Stock Price Prediction Based On LSTM And Transformer Model

The co-occurrence of high returns and risks is the typical characteristic of the stock market.Therefore,predicting the stock trend has always been a popular topic for scholars to investigate.The stock price sequence is characterized by high noise,nonlinearity,and non-stationarity.Meanwhile,the stock market is affected by many factors,so it is difficult to predict.In the early days,people used traditional time series models such as ARIMA and GARCH for modeling,but these models have strong assumptions,great limitations,and limited prediction accuracy.With the rapid development of computer technology and neural networks,two types of deep learning models have emerged that can be used to process sequence data.One is LSTM.LSTM adds a gating mechanism on the basis of RNN,which can describe the time series well,extract the characteristic information of the sequence and have memory ability.The other is Transformer,a typical Seq2Seq model that has shown remarkable performance improvement in natural language processing and advanced vision tasks in recent years.Inspired by this,this paper attempts to improve these two models and apply them to stock price prediction.This paper selects the stock data of ICBC from 2010 to 2022 as the research object,conducts forecasting modeling based on LSTM and Transformer models,adds EMD,EEMD and CEEMDAN methods in signal decomposition,and takes ARIMA as the comparison model.By comparing MAE,RMSE,MAPE and R2,this paper aims to explore better performing predictive models.The results obtained in this paper are as follows:1.Verified the feasibility and effectiveness of Transformer applied to time series tasks,established a TST(Time-Series Transformer)model for time series prediction,and R2 can reach 0.924;2.Adding LSTM into Transformer architecture and proposing LSTM-TST model can not only improve the training speed of TST model,enables faster convergence,but also improve its prediction performance significantly.The predicted R2 has increased to 0.948;3.A series of hybrid models are established by combining signal decomposition methods such as EMD with the above-mentioned neural network model.The experimental results show that the best performing models are EMD-LSTM-TST and CEEMDAN-LSTM-TST,which are significantly better than other models in each predictive evaluation index,showing the advantages of hybrid model over traditional statistical model and single neural network model.