Traditional stock price forecasting models often only consider time-series data and are limited to the models own mechanism ignoring the impact of public opinion on stock prices resulting in poor forecasting accuracy. To solve this problem a feature fusion prediction model based on Bert stock forum public opinion analysis was proposed to predict the trend of stock price. Firstly Bert natural language processing was used to classify the sentiment of stock forum opinions and company?? s announcement policies and the emotion after classification was converted into virtual variables to construct the emotion feature database of financial public opinions. Then the feature fusion matrix was constructed by merging the financial public opinion feature database and the time-series data. Finally a long short-term memory LSTM network model was used to predict the closing price of the stock and to derive the upward and downward results of the stock price. Taking Huayin Power 600744. SH as an example for empirical analysis the experimental results showed that after the introduction of stock sentiment features into the model the accuracy of the stock price trend increased by 8. 63% the regression indicators of FMAPE and FRMSE for predicting the closing price decreased by 23. 59% and 22. 9% respectively and R2 increased by 8. 11%. This demonstrates that the introduction of new sentiment features of public opinion can improve the accuracy of stock price prediction in practical forecasting and can be used as a means to accurately predict the trend of stock price.