Research On Stock Trend Forecasting Based On Investor Sentiment And Deep Learning

As an important part of the national economy,the stock market is becoming more and more important,so it is necessary to make a scientific prediction.Although the stock time series data has the characteristics of non-stationary,non-linear and high noise and the traditional way of forecasting effect is not significant,and the prediction has a certain challenge,the research still has not stopped.With the development of machine learning and deep learning,neural network has good nonlinear approximation ability and adaptive learning characteristics,which make machine learning and deep learning become a new direction of trying to predict the stock market.First of all,in deep learning,Long Short-term Memory Neural Network uses the ?gate? to memorize and forget the important time series information,which overcomes the long-term dependence problem of Recurrent Neural Network,so it is suitable for predicting the time series data of the stock market,and becoming the preferred model in this paper.Secondly,as an indispensable participant in the stock market,investors have a certain reference value to the judgment of the stock market.Therefore,it can help predict the future trend of the stock market by extracting the investor's sentiment.To sum up,the prediction of this paper adopts LSTM model based on the investor sentiment and the text sentiment analysis.Extracting the investor sentiment index through the text sentiment analysis of the title text data obtained by the web crawler,and adds the number of reading and comments as well as Baidu Index as the investor sentiment data.In addition,When selecting stock market data,combining the basic,technical and macro data,and add the denoising sequence of wavelet decomposition and reconstruction of closing price,which aims to increase sample diversity and provide high-quality data.When building model,adopt the text sentiment polarity analysis method based on the sentiment dictionary to build the investor sentiment index in the text sentiment analysis,then use the principal component analysis to get the preliminary features from preprocessed data.In the part of deep learning model,the obtained features are input into MLP,LSTM,GRU,CNN-LSTM,MFCN-LSTM and other models respectively for prediction and comparison,and the attention mechanism is added into the model to further explore whether it can improve the prediction accuracy.Among them,CNN-LSTM and MFCN-LSTM are LSTM model with convolutional neural network,in which the convolutional neural network is added to achieve the effect of feature extraction.Stock trend prediction is to predict whether the stock will go up or down,which is a binary classification problem.It uses the way of sliding window to predict the trend of the next day.Through the study of individual stocks,first,the investor sentiment is conducive to improving the accuracy of neural network prediction and providing guidance for investment strategies.Secondly,LSTM model and its variant GRU model are better than the general neural network model MLP in the prediction of stock time series,and the prediction in a short time is better than that in a long time.Through the use of gate,LSTM model can learn to remember and forget in time.Third,with the CNN can improve the overall effect of the model.Through the CNN,the feature information can be extracted effectively,which makes the prediction accuracy of CNN-LSTM and MFCN-LSTM better.Finally,the addition of attention mechanism can make the neural network model pay attention to more important information from the complex data to a certain extent,so as to improve the prediction accuracy of model,but the improvement is not obvious in the LSTM model which has joined the CNN model.In this paper,MFCN-LSTM model is the best model for the final prediction effect,and the average accuracy of the multiple stocks prediction is 61.11%.