関岡 冴俊

In recent years, unspecified messages posted on social media have significantly affected the price fluctuations of online-traded products, such as stocks and virtual currencies. In this study, we investigate whether information on Twitter and natural language expressions in tweets can be used as features for predicting market information, such as price changes in virtual currencies and sudden price changes. Our method is based on features created using Sentence-BERT for tweet data. These features were used to train the light-gradient boosting machine (LightGBM), a variant of the gradient boosting ensemble framework that uses tree-based machine learning models, with the target variable being a sudden change in closing price (sudden drop, sudden rise, or no sudden change). We set up a classification task with three labels using the features created by the proposed method for prediction. We compared the prediction results with and without these new features and discussed the advantages of linguistic features for predicting changes in cryptocurrency trends.