PDF Download

[Link]
30 December 2025
Total Citations: 0
Total Downloads: 104
.
.
Latest updates: hps://[Link]/doi/10.1145/3746252.3761517

.
.
Published: 10 November 2025
.
.
.
RESEARCH-ARTICLE

.
Citation in BibTeX format
THEME: Enhancing Thematic Investing with Semantic Stock

.
.
CIKM '25: The 34th ACM International
Representations and Temporal Dynamics Conference on Information and
Knowledge Management
HOYOUNG LEE, Ulsan National Institute of Science and Technology, Ulsan, South Korea November 10 - 14, 2025
Seoul, Republic of Korea
.
WONBIN AHN, LG Corporation, Youngdungpo-gu, South Korea

.
.
Conference Sponsors:
.
SUHWAN PARK, Ulsan National Institute of Science and Technology, Ulsan, South Korea SIGWEB
SIGIR
.
JAEHOON LEE, LG Corporation, Youngdungpo-gu, South Korea
.
MINJAE KIM, LG Corporation, Youngdungpo-gu, South Korea
.
SUNGDONG YOO, LG Corporation, Youngdungpo-gu, South Korea
.
View all
.
.
Open Access Support provided by:
.
LG Corporation
.
Ulsan National Institute of Science and Technology
.
CIKM '25: Proceedings of the 34th ACM International Conference on Information and Knowledge Management (November 2025)
hps://[Link]/10.1145/3746252.3761517
ISBN: 9798400720406
.
 THEME: Enhancing Thematic Investing with Semantic Stock
Representations and Temporal Dynamics
Hoyoung Lee∗ Wonbin Ahn∗ Suhwan Park
Ulsan National Institute of Science LG AI Research Ulsan National Institute of Science
and Technology Seoul, Republic of Korea and Technology
Ulsan, Republic of Korea [Link]@[Link] Ulsan, Republic of Korea
hoyounglee@[Link] suhwan@[Link]

Jaehoon Lee Minjae Kim Sungdong Yoo

LG AI Research LG AI Research LG AI Research
Seoul, Republic of Korea Seoul, Republic of Korea Seoul, Republic of Korea
[Link]@[Link] [Link]@[Link] [Link]@[Link]

Taeyoon Lim Woohyung Lim† Yongjae Lee†

LG AI Research LG AI Research Ulsan National Institute of Science
Seoul, Republic of Korea Seoul, Republic of Korea and Technology
[Link]@[Link] [Link]@[Link] Ulsan, Republic of Korea
yongjaelee@[Link]

Abstract Keywords
Thematic investing, which aims to construct portfolios aligned with Hierarchical Contrastive Learning; Thematic Investing; Information
structural trends, remains a challenging endeavor due to overlap- Retrieval; Stock Selection; Representation Learning
ping sector boundaries and evolving market dynamics. A promising
ACM Reference Format:
direction is to build semantic representations of investment themes Hoyoung Lee, Wonbin Ahn, Suhwan Park, Jaehoon Lee, Minjae Kim, Sung-
from textual data. However, despite their power, general-purpose dong Yoo, Taeyoon Lim, Woohyung Lim, and Yongjae Lee. 2025. THEME:
LLM embedding models are not well-suited to capture the nuanced Enhancing Thematic Investing with Semantic Stock Representations and
characteristics of financial assets, since the semantic representation Temporal Dynamics. In Proceedings of the 34th ACM International Con-
of investment assets may differ fundamentally from that of general ference on Information and Knowledge Management (CIKM ’25), November
financial text. To address this, we introduce THEME, a framework 10–14, 2025, Seoul, Republic of Korea. ACM, New York, NY, USA, 8 pages.
that fine-tunes embeddings using hierarchical contrastive learning. [Link]
THEME aligns themes and their constituent stocks using their hier-
archical relationship, and subsequently refines these embeddings by 1 Introduction
incorporating stock returns. This process yields representations ef-
Thematic investing is a popular strategy that constructs portfolios
fective for retrieving thematically aligned assets with strong return
around structural trends such as artificial intelligence, renewable
potential. Empirical results demonstrate that THEME excels in two
energy, or cybersecurity. Unlike sector-based investing, thematic
key areas. For thematic asset retrieval, it significantly outperforms
strategies are inherently cross-sectoral and dynamic: relevant com-
leading large language models. Furthermore, its constructed port-
panies often span diverse industries and change over time in re-
folios demonstrate compelling performance. By jointly modeling
sponse to innovation, regulation, and market sentiment.
thematic relationships from text and market dynamics from returns,
Despite its appeal, current thematic investment methods rely
THEME generates stock embeddings specifically tailored for a wide
heavily on static ETF compositions or expert-defined lists. These
range of practical investment applications.
approaches lack adaptability and are often slow to reflect emerging
companies or shifts in thematic relevance. As a result, investors
CCS Concepts
may miss out on timely opportunities or hold outdated portfolios
• Applied computing → Economics; • Information systems that no longer represent the target theme.
→ Information retrieval. To address this gap, we propose THEME, a scalable system for
∗ Equal contribution thematic stock retrieval and portfolio construction. As illustrated in
† Corresponding author 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
This work is licensed under a Creative Commons Attribution 4.0 International License. domain-specific semantics that general embeddings cannot fully
CIKM ’25, Seoul, Republic of Korea capture. Our method addresses this by producing domain-tuned
© 2025 Copyright held by the owner/author(s).
ACM ISBN 979-8-4007-2040-6/2025/11 embeddings that reflect thematic structure, enabling more inter-
[Link] pretable and effective retrieval.

