Financial market modeling integrating language information via deep learning

Research Project

Project/Area Number	21J11781
Research Category	Grant-in-Aid for JSPS Fellows
Allocation Type	Single-year Grants
Section	国内
Review Section	Basic Section 61030:Intelligent informatics-related
Research Institution	The University of Tokyo
Principal Investigator	DU XIN 東京大学, 工学系研究科, 特別研究員(DC2)
Project Period (FY)	2021-04-28 – 2023-03-31
Project Status	Completed (Fiscal Year 2022)
Budget Amount *help	¥1,700,000 (Direct Cost: ¥1,700,000) Fiscal Year 2022: ¥800,000 (Direct Cost: ¥800,000) Fiscal Year 2021: ¥900,000 (Direct Cost: ¥900,000)
Keywords	金融 / 自然言語処理 / 深層学習 / 株のベクトル表現 / ポートフォリオモデル / 言語のベクトル表現 / ファットテール / 時系列 / ポートフォリオ
Outline of Research at the Start	人工知能技術の急速な発展と言語処理技術の急速な革新により、大量のテキストの処理と理解が可能になった。この背景で、金融市場や金融活動を分析するための言語処理技術の使用は、近年重要な課題になった。本研究は、英語のニューステキストと米国の株価変動との関係を分析することにより、新しいポートフォリオ構築モデルを提出した。これからは、他の言語に対する有効性を確認し、株式以外の金融資産または高頻度データへの応用を検討する。
Outline of Annual Research Achievements	In recent years, the financial markets have faced significant challenges such as the 2018 bitcoin price crashes and the 2020 US stock market declines. While previous research has largely focused on analyzing price data, this researcher sought to take a novel approach by incorporating natural language data such as news articles. Deep learning techniques were employed to process both price and language data within a single computational framework, establishing a connection between these two complex social systems and ultimately enhancing our understanding of financial markets. In the past year, the researcher proposed a generalized model for stock portfolio optimization that integrates natural language data. This model represented a stock with a vector obtained from news articles and identified extreme risk correlations between stocks from these articles, effectively diversifying the risks. This work was accepted for publication in "Knowledge-Based Systems." Additionally, the researcher investigated the limitations of vector representations of stocks in describing complex phenomena like polysemy. To address these limitations, a new representation method using functions instead of vectors was proposed and validated on language data. The method will be further validated on financial markets and has been accepted for publication in "Advances in Neural Information Processing Systems 2022."
Research Progress Status	令和4年度が最終年度であるため、記入しない。
Strategy for Future Research Activity	令和4年度が最終年度であるため、記入しない。