Staging Emotions in Stock Market Data Using Natural Language Processing with Transformers

The stock market is complex and difficult to analyze and predict. One of the important factors is investors’ sentiments. In recent years, investors frequently post about the market on Social Networking Service (SNS). It is conceivable that using Nature Language Processing (NLP) to stage wmotions in SNS posts could be a feature of market analysis and forecasting. In this study, Long Short-Term Memory (LSTM), Transformer, and Transformer with sparsemax are used in the sentiment NLP task. The posts are staged in to five levels: too negative, negative, neutral, positive, and too positive. The results showed that LSTM model was superior in terms of training time and Transformer model was superior in terms of accuracy. Too positive, too negative, and neutral can be labeled with high accuracy. Therefore, the models are considered to be able to identify strong trends. The emotional labels obtained in this study could be useful in analyzing and predicting the stock market.