Paper: arXiv 2508.16936

Authors: 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.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)

Why this score: The paper employs advanced machine learning techniques like hierarchical contrastive learning and embedding fine-tuning, representing significant mathematical and methodological complexity. It is highly empirical, with a custom dataset (TRS), clear backtesting for portfolio performance, and quantitative comparisons to baseline models like LLMs.

Research Flowchart

  flowchart TD
  A["Research Goal"] --> B["Data Collection"]
  subgraph B ["Inputs"]
      B1["Thematic Text"]
      B2["Stock Returns"]
  end
  B --> C{"THEME Framework"}
  C --> D["Hierarchical Contrastive Learning"]
  D --> E["Align Themes & Stocks"]
  E --> F["Incorporate Returns"]
  F --> G["Refined Embeddings"]
  G --> H["Outcomes"]
  subgraph H ["Findings"]
      H1["Superior Asset Retrieval"]
      H2["High-Performance Portfolios"]
  end

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes THEME, a hierarchical contrastive learning framework that fine‑tunes stock embeddings using semantic and temporal supervision for thematic investing.

Source sentence
we introduce THEME, a framework that fine‑tunes embeddings using hierarchical contrastive learning.

Assumes 4

  • Portfolios are constructed as equal‑weighted baskets of the top‑K retrieved stocks. execution
    Source sentence
    For each window, we construct equal‑weighted portfolios using the
  • A 14‑day holding period is used to evaluate returns of retrieved stocks. execution
    Source sentence
    stocks is then recorded over the subsequent 14 trading days.
  • Two years of historical U.S. market data are sufficient for temporal refinement. data
    Source sentence
    For the temporal refinement stage, we use two years of historical U.S. market data, generating training samples via a rolling window.
  • Thematic ETFs with at least 10 constituents provide a reliable base for training. data
    Source sentence
    We filter for ETFs with at least 10 constituents. The remaining 969 ETFs are randomly split into training (678), validation (97), and test (194) sets.

Data

Backtest sample: 2024-04 – 2025-04 · US equities · equities · daily

  • Thematic Representation Set (TRS) text-news unknown
    Source sentence
    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.
  • Real‑world thematic ETFs benchmark unknown
    approximately 3,000 unique U.S. equities
    Source sentence
    The first component consists of 1,153 real‑world thematic ETFs covering approximately 3,000 unique U.S. equities.
  • Two‑year historical U.S. market data market-prices unknown
    two years
    Source sentence
    For the temporal refinement stage, we use two years of historical U.S. market data, generating training samples via a rolling window.

Costs

Not modeled A trading or execution result is reported without costs.

Finds 2

Reported performance (THEME (Ours)): Sharpe 0.77 (benchmark 0.48) · max drawdown 23.2%

  • THEME consistently improves portfolio quality over vanilla approaches, achieving higher Sharpe ratios and cumulative returns across top‑k cutoffs. performance
    Source sentence
    As summarized in Table 2, THEME consistently improves portfolio quality over vanilla approaches.
  • THEME outperforms extensive baselines including strong LLM and embedding‑based baselines in retrieval precision and hit rate. performance
    Source sentence
    In our experiments, THEME outperforms extensive baselines including strong LLM and embedding‑based baselines in retrieval precision and hit rate (Experiment 4.2).

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 20 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 2 papers in the archive

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (31)

  1. Linq-Embed-Mistral Technical Report (2024) — link
  2. SimStock : Representation Model for Stock Similarities (2023) — link
  3. Mapping Companies to Long-Term Themes Using Domain-Specific Language Models (2023) — link
  4. Learning Company Similarities from SEC Filings Using Transformer-Based Embeddings (2023) — link
  5. SETN: Stock Embedding via Text and Network for Thematic Portfolio Construction (2024) — link
  6. GPT-4 Technical Report (2023)
  7. BloombergGPT: A Large Language Model for Finance (2023) — link
  8. VoyageAI (May 25 version) (2025)
and 23 more
  1. FinMTEB: Finance Massive Text Embedding Benchmark (2025) — link
  2. Gemini (May 25 version) (2025)
  3. C-Pack: Packaged Resources To Advance General Chinese Embedding (2023) — link
  4. Unsupervised cross-lingual representation learning at scale (2019) — link
  5. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (2019)
  6. Augmented SBERT: Data Augmentation Method for Improving Bi-Encoders for Pairwise Sentence Scoring Tasks (2020)
  7. Multilingual E5 Text Embeddings: A Technical Report (2024) — link
  8. Jasper and Stella: distillation of SOTA embedding models (2025) — link
  9. Towards general text embeddings with multi-stage contrastive learning (2023) — link
  10. Improving Text Embeddings with Large Language Models (2023) — link
  11. Text Embeddings by Weakly-Supervised Contrastive Pre-training (2022) — link
  12. Generative Representational Instruction Tuning (2024) — link
  13. SFR-Embedding-Mistral:Enhance Text Retrieval with Transfer Learning (2024) — link
  14. Temporal Knowledge Graphs for Dynamic Financial Event Modeling (2023) — link
  15. Empirical Asset Pricing via Machine Learning (2020) — link
  16. FinLLM: Financial Large Language Models for Algorithmic Trading (2023) — link
  17. KGTransformer: Temporal Knowledge Graph Transformer for Financial Event Prediction (2023) — link
  18. LLMs analyzing the analysts: Do BERT and GPT extract more value from financial analyst reports? (2023)
  19. A survey of large language models for financial applications: Progress, prospects and challenges — → 2406.11903
  20. An Overview of Machine Learning for Asset Management. (2023)
  21. An Overview of Machine Learning for Portfolio Optimization. (2024)
  22. Your AI, Not Your View: The Bias of LLMs in Investment Analysis — → 2507.20957
  23. Lora: Low-rank adaptation of large language models. (2022)

2 of 31 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.