The prediction of financial time series data is a challenging task, mainly due to unprecedented changes in economic trends and conditions, and on the other hand, incomplete information. Market volatility in recent years has caused serious problems in economic and financial time series forecasting. The rise or fall of stock prices plays an important role in determining the returns of investors. In the case of the stock market, the resulting data is huge and highly non-linear. The deep learning model can effectively fit this type of data and can provide good predictions by analyzing the interaction and hidden patterns in the data. We can see that most of the research uses different deep learning models for stock forecasting, but these methods are applied to different experimental data and different parameter configurations, and we know that different data have different time series features. We don't know which method is the best. This study is based on 45 Taiwan-listed companies. We use the existing deep learning method to find the best parameter configuration when adjusting different parameters, compare the prediction differences between models, and finally apply deep learning. The method of predicting the construction of a stock trading system. . The empirical results show that the application of the deep learning method to predict the constructed stock trading system, after deducting the fee and securities transaction tax, its return on investment exceeds the buy-and-hold trading strategy. In addition, the risk learning indicators of the deep learning system simulation transaction are better than the buy-and-hold strategy, indicating that the trading system we construct is indeed remunerative.