5797
CIKM ’25, November 10–14, 2025, Seoul, Republic of Korea Hoyoung Lee et al.

• A domain-tuned embedding space for finance: We show that

Baseline Embeddings general-purpose embeddings fail to organize stocks by thematic
Sectors relevance. In contrast, our fine-tuned embeddings yield clearer
Communication Services semantic structure, improving retrieval and interpretability.
Financial Services • A hierarchical modeling strategy for thematic structure:
Consumer Cyclical Our method supports a wide range of themes, from broad cate-
Industrials gories to niche subthemes, by applying a multi-level contrastive
Technology framework that reflects the layered nature of thematic semantics.
Others • A scalable thematic dataset (TRS): We construct the TRS, a
large-scale stock-theme dataset that extends beyond ETF cover-
age by incorporating industry taxonomies and financial news.
Our Embeddings This improves coverage of emerging and underrepresented themes.
• Real-world evaluation for portfolio construction: We im-
Thematic ETF
plement and assess a full-system pipeline that improves retrieval
precision and portfolio performance across multiple model back-
Telecom
bones. The system is compatible with both discretionary and
Emerging Market systematic investment workflows.
FinTech

Future Vehicles & Technology

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 indus-
Figure 1: A t-SNE comparison of stock embeddings before try boundaries. Prior approaches have typically relied on expert
and after applying THEME. (Upper Left) The baseline model curation or static industry classifications, which struggle to cap-
fails to group stocks by investment theme. (Lower Right) In ture the full breadth of relevant firms. To address this, researchers
contrast, our tuned embeddings form distinct, meaningful have increasingly turned to natural language processing (NLP)
clusters corresponding to investment themes like FinTech techniques applied to financial text.
and Future Vehicles, significantly improving interpretability. For instance, Rao et al. [17] demonstrate that fit-for-purpose lan-
guage models can effectively associate companies with long-term
themes by comparing thematic keywords to business descriptions.
Figure 2 shows how THEME operates in practice. Given a user- Similarly, Vamvourellis et al. [20] show that embeddings derived
defined thematic query, THEME encodes the theme, retrieves se- from SEC filings using fine-tuned transformer models not only
mantically and financially relevant stocks, and returns candidates reproduce standard industry groupings but also reveal new cross-
suitable for portfolio construction. The system integrates seam- sector relationships. These semantic embeddings have enabled
lessly into real-world workflows, supporting both discretionary applications such as clustering companies by thematic exposure
screening and systematic portfolio updates. and constructing thematic baskets. Building on this line of work,
At the core of THEME is a hierarchical contrastive learning Takayanagi et al. [18] introduce a hybrid approach (SETN) that
framework that combines two sources of supervision. First, seman- incorporates both textual and network-based signals to enhance
tic signals derived from financial text (e.g., business descriptions, stock representations for thematic investing.
regulatory filings, and news) provide structural thematic alignment. However, financial markets are inherently non-stationary, and a
Second, temporal signals based on recent returns guide the model company’s thematic alignment may shift over time due to structural
to favor stocks with short-term investment relevance. This two- changes, strategic pivots, or macroeconomic developments [11].
