Exploring the impact of investor’s sentiment tendency in varying input window length for stock price prediction

Stock price prediction is one of the most important aspects of business investment plans, and has been an attractive research topic for both researchers and financial analysts. Many previous studies indicated the effectiveness of social media sentiment in stock price predictions through time series modelling. However, the time series information hidden in consecutive trading days has not been fully explored. In this paper, we build a stock price prediction model based on attention-based Long Short Term Memory (ALSTM) network using price data, technical indicators and sentiment information from social media. We employed a novel method to feed the deep network with long time series data to learn the deep sequential information of stock price movement. A fine-tuned BERT sentiment classification model and a sentiment lexicon are proposed to extract deep sentiment tendency of social media posts. We conducted experiments on 28 stocks within three years’ transaction period, and the results show that: (1) evaluated by the indicators of the Mean Absolute Error (MAE), the Root Mean Square Error (RMSE) and the accuracy, our proposed method outperforms the baseline models in both validation and test data sets; (2) models incorporating stock prices, technical indicators and sentiment features perform better than models that only use partial data source; (3) the fine-tuned BERT model performs better in sentiment classification task, and the exploitation of the sentiment features computed with the use of BERT model also led to higher predicting accuracy compared with the features calculated using sentiment lexicon; and (4) setting the input window length to 5-day achieves the best performance in average prediction accuracy.