Financial sentiment analysis with Deep Ensemble Models (DEMs) for stock market prediction

Abstract

The stock market forecasting is popular research topic for analysts. In this study, it is proposed to estimate direction of Bist100 index by financial sentiment analysis. To our knowledge, this is the first study in literature using Twitter for forecasting stock market direction and doing this with deep ensemble models. The contributions of study are fourfold: First, feature set is enriched semantically to eliminate size limitation problem in Twitter. In first stage, meaningful features that express dataset are selected by means of information gain and ant colony optimization. Next, features are enriched in meaning, context, syntax using document representation models such as Avg(Word2vec), Avg(Glove), Avg(Word2vec)+Avg(Glove), TF-IDF+Avg(Word2vec), TF-IDF+Avg(Glove). Secondly, it is proposed to improve system performance performing classification with multiple learning algorithms. Instead of traditional classification algorithms, a deep ensemble model (DTM) is constructed blending deep learning architectures such as Convolutional Neural Networks, Recurrent Neural Networks, Long Short-Term Memory Networks. Third, majority voting and stacking methods are used to obtain final decision of deep ensemble model. Fourthly, Turkish and English Twitter datasets are employed to demonstrate that proposed approach improves classification performance. Consequently, experimental results show that proposed model is significantly superior to previous studies when compared with literature studies.