K2Q Capital

01AI & Corporate Disclosures

PeerJ Computer Science, 2025Peer-Reviewed

From Words to Returns: Sentiment Analysis of Japanese 10-K Reports Using Advanced Large Language Models

Katsuhiko Okada, Moe Nakasuji, Yasutomo Tsukioka, Takahiro Yamasaki

Summary

This research examines whether advanced Large Language Models can extract investment-relevant information from corporate disclosures that conventional textual-analysis methods fail to identify.

Using large-scale Japanese corporate filings, the study compares LLM-based analysis with traditional dictionary and machine-learning approaches and demonstrates that sophisticated language models can uncover economically meaningful information in publicly available financial text.

Investment Relevance

Corporate disclosures contain far more information than conventional numeric datasets capture. This research demonstrates how unstructured financial information can be transformed into systematic research signals and how new AI-based signals can be rigorously benchmarked against traditional methodologies.

Read the Peer-Reviewed Research

02Dynamic Investment Themes

Research in Progress

Momentum Without Momentum: Text-Based Themes and Return Predictability in Japan

Summary

Traditional sector classifications assume that economically related companies belong to the same industry. This research asks whether that assumption itself may obscure important investment information.

Using semantic analysis of corporate business descriptions, companies are grouped according to what they actually do rather than according to their official industry classification. These data-driven themes connect economically related firms across conventional sector boundaries.

The evidence suggests that return predictability becomes visible when economically persistent themes are identified, even in a market where conventional stock and industry momentum are weak.

Investment Relevance

The research suggests that the way investors classify companies can itself determine what information they see. AI-based classifications may reveal latent economic relationships — including shared business activities, supply-chain exposures, technologies, or sources of fundamental risk — that static industry classifications fail to capture. For asset managers, this opens a broader research question: what other investment signals become visible when we change the way the investment universe is defined?

03Deep Learning & Market Information

Research in Progress

Decoding the Unique Price Behavior in the Japanese Stock Market with Convolutional Neural Networks

Summary

This research investigates whether the visual structure of historical price movements contains information that conventional return-based models fail to capture.

By treating financial price histories as images and applying convolutional neural networks, the study explores a fundamentally different representation of market information.

Investment Relevance

Predictive information may depend not only on which variables are used, but also on how financial information is represented. This illustrates K2Q's broader research philosophy: unconventional representations of familiar data can reveal patterns invisible to conventional models.

Featured Research & Applications

From Words to Returns: Sentiment Analysis of Japanese 10-K Reports Using Advanced Large Language Models

Momentum Without Momentum: Text-Based Themes and Return Predictability in Japan

Decoding the Unique Price Behavior in the Japanese Stock Market with Convolutional Neural Networks

Additional academic research