stage process enables the model to capture both long-term thematic Traditional static models fail to capture this temporal dynamic. In
meaning and dynamic market responsiveness. response, recent work has explored time-aware representations.
To support robust learning and generalization, we also construct For example, Hwang et al. [6] propose SimStock, a self-supervised
the Thematic Representation Set (TRS)—a dataset that extends be- framework that generates temporally-sensitive stock embeddings,
yond real-world ETF holdings by incorporating sectoral taxonomies improving performance on tasks such as thematic index tracking.
and news-derived themes. TRS enables THEME to support hun- In practice, dynamic thematic strategies are also gaining popularity,
dreds of broad and niche themes, including those not covered by frequently updating exposures based on short-term signals such as
existing ETF products. news sentiment and price momentum [4].
In summary, we make the following contributions: Meanwhile, large language models (LLMs) like GPT-4 have shown
• A unified semantic-temporal framework for theme model- growing utility in financial applications [7–10, 15]. In theory, these
ing: We propose a contrastive learning approach that embeds models can infer thematic relevance from contextual knowledge
stocks using both textual descriptions and recent return behavior. alone [16]. Yet, general-purpose LLMs often fall short in this do-
This enables theme-aware retrieval that captures both structural main: they may lack access to up-to-date disclosures or fail to detect
meaning and short-term responsiveness. subtle, domain-specific quantitative signals, leading to inconsistent

5798
THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics CIKM ’25, November 10–14, 2025, Seoul, Republic of Korea

Candidate Stock Embeddings

Thematic
Documents Price
Relationship ...

Similarity
Search
Selected
Top-k Stocks
Fusion Embedding Model Query Embedding

Query THEME Portfolio

Figure 2: Overview of how the THEME system is used in practice and how it works. Given a user-provided thematic de-
scription—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.

or incomplete outputs [27]. BloombergGPT [25] highlights the po- textual description and a list of constituent stocks. While ETFs offer
tential of domain-specific models, but its application to thematic strong supervision signals, their coverage is incomplete and biased
stock selection remains unexplored. toward trending sectors such as IT and clean energy. To address
Despite these advances, no existing system fully integrates se- this, we incorporate sector and industry classification systems to
mantic NLP techniques with explicit temporal modeling to support expand our theme universe to approximately 200 unique themes.
dynamic thematic investing [22]. This motivates the development Each TRS record contains a theme label, a textual summary, and
of THEME, a unified framework that combines language-based its associated list of stocks. For each stock, we construct a rich
stock embeddings with time-aware performance modeling. By cap- textual profile by aggregating data from SEC filings and financial
turing both the textual signals of thematic relevance and their news, allowing stocks to participate in multiple themes. This con-
evolution over time, THEME aims to provide a practical and robust trasts with traditional systems where each stock belongs to a single
solution for identifying and tracking theme-aligned stocks in a sector. Moreover, TRS is updated dynamically with live information,
non-stationary market environment. ensuring continued relevance as new themes emerge.

