Predicting stock price trends based on financial news articles and using a novel twin support vector machine with fuzzy hyperplane

• Published in: Applied soft computing 2021-01, Vol.98, p.106806, Article 106806
• Main Authors: Hao, Pei-Yi, Kung, Chien-Feng, Chang, Chun-Yang, Ou, Jen-Bing
• Format: Article
• Language: English
• Subjects: Fuzzy set theory; Latent Dirichlet allocation; Sentiment analysis; Stock price trend prediction; Text mining; Twin support vector machine
• ISSN: 1568-4946; 1872-9681
• DOI: 10.1016/j.asoc.2020.106806

This study extracts the hidden topic model and emotional information from news articles. A novel fuzzy twin support vector machine is also developed to merge the large volume of information from on-line news, using this to predict stock price trends. Fuzzy set theory is very useful for this approach because the texts use fuzzy terminology, such as high/low and big/small, and the boundary between rising and falling categories is poorly defined. By using fuzzy partial ordering relation, the decision function of our approach is generalized such that the values assigned to the stocks fall within a specified range and indicate the membership grade of these stocks in the positive class (rising trends). The proposed approach is useful for dealing with outliers. Outliers only slightly increase the spread of the fuzzy decision boundary and the estimation of the mode of fuzzy hyperplane remains unchanged. Therefore, our model is more robust compared with other methods when the data contain some outliers. Besides, the membership grade estimated by the fuzzy decision boundary gives a better insight into the degree of confidence in the predicted outputs. The explicable characteristic for the degree of confidence is a vital aspect for decision-making applications. •This study combines the superior power of SVM and fuzzy set theory.•This study acquires high-level expression features in financial news articles.•The proposed fuzzy decision function is based on fuzzy partial ordering relation.•The proposed method gives an insight into the confidence level in predicted outputs.•The proposed approach is useful for dealing with outliers.•Our model is more robust compared with other methods when the data contain noises.