Stock Price Prediction Of Listed Companies In The Power Battery Industry Chain Based On ALBERT And GRU

Stock market is one of the most important components of the financial market,and stock price forecasting has always been a hot research field spanning computer science and technology,financial economy,and other disciplines.Effectively predicting the fu-ture trend of the stock market has great practical significance for individual investors,institutional investors,and government decision-making departments.Creating a scien-tific and effective method to achieve accurate prediction of stock prices is a highly diffi-cult but important academic and practical work.Manufacturing industry occupies a major position in the national economy,which is the foundation of establishing a country,the tool of rejuvenating the country,and the foundation of a powerful country.Currently,China’s economy is in the transition period from high-speed development to high-quality development.To complete this historic transformation,we must rely on high-end advanced manufacturing industry to support and drive.Building a high-end and advanced manufacturing industry with international competitiveness is an urgent need for China to enhance its comprehensive national strength,and is the only way for China to ensure national security and build itself into a world power to achieve the great rejuvenation of the Chinese nation.As a typical repre-sentative of China’s high-end manufacturing industry,the power battery industry has emerged as high-end manufacturing demonstration enterprises that master core technol-ogies and lead the development of the global industry,represented by BYD and Ningde Times.Therefore,this article selects 24 A-share listed companies in the power battery industry chain as the research object,and the main research work is as follows:1.Using crawler technology,obtain a total of 233280 daily stock review text data for all stock market trading days(486)of the research object from the Oriental Fortune Stock Bar from 2020-2021(20 randomly selected from each stock market trading day of all stocks of the research object);Obtain historical stock price data from Wind Financial Database on all stock market trading days(486)from 2020 to 2021,covering seven in-dicators:daily opening price,daily closing price,daily intraday maximum price,daily intraday minimum price,daily turnover,daily trading volume,and daily turnover.The data from the four stocks of Longpan Technology,Putailai,Zhongke Electric,and Shanshan Stock in 2021 were used as the test set,and the remaining data was used as the training set.The ratio of training set to test set is 11:1.After dividing the dataset,the data is preprocessed to convert it into the required input format for the model to be built.2.The idea of using historical stock price data to extract stock price features,using stock review text to extract emotional features,and using feature fusion methods to fuse the two features to create a stock price prediction model is proposed.Based on this idea,the ALBERT&GRU-Attention model is constructed.The model consists of three parts:an emotional feature extraction module,a stock price feature extraction module,and a feature fusion output module.The emotional feature extraction module is based on the ALBERT.Chinese.tiny model and a multi-layer perception mechanism.AL-BERT.Chinese.tiny is based on the pre training of large Chinese corpora at home and abroad,such as Chinese Wikipedia and Chinese Baidu Encyclopedia.Compared to the native BERT model,it is more suitable for Chinese natural language processing applica-tions,with a significant reduction in parameters,smaller requirements for hardware re-sources,and faster training speed.The stock price feature extraction module is con-structed based on the GRU model,and the feature fusion output module uses the atten-tion mechanism to fuse the extracted stock price features and emotional features and access the full connection layer to ultimately output the prediction results.3.Select MAE(mean absolute error),RMSE(root mean square error),MAPE(mean absolute percentage error),and R~2(coefficient of determination)as quantitative evaluation indicators for the model,and add a comparison chart of real and predicted values,a residual chart of real and predicted values,a loss change chart during training,and a R~2 change chart during training as auxiliary evaluation.4.Construct three sets of eleven comparative models and conduct comparative ex-periments using the same dataset to prove the effectiveness of the ALBERT&GRU-Attention model.Construct a single neural network prediction model that only models historical stock price data:GRU,LSTM,Bi LSTM,Transformer;A composite prediction model is constructed to model historical stock price data and stock review text,and to use the concat method for simple splicing and feature fusion:ALBERT&GRU,AL-BERT&LSTM,ALBERT&Bi LSTM,ALBERT&Transformer;A composite prediction model that models stock price data and stock review texts and uses attention mecha-nisms for feature fusion,but the stock price feature extraction module is not GRU,has been constructed:ALBERT&LSTM Attention,ALBERT&Bi LSTM Attention,AL-BERT&Transformer Attention.The final experimental results show that the main exper-imental model ALBERT&GRU-Attention originally constructed in this article has the best prediction effect among all twelve prediction models in this article,with MAE,MSE,MAPE,and R~2 reaching 1.73193,5.51327,0.05027,and 0.96565,respectively.