Incorporating stock prices and news sentiments for stock market prediction: A case of Hong Kong

• Published in: Information processing & management Vol. 57; no. 5; p. 102212
• Main Authors: Li, Xiaodong, Wu, Pangjing, Wang, Wenpeng
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 01.09.2020; Elsevier Science Ltd
• Subjects: Data analysis; Data mining; Deep learning; Dictionaries; Indicators; Machine learning; Neural networks; News; News sentiment analysis; Prediction models; Predictions; Prices; Pricing; Securities markets; Sentiment analysis; Stock exchanges; Stock prediction; Stock prices; Test sets; Hong Kong China; China; Deep learning; 99-00; Stock prediction; News sentiment analysis; 00-01
• DOI: 10.1016/j.ipm.2020.102212
• ISSN: 0306-4573

•Based on both technical indicators and news sentiments, the LSTM models outperform the MKL and the SVM in both prediction accuracy and F1 score.•The LSTM models incorporating both information sources outperform the models that only use either technical indicators or news sentiments, in both individual stock level and sector level.•Among the four sentiment dictionaries, finance domain specific sentiment dictionary (Loughran-McDonald Financial Dictionary) models the new sentiments better, which brings at most 120% prediction performance improvement, comparing with the other three dictionaries (at most 50%). Stock prediction via market data analysis is an attractive research topic. Both stock prices and news articles have been employed in the prediction processes. However, how to combine technical indicators from stock prices and news sentiments from textual news articles, and make the prediction model be able to learn sequential information within time series in an intelligent way, is still an unsolved problem. In this paper, we build up a stock prediction system and propose an approach that 1) represents numerical price data by technical indicators via technical analysis, and represents textual news articles by sentiment vectors via sentiment analysis, 2) setup a layered deep learning model to learn the sequential information within market snapshot series which is constructed by the technical indicators and news sentiments, 3) setup a fully connected neural network to make stock predictions. Experiments have been conducted on more than five years of Hong Kong Stock Exchange data using four different sentiment dictionaries, and results show that 1) the proposed approach outperforms the baselines in both validation and test sets using two different evaluation metrics, 2) models incorporating prices and news sentiments outperform models that only use either technical indicators or news sentiments, in both individual stock level and sector level, 3) among the four sentiment dictionaries, finance domain-specific sentiment dictionary (Loughran–McDonald Financial Dictionary) models the news sentiments better, which brings more prediction performance improvements than the other three dictionaries.