Aimed at better modeling stock returns and finding robustly optimal investment decisions, a new portfolio selection model is proposed in this paper. The model differs from existing ones in following ways: multiple market frictions are taken into account simultaneously; the adopted multivariate t-distribution can capture the well-recognized fat tails in the return data by adding only one more parameter relative to the normal; the downside loss risk is controlled by a chance constraint which, including VaR as a special case, is flexible in terms of adjusting the threshold return and the loss probability level; one important advantage about the combination of the latter two innovations is that the derived asset allocation model can be transformed into a second-order cone program or a linear program, which can be easily solved in polynomial time. Empirical results based on some S&P 500 component stocks not only demonstrate the practicality of our new model, but show how different model parameters could affect the optimal portfolio selection. This is very useful in guiding investors to choose a correct model and to find the investment strategy most suitable for their specific purpose.