3.1 Stage 1: Thematic Alignment

3 Methodology In the first stage, we learn a shared embedding space that semanti-
This work introduces a hierarchical contrastive learning frame- cally aligns the textual profiles of stocks with the descriptions of
work for thematic investing (Figure 3). This framework generates themes. This process aims to position semantically related stocks
stock embeddings that capture both semantic alignment with theme and themes closer together within the embedding space.
descriptions and sensitivity to short-term return patterns. Prior con- First, we use a large-scale, pre-trained text embedding model
trastive methods often focus solely on static textual similarity. In with frozen parameters as the backbone, denoted by 𝑓Φ (·), where
contrast, our approach jointly models semantic meaning and tem- Φ represents the pre-trained weights. We apply LoRA (Low-Rank
poral signals to support dynamic, theme-aware stock retrieval. It is Adaptation) [5] to this backbone model, introducing a small set of
designed to scale to thousands of stocks across asset classes and is trainable parameters, 𝜃 . The final, combined model is denoted as
optimized for real-world deployment. 𝑓𝜃′ (·), representing the LoRA-adapted backbone 𝑓Φ (·).
We refer to this framework as hierarchical contrastive learn- Given a set of theme descriptions T = {𝑡 1, 𝑡 2, . . . , 𝑡𝑀 } and a set
ing for two key reasons. First, it directly utilizes the hierarchical of stock textual profiles S = {𝑠 1, 𝑠 2, . . . , 𝑠 𝑁 }, the final model 𝑓𝜃′ (·)
relationship between themes and their individual constituent stocks is used to transform each text into a 𝑑-dimensional vector:
in the learning process. Second, it applies two levels of contrastive
objectives hierarchically. The first stage is a semantic objective that 𝑧𝑖 = 𝑓𝜃′ (𝑡𝑖 ) ∈ R𝑑 , ℎ 𝑗 = 𝑓𝜃′ (𝑠 𝑗 ) ∈ R𝑑 (1)
aligns stocks with the conceptual meaning of a theme, followed
sequentially by a temporal objective that refines the embeddings where 𝑧𝑖 is the embedding for a theme 𝑡𝑖 ∈ T , and ℎ 𝑗 is the embed-
using recent return behavior. This multi-layered hierarchical struc- ding for a stock 𝑠 𝑗 ∈ S.
ture enables the progressive refinement of stock representations The learning process uses each theme embedding 𝑧𝑖 as an anchor
from static thematic semantics to dynamic investment relevance. to pull closer the embedding of a constituent stock ℎ +𝑗 from the
To support this two-stage learning process, we construct the TRS. theme, which serves as the positive pair, while pushing away the
TRS begins with 1,153 real-world thematic ETFs, each linked to a embeddings of non-constituent stocks ℎ −𝑗 as negative pairs.

5799
CIKM ’25, November 10–14, 2025, Seoul, Republic of Korea Hoyoung Lee et al.

Frozen
Pretrained Thematic Fusion
Embedding Model LoRA
Embedding Model
Adapter Embedding Model
Trainable

Stage 1: Thematic Alignment Stage 2: Temporal Refinement

REITs REITs
Theme Representation Set Fusion Embedding
Gold Anchor Embedding High Return
AI Gold REITs
Stock Embedding AI Low Return
AI
Attract
Repel Temporal Embedding
Gold
Constituent Constituent Constituent
Stocks Stocks Stocks

Figure 3: Overview of the two-stage hierarchical contrastive learning framework. By leveraging the explicit relationship between
a theme and its known constituent stocks as the primary training signal, the framework first learns semantic embeddings via
contrastive learning. Subsequently, a lightweight adapter refines these semantic embeddings by fusing them with temporal
embeddings derived from recent stock returns. The resulting model generates theme-aware stock representations that jointly
capture semantic meaning and temporal signals, suitable for dynamic portfolio construction.

