Diffusion Model Based Data Augmentation For Stock Movement Predictions

In recent years,the deep learning method has achieved unprecedented development,and a large number of deep learning methods have been applied in many fields including stock price and movement prediction.Financial data are usually nonlinear,high-dimensional,high-noise and unstable.Therefore,due to the characteristics of time series data,that is,only one observed value can exist at each time point,the data size of various investment targets is too samll compared with its complexity,.This paper focuses on the data augmentation of stock data based on the emerging deep generative model,the diffusion model,and explores whether an appropriate deep neural network architecture can effectively generate simulative data,so as to improve the accuracy of deep learning model in the task of stock movement predictions.This paper mainly carries on two aspects of research.On the one hand,this paper combines the LSTM network structure with the idea of diffusion model to construct a sequential diffusion model,which can be applied to the generation of stock trading data,and then used as a data augmentation method for stock movement predictions.On the other hand,the LSTM+ResNet stock movement prediction model is constructed by using a residual network to replace the fully connected network classifier of the traditional LSTM prediction network.In terms of empirical analysis,this paper tests the validity of the data augmentation method and the LSTM+ResNet stock movement prediction model on the real trading data of Shanghai Composite Index,Hang Seng Index and Nasdaq Composite Index in 1000 trading days up to February 9,2023 respectively.The results show that no matter what model is used as the movement prediction model,the LSTM-based sequential diffusion model can be effectively used as a stock data augmentation method to improve the prediction performance of the model.No matter what kind of training set is used,the LSTM+ResNet model proposed in this paper has better performance than the traditional LSTM+Linear model.As the main conclusion,this paper believes that for different prediction models,the simulation data generated by the sequential diffusion model can be regarded as the supplement of real historical data,as a means of data augmentation,to deal with the problem of the shortage of financial data,and improve the performance of the model in the task of stock movement predictions.In addition,this paper argues that the use of the residual network to replace the fully connected network classifier used in the traditional LSTM prediction model can effectively improve its performance in the task of stock movement predictions.