The direction of stock returns is, to some extent, predictable. In this paper, we examine the predict ability of a new model in predicting the direction of stock returns. This forecasting model is based on applying and extending the time-varying probability density function theory proposed by Andrew and Vitaliy (2011).The empirical work in China stock market shows that our model have statistically significant out-of-sample predict ability of the directions of stock returns. A simple trading strategy simulation based on our model can yield a much higher investment returns than the buy-and-hold trading strategy while undertaking a lower risk. Furthermore, our model can control the correct prediction ratio of positive direction by setting different action thresholds. This allows us to have a flexible investment portfolio selection with various matches of returns and corresponding risks.