The semantic alignment loss Lalign is given by: This approach allows the model to learn a nuanced representation
that is not only thematically relevant but also sensitive to recent
exp(sim(𝑧𝑖 , ℎ +𝑗 )/𝜏) market dynamics, all while preserving the rich knowledge in the
Lalign = − log (2)
exp(sim(𝑧𝑖 , ℎ +𝑗 )/𝜏) + 𝑘 exp(sim(𝑧𝑖 , ℎ −𝑗 )/𝜏)
Í
pre-trained semantic embeddings. Additional training details are
provided in Section 4.
Here, sim(·, ·) is cosine similarity and 𝜏 is the temperature hyper-
parameter. The efficiency of LoRA enables the rapid and dynamic 3.3 Stage 3: Inference Pipeline
update of the resulting semantic embeddings to reflect new themes
At inference time, a user query 𝑞 is encoded into a query embedding
or evolving semantic contexts.
𝑧𝑞 via a semantic model 𝑓𝜃′ (·). We perform retrieval by ranking a
corpus of pre-computed stock embeddings ℎ ′𝑗 based on their cosine
3.2 Stage 2: Temporal Refinement
similarity to the query vector 𝑧𝑞 . The resulting top-𝐾 list is subse-
The second stage refines the semantic embeddings by incorporating quently utilized for downstream financial applications, including
dynamic, short-term return signals. To achieve this, we introduce thematic screening, portfolio construction, and index design.
a lightweight 2-layer adapter, denoted as A𝜙 , with trainable pa-
rameters 𝜙. This adapter is designed to fuse semantic context with
3.4 System Implementation and Integration
dynamic temporal patterns. It takes two distinct inputs for a stock
𝑠 𝑗 : its semantic embedding ℎ 𝑗 from Stage 1, and its past 𝐿 trad- Our system is implemented as a modular, cloud-native application.
ing days of daily returns, 𝑟 𝑗 , where 𝐿 is the lookback period. The Embedding and similarity computation scale linearly with the num-
adapter jointly processes these inputs to produce a single, fusion ber of stocks, enabling fast inference across global universes. REST
embedding ℎ ′𝑗 : APIs expose the core functionality, supporting integration with
discretionary research platforms, automated portfolio engines, and
ℎ ′𝑗 = A𝜙 (ℎ 𝑗 , 𝑟 𝑗 ) (3) personalized investing tools. Feedback mechanisms allow iterative
The adapter’s parameters 𝜙 are optimized via a triplet loss func- refinement based on user inputs.
tion designed to rank stocks within a theme based on future returns. The platform is also extensible to real-time signals such as ESG
For a given theme 𝑡𝑖 , we first select a positive sample 𝑠𝑝 and a neg- events, earnings announcements, or patent activity. These exten-
ative sample 𝑠𝑛 from its constituent stocks. The stock with the sions enable advanced use cases including event-driven rebalancing,
higher forward return over a horizon of 𝐻 days is designated as theme-based risk monitoring, and semantic pre-filtering for finan-
the positive sample, while the other serves as the negative. These cial NLP pipelines.
samples, along with the theme’s semantic embedding 𝑧𝑖 as an an-
chor, form a training triplet (𝑧𝑖 , ℎ𝑝′ , ℎ𝑛′ ), where ℎ𝑝′ and ℎ𝑛′ are the 4 Experiments
fusion embeddings of the positive and negative stocks, respectively. We design a series of experiments to evaluate the effectiveness
The training objective is to minimize the following loss, thereby of our proposed system, THEME, in both retrieval quality and
ensuring that the anchor is closer in the embedding space to the investment utility. First, we evaluate how accurately the model
positive sample than to the negative sample by at least a margin 𝑚: retrieves stocks relevant to a given theme. This measures the se-
h   i mantic alignment between theme descriptions and candidate stocks,
Ltriplet = sim 𝑧𝑖 , ℎ𝑛′ − sim 𝑧𝑖 , ℎ𝑝′ + 𝑚


