Abstract: A novel dual-stage mean-CVaR investment strategy based on long-short term memory networks (LSTM) is proposed, namely, LSTM+CVaR model. Specifically, in the first stage, the LSTM is used to predict the return of stocks and select stocks. In the second stage, the mean-CVaR model is used to determine the weight of every stock. Finally, using China Securities 300 Index component stocks as sample, considering the influence of threshold constraints and transaction costs, the characteristics of risk and return, cumulative return and Sharpe Ratio of portfolio models are compared in this paper, and the portfolio models include LSTM+CVaR model, equal weight model based on LSTM prediction, CVaR model based on random stock selection, equal weight model based on random stock selection and China Securities 300 Index component stocks. The empirical results demonstrate that the LSTM+CVaR model achieves higher average returns, return-risk ratio, cumulative returns, Sharpe Ratio, and decreasing the transaction costs or increasing the threshold constraint can improve the performance of the above model.