Stock Price Prediction with Bert-BiLSTM Fusion Model in Bimodal Mode

In order to comprehensively consider the impact of social opinion and public sentiment on stock prices, this paper proposes to incorporate a bimodal fusion model based on Bert and BiLSTM, which takes into account both structured data (stock price data) and unstructured data (textual information) in order to improve the accuracy of stock price prediction. The Bert-BiLSTM based stock prediction model is established by collecting the stock price data of CSI 300 sector from 1 May 2021 to 1 May 2024 and obtaining the related news text using crawler technology. After preprocessing the data, Bert is used to transform the text in each time window into word vectors, which are fused with the processed stock price data and finally passed into BiLSTM for prediction. After 40 rounds of training, the mean square error MSE of the model is reduced to 0.224, and the goodness of R2 fit is as high as 0.987. The experimental results show that the model has good fitting ability and can be applied to long-term stock price prediction.