Research On Time Series Trend Prediction By Fusing Multimodal Data

In recent years,time series trend prediction has become a hot topic in industry and academia.It has been widely used in various practical applications,such as stock trend prediction,temperature prediction,and traffic flow prediction.With the rapid development of the Internet,multi-modal data related to time series data is increasing rapidly.Trend prediction based on only a single time series data has its limitations,which makes more applications involve multi-modal information.Due to the high randomness of multi-modal time series data,the generalization ability of some existing prediction models is weak.Furthermore,some models suffer from the problem of gradient explosion or gradient vanishing in the training process.To address the above problems,this thesis proposes the following two works:(1)To improve generalization ability of some existing prediction models and alleviate gradient explosion or gradient vanishing in the training process,we propose a novel stock movement prediction model(named as SVGN)based on Gated Orthogonal Recurrent Unit(GORU)and Variational Auto-Encoder(VAE).Specifically,GORU encodes the text information,and then VAE infers and decodes the information with standardized historical price information on the market information formed by the splicing of the encoded text.At the same time,the orthogonality introduced by GORU can alleviate the problem of gradient explosion or gradient disappearance,and enhance the generalization ability of the model.We evaluate the relative contributions of text information and historical prices with respect to prediction accuracy by the ablation experiments.The experimental results on publicly available datasets show that SVGN is better than several state-of-the-art models such as Stock Net and Adv-LSTM,which indicates that the GORU and VAE can effectively improve the model’s generalization ability and accuracy for predicting stock trends.(2)To address the problem that the SVGN model cannot effectively handle the interdependence of multivariate time-series data,we propose a multivariate time-series trend prediction model(named as MTTM)based on multi-level context.First,in order to capture the time dependence within each time series data,MTTM performs trend decomposition on the original time series data and encodes it to obtain the hidden state.Then,the hidden state of a single time series and the hidden state of the comprehensive time series are fused by using the multi-modal data fusion module.Finally,the selfattention mechanism is used to combine the effects of different time series and generate the final prediction for each time series trend.Experimental results on three publicly available datasets show that MTTM achieves higher accuracy,MCC and F1 scores than several state-of-the-art trend prediction methods such as Informer and Adv-LSTM,indicating that the MTTM model has better predictive performance in the field of multivariate time series prediction.The SVGN model mainly predicts the time series data of stocks and achieves more accurate prediction results.However,this method needs to use stock text information for prediction,but its prediction ability for other time series data without text information is weak.And SVGN cannot effectively deal with the interdependence of multivariate time series data.To address the problem of SVGN,the MTTM method makes use of Transformer’s strong feature extraction capability and parallel processing ability to predict multivariate time-series data,and obtains good prediction results.Both SVGN and MTTM provide effective methods for trend prediction,and also provide some possible directions for future work.