With the rapid development of economy and technology,the stock market has become an important part of the current financial market.How to accurately predict stock trends has become a topic of concern for investors.The rise and fall of stock prices are affected by many factors,and the stock market itself has strong volatility and instability.So how to tap the deep market rules of the stock market has become a big challenge.Traditional machine learning methods have great limitations when faced with non-linear,high-noise,and highly volatile stock timing prediction problems.In recent years,the rise of deep neural networks has provided new solutions to stock trend forecasting problems.In order to solve the problem of vanishing gradients of long-term sequences,this paper is based on the long-term and short-term memory network(LSTM),and selects the stock trading data of the A-share section for nearly 10 years as the original data set.And the main work and results are as follows:(1)Constructed a multi-category feature system as input of long-term and short-term memory networks for training,totally including 38 input features such as basic indicators,technical indicators,various key turning point features,and real event information of individual stocks.Introduced the basic concepts and calculation methods of various indicators,and the data pre-processing process.Meanwhile,the experimental part comprehensively analyzes the effectiveness of various features for stock trend prediction,and the comparison results show that the multi-category features system performs well in the prediction,and can reach the 68.77% accuracy rate of short-term fluctuation prediction.(2)Introduced the principle and training process of neural network,and the LSTM model is optimized in many aspects: The number of hidden layers of the LSTM network and the sliding window size of the input features were determined through experiments.In the back-propagation process,a batch gradient algorithm is combined with the Adam optimization algorithm that adds an exponential decay learning rate to update the model weights,and an attention model is introduced to assign weights to different input features to enhance the extraction of important features.Regularization and early stopping strategy are added to optimize the overfitting problem.(3)The LSTM model is compared with commonly used models such as Convolutional Neural Network(CNN),Recurrent Neural Network(RNN),Multi-layer Perceptron(MLP),and the experimental results show that the deep neural network has significantly improved the prediction effect compared with the previous shallow structure model,and LSTM is superior to other deep neural network models in the experiment in solving stock timing prediction problems. |