed price and then the noise path is used to improve the traditional sampling scheme. For further improvement, our new sampling scheme is combined with the transaction time sampling scheme and the first-order correction RV. Experiments on the stock price data from both stock exchanges in China demonstrate that although our estimator may produce an outlier occasionally, it can pass the robustness tests in general.

Key words: high-frequency data, realized variance, microstructure noise, compound Poisson process