In today’s economic era, investment in the stock market has become increasingly widespread, and the philosophy of “Don’t put all your eggs in one basket!” is commonly adopted by investors. In other words, they will have appropriate investment portfolios to diversify their investment plan, aiming to reduce risks and improve returns. Many models, from machine learning to deep learning, from traditional to modern methodologies, have been proposed and applied to the “portfolio optimization” problem. This study uses several deep learning models, combining the attention mechanism to predict the investment portfolios with the highest Sharpe ratio over a certain period. Our research investigates whether long-term investing (forecasting a portfolio over a month or longer) is more effective than Short-term investment. We obtained some initial results by experimenting with Vietnam’s stock data set from 2016 to 2020, with 50 stocks with a high market capitalization (data as of the end of 2020). Long-term investments tend to provide more stable returns, and the Sharpe ratio of the portfolio is higher. Moreover, the bidirectional GRU model, combined with the attention mechanism, gives quite good results compared to other models. In addition, processing the missing data using the forward fill method also shows better results than when the missing data problem is not addressed.