THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference

Hoyoung Lee, Wonbin Ahn, Suhwan Park, Jaehoon Lee, Minjae Kim, Sungdong Yoo, Taeyoon Lim, Woohyung Lim, Yongjae Lee · 2025

Read THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference by Hoyoung Lee, Wonbin Ahn, Suhwan Park, Jaehoon Lee, Minjae Kim, Sungdong Yoo, Taeyoon Lim, Woohyung Lim, Yongjae Lee (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-16936.txt.

Introduction

Hoyoung Lee and eight co-authors wrote a 2025 paper on THEME, a system that builds portfolios around investment themes like clean energy or robotics. Thematic investing is hard because themes overlap and markets shift. Text-based theme descriptions offer a route to finding matching stocks, but general language model embeddings miss financial nuance.

THEME fine-tunes embeddings with hierarchical contrastive learning. Themes are aligned with their member stocks through the hierarchy, then the embeddings are refined using actual stock returns. The result retrieves thematically aligned assets that also carry return potential. In tests, THEME beat leading large language models at thematic retrieval by a wide margin. Its constructed portfolios also performed compellingly. The framework fuses textual meaning with market dynamics in one representation.

Catalog record

CREATORHoyoung Lee, Wonbin Ahn, Suhwan Park, Jaehoon Lee, Minjae Kim, Sungdong Yoo, Taeyoon Lim, Woohyung Lim, Yongjae Lee
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-16936

