In the literature, the mean-variance model has been shown to perform poorly on out-sample data. The portfolio derived from the mean-variance model is very sensitive to its parameters, i.e., the expected return of each asset and the covariance matrix between assets’ returns. However, these parameters calculated from historical data often fail to reflect the dynamics of assets’ future returns. To overcome this problem, this thesis proposes several alternatives to estimate these parameters. Firstly, a dynamic model and an optimistic expected return are proposed to replace the traditional way of using mean for calculating the expected return. Then, the calculation of the covariance matrix is based on the optimistic expected return. Our experimental results show that using the dynamic model and the optimistic expected return can derive portfolio with better performance on out-sample data.