Loading...

Proposing a Hybrid Approach based on Deep Learning Algorithms for Stock Market Prediction

Mobasseri, Niloofar | 2021

307 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 53995 (01)
  4. University: Sharif University of Technology
  5. Department: Industrial Engineering
  6. Advisor(s): Khedmati, Majid
  7. Abstract:
  8. Now a day, stock price prediction is known as one of the most challenging activities in the financial field. Research in price prediction models in financial markets, despite its many challenges, is still one of the most active areas for research. The price of non-linear financial assets is dynamic and unpredictable. Therefore, it is very difficult to arrangement and predict financial time series. Recently, many studies demonstrate that checking the news published in relation to a stock can significantly improve the accuracy of the prediction model.Among the latest techniques available for stock price prediction, we can mention deep learning models, which due to their high ability to recognize complex patterns in various fields, are among the topics that have accounted for a large volume of research. In this study, news data related to the Tesla Automotive Company's stock have been collected from reputable news websites for 11 months since the beginning of 2020, and using the FinBERT algorithm, which is an extended version of the BERT algorithm and a set of transformer models and deep learning. The effect of this News on the movement of the company's stock price has been examined. For this purpose, the multi-label classification model has been introduced in comparison with the multi-class model, which is commonly used in articles, and also the effect of news release time on stock price changes has been investigated by considering the time lag. Based on the results, the accuracy of the multi-label model introduced in this study using the F1 score is 0.846, while the multi-class model at its best has reached an accuracy of 0.822. Also, based on the results, it is proved that the effect of the published news on the day of publishing the news is more than the day after its publication
  9. Keywords:
  10. Text Mining ; Natural Language Processing ; Deep Learning ; Stock Prediction ; Bidirectional Encoder Representations from Transformers (BERT)Model

 Digital Object List

 Bookmark

No TOC