Analysis And Prediction Of Financial Time Series Based On Deep Learning And Network Big Data

The investigation of time series is challenging due to the non-linear,complex,selfsimilarity and time-varying nature of time series data.In the background of the rapidly developing big data technology as well as artificial intelligence,mining the information hidden in the time series data becomes an important step in the decision making of various industries.In this paper,we focus on typical financial time series as the research object,based on behavioral economics,and take web big data represented by Baidu index and Google trends as the exogenous data for investigating financial time series,and use deep learning and other methods to carry out research on the prediction and correlation analysis of financial time series.This paper includes the following three aspects of research.The MF-DCCA method is applied to analyze the cross-correlation between Baidu index and RMB exchange rate and the change of correlation before and after the epidemic,and the predictability of Baidu index on RMB exchange rate is discussed.Then,the Baidu index was used to measure the public's online attention to the RMB exchange rate as a potential indicator for the RMB exchange rate prediction using a deep learning model.The decomposition integrated deep learning model WOA-STL-LSTM with optimization is constructed by combining the whale optimization algorithm,STL and LSTM.The experimental results show that there is a strong anti-persistent long-range cross-correlation between RMB exchange rate and Baidu index,and this relationship is less affected by the epidemic.In addition,the Baidu index can be used as an useful feature for RMB exchange rate forecasting,and the combined WOA-STL-LSTM model has better performance in RMB exchange rate forecasting compared with traditional models.Using Google Trends as a potential indicator to research the price of WTI crude oil,LSTM is used to forecast the price of WTI for the next trading day.Then,complexity analysis theory was applied to explain the role of Google Trends in improving WTI forecasts,including the self-similarity and complexity of time-series data using Hurst index,fractal dimension and Lyapunov index,and the directionality and causality of information transmission between Google Trends and WTI were quantified using transmission entropy.The experimental results show that the ability of Google Trends to improve WTI forecasts and the weak dynamic impact are attributed to the nonlinear,chaotic behavior and long-term memory characteristics of the time-series data of Google Trends and WTI at different time scales,and that the information flow between the two time-series variables is bidirectional,asymmetric,and time-varying.Five small samples of financial time series datasets are investigated using deep learning methods and data augmentation techniques.Linear interpolation and fractal interpolation algorithms were applied to the training datasets,respectively,by which the data length and granularity were increased.The interpolated data are also used for the training of deep learning models,and the improvement effect of data enhancement methods on the prediction of small sample time series data is explored.Besides,the matching mechanism of different interpolation algorithms to deep learning algorithms is discussed in this paper.The experimental results show that the time series prediction after linear interpolation and fractal interpolation is significantly better than the prediction of the original data,and the fractal interpolation is better than the linear interpolation.Based on the concept of self-similarity,this study provides a feasible solution and evidence for improving the prediction of small-sample time-series data.