Research On CNN-BiLSTM Stock Price Forecast Model And Quantitative Trading Strategy Based On Attention Mechanism

With the maturation of China's stock market,stock price forecasting has gradually become a research hotspot in the financial and economic fields.As is known to all,although the profits investors can get in the bull market are very considerable,the stock market is complicated and volatile,so investors need to take huge risks while getting profits.Therefore,in order to minimize investment risks and maximize profits,researchers have proposed a variety of stock price forecasting methods to avoid investment risks as much as possible.Among the research methods of stock price prediction,the deep learning method can enhance the ability of feature expression and extract effective information efficiently.It is applicable to the high-frequency data of the stock market with short interval and large scale,and has significantly improved the calculation speed.In addition to advantages in computational efficiency,compared with traditional time series methods,deep learning can better approximate nonlinear functions,and then solve nonlinear problems that frequently occur in practice.Based on deep learning method,this paper proposes a new stock price forecasting framework CNN-Bi LSTM-Attention.The model firstly extracts stock characteristics of historical information through CNN,and then uses bidirectional LSTM network to train and predict the minute data set of stocks.At the same time,the attention mechanism is introduced in the network training process to assign different weights to each stock feature.Firstly,this paper selects the 1-minute data sets of 600019 Baosteel,600036 China Merchants Bank,600048 Poly Real Estate and 600079 Renfu Pharmaceutical to make regression prediction on this model.From the evaluation index MAPE is less than 95%,the prediction effect of the model is very good.Secondly,compared with the single-layer LSTM model,it is found that the generalization ability and prediction accuracy of CNN-Bi LSTM-Attention model are better than that of the single-layer LSTM model,and the test set fitting results also show that the model can effectively eliminate or alleviate the phase difference problem of the single-layer LSTM model.Finally,the predicted value of the model is used to generate trading signals,and the simple moving average strategy and moving average convergence and divergence strategy are respectively used to backtest the minute test set of the four selected stocks.The backtest results show that the strategy can capture the signal points of A and B tags with the best returns in single stock transaction.Thereby,achieve the goal of four stocks are earnings,and baosteel,China merchants bank and poly real estate stock actual total yields of 30% or more,including baosteel in SMA strategy to measure the total yields the highest reached 58.24%,even in the solid plate filter out most of the losses or gains little trading posts,has higher application value.