Research On Financial Data Prediction Based On Deep Learning

The complex market behavior mechanism makes financial time series characterized by much noise,low signal-to-noise ratio,nonlinearity and nonsmoothness,and volatility clustering,which makes the traditional econometric forecasting models more challenging in financial data forecasting.Using neural network algorithms and deep learning techniques for time series forecasting of financial markets has important practical significance and theoretical value.Therefore,this thesis focuses on the following two aspects based on deep learning models:(1)This thesis introduces the VAT regularization method in LSTM that can increase the robustness of the model,and proposes the VAT-LSTM financial data forecasting model to solve the problems of nonlinear distribution characteristics and generalization errors in financial data.Through the empirical analysis of the 2015-2022 Shanghai Stock Exchange(SSE)Composite Index data,it is found that the model performs well in short-,medium-,and long-term forecasting,and has higher forecasting accuracy compared with classical time-series forecasting models such as LSTM,RNN,GRU,BP,SVR,ARIMA,etc.It is suitable as a forecasting model for financial data and can be extended for application to other fields.(2)This thesis proposes the Transformer-Encoder financial data prediction model based on attention mechanism for improving the prediction accuracy of financial data.This thesis uses the SSE Composite Index data from 2015-2022 for empirical analysis,and also compares with classical time-series forecasting models such as LSTM,RNN,GRU,BP,SVR,ARIMA,etc.The results show that the Transformer-Encoder financial data forecasting model achieves the best forecasting results in short-,medium-and long-term forecasting of financial data.The method effectively improves the accuracy of financial data prediction,is suitable as a prediction model for financial data,and can be extended to other fields.The VAT-LSTM financial data prediction model and Transformer-Encoder financial data prediction model proposed in this thesis show better prediction results in short-,medium-,and long-term prediction of financial data compared with traditional time series models(ARIMA),machine learning models(BP neural networks,SVR),and deep learning models(LSTM,RNN,GRU).However,in the long-term prediction of financial data,its prediction accuracy and model fitting degree have a certain difference from its performance in short-term and medium-term prediction,and need to conduct further research and improvement.