Research On Financial Time Series Forecasting Based On Multi-source Fusion Data

Investing in the financial market is to achieve higher profits.By predicting the trend of financial time series changes,investors can better formulate investment strategies,control risks,and improve investment returns.However,the financial market is influenced by a large number of events,and its future changes are dynamic.Therefore,the prediction of financial time series is extremely challenging.However,traditional methods typically use historical data from financial time series to predict their future development trends.Historical data is essentially a random variable,and predicting future market changes solely through it cannot intuitively reflect other influencing factors.It also ignores the correlation between influencing factors and is insufficient to obtain reliable prediction results.With the rapid growth of information in the Internet era and the development of natural language processing technology,it is possible to quantify investor sentiment from massive news text data and further explore financial market trends and volatility.Nevertheless,there are still challenges in conducting public opinion analysis on financial texts:firstly,sentiment analysis based on deep neural networks requires large-scale labeled data,and labeling financial text fragments is expensive;Second,financial texts have their own unique terminology and tend to use fuzzy expressions.However,most financial text analysis methods so far are based on the manual feature selection of "word counting",unable to understand the deeper Semantic information of financial texts;Third,because the natural language processing model restricts the length of text input,the public opinion analysis task of most financial texts only uses the news title and ignores the news text content,which makes it difficult to focus on the subjective data that can truly reflect investor sentiment,and lacks the correct processing and analysis of f the news text content.Based on the deep learning algorithm,this paper studies the related tasks of financial time series prediction.The main contributions are as follows:(1)Build a financial long text public opinion analysis network based on transfer learning(Rank-Layered Finbert-LSTM,Rank-LFLSTM)Finbert,a pre-training language model for the financial field,is used to initialize the downstream public opinion analysis model with the weight learned from the language modeling task,thus solving the problem of tag data scarcity;The method based on transfer learning carries out pre-training on the corpus of specific financial field,so as to learn the deep semantic information of financial text;Based on the Textrank algorithm,the long text is dynamically divided according to the semantic features,and the context features of the text are deeply mined,so as to accurately capture the changes in market sentiment.The experiment shows that the public opinion analysis network has made significant progress in extracting the deep semantic features of financial text,and the classification accuracy is far higher than the general model.(2)Build financial time series prediction model based on multi-source fusion dataBased on the financial text public opinion analysis task,a multi-source fusion data combining the historical price data of financial entities,financial news sentiment scores,and financial related technical indicators is proposed to solve the problem of single data feature;A bidirectional LSTM model combining attention mechanism is proposed,which makes full use of temporal features and divides different influence weights for different features.In this paper,through AUC and F1_Score fully verifies the effectiveness of this model.