Abstract: The classical Markowitz's mean-variance model in modern investment science uses variance as  risk measure while it ignores the asymmetry of the return istribution. When skewness is adopted to measure the asymmetry,  the portfolio optimization model becomes very difficult due to the nonconvex and non-quadratic features of skewness.In this paper,  we propose a mean-variance-approximate skewness (MVAS) model which measures the asymmetry by the radio of positive semi-variance and negative semi-variance. Empirical analysis of Chinese stock market shows that the portfolio models which consider  the asymmetry of the return distribution  outperform  MV and MAD models when the market has non-normal feature.