[4Yin-A] Poster 6

[4Yin-A-63] Proposal for an Automated Thematic Stock Selection Method Using an Autonomous Web-Exploration Agent

〇Mitsutoshi Matsudo¹, Takehiro Takayanagi^1,3, Noriaki Saito^1,2, Yuri Murayama¹, Kiyoshi Izumi¹ (1. Univ. of Tokyo, 2. Global X Japan Co. Ltd., 3. Simulacra Inc.)

Keywords:

Thematic stock,AI agent,Web search

In recent years, the growing popularity of thematic funds has increased the time and labor of identifying theme-related listed companies and organizing investment candidates. However, compiling a comprehensive set of candidate firms while accounting for the broader economic ripple effects of a theme is challenging. Moreover, text-embedding similarity computed from disclosure documents alone may retrieve firms that merely mention buzzwords such as “generative AI” to signal innovativeness or appeal to investors, without corresponding business substance. To address these issues, we propose an automated method for thematic stock selection that (i) uses a web-search agent to broadly explore related business domains, including indirect links and spillovers, (ii) verifies relevance using web evidence, and (iii) estimates the revenue share attributable to theme-related businesses to rank candidates. Experiments using real-world thematic ETF holdings as a benchmark set show that our method achieves higher recall than the baselines, especially at larger K. Expert assessment also confirmed the practical value of expanding the search space and presenting evidence with citations. Remaining errors stem from ambiguous theme definitions; we therefore suggest explicitly specifying theme scope and adopting a human-in-the-loop workflow.