Research On Financial Time Series Forecasting Based On Deep Learning

With the development of the financial market,a large amount of financial time series data is generated in the industry,especially stock price time series data.Traditional linear models cannot satisfy nonlinear stock price predictions.With the development of neural networks,it is possible to predict complex stock price data,but neural network prediction will bring uncertainty to the prediction results.The most common method of improvement is to combine the strengths of existing individual models.The first part of this thesis realizes the prediction of the Shanghai Composite Index closing by VMD and Bi LSTM fine-tuned based on transfer learning.The second part uses Bootstrap combined with the combined model of VMD and Bi LSTM proposed in the first part to accomplish interval prediction to realize the quantification of uncertainty,and the experiment adds multi-step forward prediction method.The main work of this thesis is as follows:(1)The VMD of transfer learning and bidirectional long short-term memory network algorithm are proposed to predict the closing price of the Shanghai Composite Index.First,use VMD to decompose the opening price of the Shanghai Composite Index to obtain IMFs.Then use Bi LSTM to predict each IMF at the opening price,and save the Bi LSTM model.The closing price is decomposed by VMD to obtain the same number of IMFs next.At this time,each IMF is fine-tuned with the saved Bi LSTM model to obtain the final Bi LSTM model.The experimental results verify the accuracy and effectiveness of the proposed model by comparing and analyzing the prediction results of the proposed TL-VMD-Bi LSTM model and seven models.(2)This thesis proposes the Bootstrap method combined with the hybrid algorithm of VMD and Bi LSTM to forecast the closing price of the Shanghai Composite Index in an interval.First,the Bootstrap method is used to simulate the closing price to generate 1000 sub-data sets.In this thesis,the generated subsets are reordered in chronological order,and 1000 new subsets are obtained.Then,models such as VMDBi LSTM were used to predict the 1000 data sets respectively,and the prediction interval was calculated using the standard deviation interval estimation and the percentile interval estimation,and the predicted intervals were evaluated by PICP,MPIW and CWC respectively.In the experimental process,the fine-tuning method in transfer learning is used to transfer the model trained on the closing price of the original Shanghai Composite Index to these 1000 subsets,which shortens the prediction time.The experiments finally verify the effectiveness and robustness of the proposed model through 3-day,6-day and 9-day multi-step forward prediction methods.