Financial Time Series Forecasting Based On CEEMDAN And Deep Learning

In the modern market economy,the financial market is in a central position and plays a very important role in the stable operation of the economy.With the development of the internet and the improvement of financial markets,the amount of financial data is becoming increasingly large and the fluctuations are more complicated.How to accurately predict the trend of the financial time sequence has attracted investors’ widespread attention.In recent years,more and more scholars have made multi-angle predictions in the financial market.Practice has proved that the intelligent financial prediction model can effectively improve the accuracy of diversified non-linear financial data predictions and play an important role in promoting financial market investment decisions.This thesis combines the CEEMDAN algorithm with deep learning models to establish a new multi-scale mixed time sequence of stock index prediction methods.First,the CSI 300 index is decomposed into a series of different characteristics.It represents short-term investors’ emotions,policies and regulations,and the long-term trend of stock movements.According to the prediction results by various methods under different frequencies data,finally determine the combined prediction method: using CNN-LSTM prediction high-frequency sequence,GRU prediction medium frequency sequence,ARIMA forecast low-frequency sequence;integrating various component prediction results as the closing price output.Select RMSE,MAPE,MAE as evaluation indicators,and compare the proposed models with single models(ARIMA,GRU,LSTM)and mixed models(CNN-LSTM,CEEMDAN-GRU,CEEMDAN-LSTM).The results show that the proposed model prediction is closer to the real value and the prediction accuracy is better.Finally,in order to verify the universality of the model,stock data of ten representative companies in the Chinese stock market have experimented as research objects,and they also achieve better prediction results.Non-linear,incredible financial time sequences are decomposed into multiple stable sub–sequences through CEEMDAN,and reconstruct them into components of different frequency to reduce computational scale,and explained the economic meaning better.The model is trained to obtain more accurate prediction result than traditional methods and neural network methods.Therefore,the combination prediction model based on CEEMDAN and deep learning has potential application value in financial time series prediction.