Research On Financial Time Series Prediction And Application Based On Deep Learning

Along with the rapid development of information technology,the cross-fertilization between computer technology represented by artificial intelligence and various industries in society has become increasingly close.Among them,intelligent models based on machine learning are widely used in various researches in the field of finance.Financial time series forecasting is a method to study the future operation trend of financial market by using relevant technical means,which has important theoretical significance and application value for its research.However,financial time series data usually have typical characteristics such as large volatility,high noise and strong nonlinearity,and there are many difficulties in forecasting research on such data.Deep learning has been developed rapidly in recent years and has shown powerful feature learning ability and model performance in many fields.In this thesis,we construct various types of deep learning models for feature modeling,forecasting and their applications on financial data such as stock price series,and the main contributions and innovations are as follows:(1)A multidimensional financial time series forecasting model based on spatio-temporal hypergraph learning is proposed to address the problem of how to effectively model the spatio-temporal correlation that exists between multiple stock series.For time-series feature learning,the LSTM network + Attention mechanism is used to model the time-series dependencies of multidimensional stock sequences.For spatial feature learning,the a priori correlations that exist between stocks are modeled using hypergraph models by employing a priori knowledge such as industry relationships and wiki relationships.To compensate for the weakness of a priori knowledge in terms of timeliness,the correlation between the time-series price series of stocks is further measured by introducing mutual information technology through a data-driven approach,and fusion learning of features is achieved through a graph learning framework with a hypergraph model.(2)To address the problem of how to effectively learn time-sensitive features in financial time series,a multi-feature fusion Conv LSTM network prediction model is proposed using the research idea of converting time series into two-dimensional images.First,the time series and its features are converted into visibility maps by the Gramian Angular Difference Fields technique.Second,a novel multi-feature fusion method is proposed to learn the long-term time dependence in financial time series.Finally,a fuzzy control module is introduced into gate mechanism of Conv LSTM,where self-similarity and time/date sensitive fluctuation patterns in data are processed.(3)Using the above proposed financial time series forecasting model,a multi-stock recommendation system and a foreign exchange trading system are constructed respectively,and the proposed model is analyzed empirically based on a real financial time series dataset in a relevant way.