Abstract:

Sentiment can influence investor behavior and then affect the stock market. However， not all sentiment at all points in time can cause stock price fluctuations， so the effectiveness of sentiment profiles needs to be further explored before using them for stock market prediction. The aim of this article is to use an LSTM neural network based approach for mining effective sentiment that can help stock market forecasting.First of all， text analytics is used to construct indicators for reflecting daily investor sentiment based on text data from the East Money Stock Bar. Then LSTM neural model is used to predict individual stocks and the effectiveness of the sentiment under the time points is analyzed which model predicts anomalies. SVM algorithms are used to mine from anomalous time points to obtain effective sentiment. Finally， the effective sentiment features are added to the LSTM model for prediction， comparing the difference between the normal model as well as the model with the addition of unmined sentiment features.The results of the reasearch show that the effectiveness of sentiment under the time points which model predicts anomalies is stronger； the performance of the model with the addition of the effective sentiment is significantly improved， indicating that the mining of valid sentiment can help the model to predict the stock market.Overall， an effective sentiment mining methodology is provided that can be used for stock market prediction. Help and insight are hoped to be given to those who use sentiment features to predict the stock market.

Key words: LSTM neural network, stock market prediction, effective sentiment, SVM, text analysis