Forecast Declines Based On Stock Data At Different Points In Time From LSTM-GAT

With the development of artificial intelligence technology,researchers have tried to apply deep neural networks in production research in various industries.In the field of financial quantitative trading,the use of technology to predict stock market patterns and thus make decision recommendations for investment risk management is one of the popular research directions.In this paper,by analyzing the shortcomings of conventional deep neural network models to predict the decline of a single stock method,we introduce a graph attention mechanism to improve deep neural network models,and the main contributions are as follows:1.To address the problem of single dimensionality of time series prediction samples,this paper constructs an EMD-LSTM-GAT model to learn the temporal relationships between different stocks by learning other stock data as variables together into the model for training and learning.The signal noise reduction processing is used to decompose the sample data into different IMF branches by EMD method,and then the temporal relationship of stock data is learned by LSTM model.The stock data of other companies are used as samples to learn the spatial relationships between different stocks using GAT as the training of features for prediction,and more robust prediction results are obtained.2.To address the problem of finding the best feature vector construction method,this paper uses three connection methods to connect different stock data-global graph,explicit connection and implicit connection to construct feature vectors into the model,generate feature matrices,and train the model for each feature matrix to find the best connection method of nodes between different stocks.The experiments prove that the feature matrices constructed by the display graph approach increase the correlation weights of the stocks of companies belonging to the same industry classification by considering industry correlation,which makes the prediction results calculated more rapidly and accurately.3.For the model network to adapt to the changing situation of graph structure generated over time,this paper adopts the stock nodes at the moment of T-1 as virtual nodes and selectively incorporates them into the stock nodes at the moment of T.Through the propagation aggregation of features,the local nodes can learn more information about the history.Finding the spatio-temporal relationships between different stocks at different moments through model training can cope with more complex and realistic stock market changes and improve the reference value of prediction results.4.In order to find the best way to construct the feature matrix and the effective information gain brought by the virtual nodes,this paper experimentally predicts the rise and fall of stocks in the next 5 days from 2018 to 2023,and successively tests and compares the feature matrices of the global graph,explicit graph,and implicit graph,as well as the training prediction after adding virtual nodes with different time points to get the best training scheme for the entry model.Through a combination of theory and experiments,this paper demonstrates that the proposed improved model is more accurate and robust in predicting stock declines by optimizing the prediction results,and proves that the model has improved to a certain extent in terms of noise immunity and credibility.