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM ’25), November 10–14, 2025, Seoul, Republic of Korea ISBN: 979-8-4007-2040-6/2025/11 DOI: 10.1145/3746252.3761517 CCS: Applied computing Economics CCS: Information systems Information retrieval Hoyoung Lee, Wonbin Ahn, Suhwan Park, Jaehoon Lee, Minjae Kim, Sungdong Yoo, Taeyoon Lim, Woohyung Lim, Yongjae Lee Abstract . Thematic investing, which aims to construct portfolios aligned with structural trends, remains a challenging endeavor due to overlapping sector boundaries and evolving market dynamics. A promising direction is to build semantic representations of investment themes from textual data. However, despite their power, general-purpose LLM embedding models are not well-suited to capture the nuanced characteristics of financial assets, since the semantic representation of investment assets may differ fundamentally from that of general financial text. To address this, we introduce THEME, a framework that fine-tunes embeddings using hierarchical contrastive learning. THEME aligns themes and their constituent stocks using their hierarchical relationship, and subsequently refines these embeddings by incorporating stock returns. This process yields representations effective for retrieving thematically aligned assets with strong return potential. Empirical results demonstrate that THEME excels in two key areas. For thematic asset retrieval, it significantly outperforms leading large language models. Furthermore, its constructed portfolios demonstrate compelling performance. By jointly modeling thematic relationships from text and market dynamics from returns, THEME generates stock embeddings specifically tailored for a wide range of practical investment applications. † † cc-license: by 1. Introduction Figure 1. A t-SNE comparison of stock embeddings before and after applying THEME. (Upper Left) The baseline model fails to group stocks by investment theme. (Lower Right) In contrast, our tuned embeddings form distinct, meaningful clusters corresponding to investment themes like FinTech and Future Vehicles, significantly improving interpretability. Figure 2. Overview of how the THEME system is used in practice and how it works. Given a user-provided thematic description—such as "AI software and chipmakers"—THEME semantically embeds the input and retrieves stocks that are both thematically aligned and investment-suitable. These selected stocks can then be used to construct a portfolio for real-world investment. Thematic investing is a popular strategy that constructs portfolios around structural trends such as artificial intelligence, renewable energy, or cybersecurity. Unlike sector-based investing, thematic strategies are inherently cross-sectoral and dynamic: relevant companies often span diverse industries and change over time in response to innovation, regulation, and market sentiment. Despite its appeal, current thematic investment methods rely heavily on static ETF compositions or expert-defined lists. These approaches lack adaptability and are often slow to reflect emerging companies or shifts in thematic relevance. As a result, investors may miss out on timely opportunities or hold outdated portfolios that no longer represent the target theme. To address this gap, we propose THEME, a scalable system for thematic stock retrieval and portfolio construction. As illustrated in Figure 1 , traditional embedding models trained on general-purpose text fail to produce meaningful clusters of thematically related stocks. This highlights a key limitation: financial texts contain domain-specific semantics that general embeddings cannot fully capture. Our method addresses this by producing domain-tuned embeddings that reflect thematic structure, enabling more interpretable and effective retrieval. Figure 2 shows how THEME operates in practice. Given a user-defined thematic query, THEME encodes the theme, retrieves semantically and financially relevant stocks, and returns candidates suitable for portfolio construction. The system integrates seamlessly into real-world workflows, supporting both discretionary screening and systematic portfolio updates. At the core of THEME is a hierarchical contrastive learning framework that combines two sources of supervision. First, semantic signals derived from financial text (e.g., business descriptions, regulatory filings, and news) provide structural thematic alignment. Second, temporal signals based on recent returns guide the model to favor stocks with short-term investment relevance. This two-stage process enables the model to capture both long-term thematic meaning and dynamic market responsiveness. To support robust learning and generalization, we also construct the Thematic Representation Set (TRS) —a dataset that extends beyond real-world ETF holdings by incorporating sectoral taxonomies and news-derived themes. TRS enables THEME to support hundreds of broad and niche themes, including those not covered by existing ETF products. In summary, we make the following contributions: • A unified semantic-temporal framework for theme modeling: We propose a contrastive learning approach that embeds stocks using both textual descriptions and recent return behavior. This enables theme-aware retrieval that captures both structural meaning and short-term responsiveness. • A domain-tuned embedding space for finance: We show that general-purpose embeddings fail to organize stocks by thematic relevance. In contrast, our fine-tuned embeddings yield clearer semantic structure, improving retrieval and interpretability. • A hierarchical modeling strategy for thematic structure: Our method supports a wide range of themes, from broad categories to niche subthemes, by applying a multi-level contrastive framework that reflects the layered nature of thematic semantics. • A scalable thematic dataset (TRS): We construct the TRS, a large-scale stock-theme dataset that extends beyond ETF coverage by incorporating industry taxonomies and financial news. This improves coverage of emerging and underrepresented themes. • Real-world evaluation for portfolio construction: We implement and assess a full-system pipeline that improves retrieval precision and portfolio performance across multiple model backbones. The system is compatible with both discretionary and systematic investment workflows. 2. Related Works The rise of thematic investing in recent years has prompted research into systematic methods for identifying theme-relevant companies. A key challenge is that themes often cut across conventional industry boundaries. Prior approaches have typically relied on expert curation or static industry classifications, which struggle to capture the full breadth of relevant firms. To address this, researchers have increasingly turned to natural language processing (NLP) techniques applied to financial text. For instance, Rao et al. ( Rao et al. 2023 ) demonstrate that fit-for-purpose language models can effectively associate companies with long-term themes by comparing thematic keywords to business descriptions. Similarly, Vamvourellis et al. ( Vamvourellis et al. 2023 ) show that embeddings derived from SEC filings using fine-tuned transformer models not only reproduce standard industry groupings but also reveal new cross-sector relationships. These semantic embeddings have enabled applications such as clustering companies by thematic exposure and constructing thematic baskets. Building on this line of work, Takayanagi et al. ( Takayanagi et al. 2024 ) introduce a hybrid approach (SETN) that incorporates both textual and network-based signals to enhance stock representations for thematic investing. However, financial markets are inherently non-stationary, and a company’s thematic alignment may shift over time due to structural changes, strategic pivots, or macroeconomic developments ( Li et al. 2023a ) . Traditional static models fail to capture this temporal dynamic. In response, recent work has explored time-aware representations. For example, Hwang et al. ( Hwang et al. 2023 ) propose SimStock, a self-supervised framework that generates temporally-sensitive stock embeddings, improving performance on tasks such as thematic index tracking. In practice, dynamic thematic strategies are also gaining popularity, frequently updating exposures based on short-term signals such as news sentiment and price momentum ( Gu et al. 2020 ) . Meanwhile, large language models (LLMs) like GPT-4 have shown growing utility in financial applications ( Lee et al. 2023 ;

Read the complete text