A multi-feature selection fused with investor sentiment for stock price prediction

Abstract

The stock market data has the characteristics of high dimensionality and multiple noise, and the investment behavior of investors is easily influenced by emotions, which poses a great challenge to stock price prediction. To improve the accuracy of stock price prediction, this study proposes a combined modeling approach based on multiple feature selection algorithms and incorporating investor sentiment. First, we collected stock trading data of the Chinese A-share market from 2018 to 2022, and three types of investor sentiment data sourced from social media, Internet news and newspaper news. Then, we used five feature selection algorithms to select dozens of important features from hundreds of features in the stock trading data. Based on three types of investor sentiment data, five sentiment indicators were constructed and included in the subsequent feature selection along with the previously selected important features. Finally, five deep learning models were used to predict stock prices using feature sets with sentiment indicators. A total of 1030 stocks from 10 industries such as pharmaceutical and biological, leisure service, food and beverage were selected for the experiment. The results show that in 10 industries, LSTM-CNN-Attention model has the best effect on stock price prediction, and after incorporating the sentiment indicator constructed by the principal component, the effect of the model is significantly improved, and the performance is the best in 7 industries. This method explores a new way of stock price prediction by integrating investor sentiment, and can provide further reference for the research of stock price intelligent prediction.