Stock Trend Forecasting And Trading Research Based On Multimodal Fusion

With the rapid development of China’s economy and the continuous improvement of China’s stock market,more and more people choose to meet the demand for wealth appreciation by investing in stocks.The characteristics of the stock market are high-risk and high returns.However,many investors lack relevant investment theory knowledge and practical experience,which will exacerbate their investment risks to reduce their investment income.How to invest in stocks reasonably has become an important issue for current stock research.More and more studies have shown that they can predict stock trends through numerical data or news review data,thereby helping investors to avoid risk increases their income reasonably.However,the existing research is mostly based on text or numerical modal state,and there is no information about another mode.In order to make full use of text value dual-modal data to improve the predictive effect on stock trends,this article proposes a multi-mode fusion model for stock trend prediction based on stock historical data and stock review text,and has achieved the following results:(1)A text extraction module based on LERT-Text CNN-ATT is proposed.This module consists of three parts: LERT pre-training model,Text CNN,and attention mechanism.Using LERT’s powerful semantic ability to understand the vector transformation of a single comment and extract different levels of semantic characteristics from it.Using Text CNN’s capture ability of keywords to comprehensively extract the semantic semantic semantic semantics from the 12-layer network of LERT to obtain the semantic characteristics of a single comment text.Finally,the text characteristics of the focus mechanism focus on the key comments are focused on the sum of the text characteristics of the day.Numerical data pre-processing is to calculate stock technical indicators through stock historical data,and form stock numerical data with stocks as historical data.Extract numerical features through CNN networks to capture short-term dependencies and local models of stock data data.(2)Proposed a characteristic fusion method based on dual-linear poolization.This method maps them to high-dimensional space by conducting external accumulation of multi-mode,and the purpose of learning the correlation between different modes in the high-dimensional space to achieve the purpose of fully tapping and fusion of multi-modal information.The two modes of text value are transformed by dual-linear poolization,which obtains the fusion characteristics of a variety of modular information,and then the prediction label of the stock trend can be obtained through LSTM.A horizontal model comparison experiment and vertical model ablation experiments were performed on the data set composed of stock historical data and stock comments.Results show that the model of the model of this article compared the CNN+LSTM model that does not use text modal information.The accuracy can be increased by 3.43%,and the Macro F1 increases to 2.01%.The use of multimodal information can improve the predictive effect of the model on the stock trend.Compared with the CNN+LSTM+BERT model using the text emotional tendency index,the accuracy is increased by 1.49%,and the Macro F1 increases to 1.46%.It proves that the comparison indicator of text information represented by quantification can improve the model’s predictive prediction of the model for the stock trend.Finally,the average excess yield of 6.04%and an annualized yield of18.48%were obtained at the return test range of the multi-mode fusion model in June to December.It is proved that the model is feasible in constructing quantitative timing strategy and has certain practical value.