抄録
In stock investment and asset management, numerical data have been mainly used for fundamental analysis and
technical analysis. However, according to recent development of natural language processing techniques, text data can also be applied to investment. For example, news texts can be divided into some keywords such as morphemes and phrases, and then the relationship between their occurrence patterns and the following stock price movements can be learned by machine learning techniques for stock price prediction. However, the increasement of keywords often leads to large dimensional state spaces, which makes it difficult for machine learning to extract useful information. Therefore, we only use headlines included in news articles and try to generate short vectors to represent headlines by using the BERT model. To evaluate this approach, we constructed a prototype of AI investment system based on the BERT model, Word2vec, or Bag-of-Words, and confirmed the superiority of the BERT model as compared to the other models.