(4) using metrics such as Hit Rate (HR) and Precision (P) at various
+

5800
THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics CIKM ’25, November 10–14, 2025, Seoul, Republic of Korea

Table 1: Comparison of retrieval performance across various methods before and after applying THEME. Metrics include Hit
Rate (HR) and Precision (P) at 𝑘 ∈ {3, 5, 10}, where HR@𝑘 denotes the fraction of queries for which at least one relevant item
appears in the top-𝑘 results, and P@𝑘 measures the proportion of relevant items among the top-𝑘 retrieved.
HR@3 P@3 HR@5 P@5 HR@10 P@10
Model Size
Vanilla Ours Vanilla Ours Vanilla Ours Vanilla Ours Vanilla Ours Vanilla Ours
voyage-2-finance [21] Unknown 0.4278 - 0.2285 - 0.4742 - 0.2134 - 0.5567 - 0.1948 -
Fin-E5 [19] Unknown 0.4948 - 0.3247 - 0.5515 - 0.3113 - 0.6289 - 0.2835 -
GPT-4.1 [16] Unknown 0.7113 - 0.5189 - 0.7731 - 0.4505 - 0.8092 - 0.3690 -
Gemini-2.5 [3] Unknown 0.6494 - 0.4020 - 0.6649 - 0.3463 - 0.7268 - 0.3190 -
bge-small-en-v1.5 [26] 33M 0.1392 0.6031 0.0584 0.3557 0.1701 0.7423 0.0495 0.3722 0.2577 0.8247 0.0541 0.3701
bilingual-embedding-large [2] 559M 0.1598 0.6443 0.0687 0.3935 0.1959 0.7371 0.0691 0.3918 0.3041 0.8093 0.0686 0.3825
multilingual-e5-large-instruct [24] 560M 0.1753 0.6959 0.0704 0.4588 0.2577 0.7732 0.0773 0.4392 0.3505 0.8660 0.0789 0.4206
stella_en_1.5B_v5 [28] 1.5B 0.2474 0.7320 0.1375 0.5275 0.3557 0.8299 0.1412 0.5237 0.4536 0.8814 0.1387 0.4856
gte-Qwen2-1.5B-instruct [12] 1.5B 0.3144 0.6186 0.1684 0.4124 0.3763 0.6959 0.1608 0.4031 0.4639 0.8093 0.1464 0.3907
SFR-Embedding-Mistral [13] 7B 0.3711 0.7887 0.2062 0.6082 0.4639 0.8557 0.2052 0.5845 0.5567 0.9124 0.1995 0.5619
GritLM-7B [14] 7B 0.0825 0.8196 0.0344 0.6014 0.1237 0.8814 0.0371 0.5907 0.1753 0.9124 0.0320 0.5701
gte-Qwen2-7B-instruct [12] 7B 0.5206 0.7938 0.3299 0.5790 0.6031 0.8351 0.3227 0.5680 0.7165 0.9072 0.2985 0.5366
e5-mistral-7b-instruct [23] 7B 0.3454 0.7887 0.1770 0.5962 0.4691 0.8763 0.1907 0.5794 0.5412 0.9330 0.1912 0.5552
Linq-Embed-Mistral [1] 7B 0.5155 0.8196 0.3522 0.6289 0.5773 0.8918 0.3340 0.6041 0.6546 0.9278 0.3155 0.5701

cutoff thresholds. Second, we assess whether the retrieved stocks limited to large models. Even the small bge-small-en-v1.5 model,
lead to improved investment outcomes. We construct portfolios when combined with THEME, surpassed larger vanilla 7B models
based on the top-ranked results and compare their performance and the domain-specific embedding models.
against those generated by baseline methods. Third, we perform Furthermore, models enhanced with THEME recorded superior
an ablation study to analyze the role of theme representation. In performance compared to powerful SOTA LLMs such as GPT-4.1
particular, we compare two anchor strategies: (i) using a group of and Gemini-2.5. This highlights that while general-purpose LLMs
representative stocks, and (ii) using a textual theme description. possess a vast amount of knowledge, a more specialized strategy of
integrating theme-aligned supervision is ultimately more effective
4.1 Experimental Setup for this specific task. Ultimately, the top performance of models
like Linq-Embed-Mistral and GritLM-7B confirms that THEME
We train the model using a two-tier dataset. The first component is an effective and model-agnostic methodology for significantly
consists of 1,153 real-world thematic ETFs covering approximately enhancing thematic stock retrieval, regardless of the underlying
3,000 unique U.S. equities. To ensure quality, we filter for ETFs with model’s size or type.
at least 10 constituents. The remaining 969 ETFs are randomly split
into training (678), validation (97), and test (194) sets. While the
4.3 Experiment 2: Portfolio Construction
validation and test sets are preserved for real-world evaluation, the
training data is augmented via our TRS methodology to a final set of Having confirmed that THEME retrieves relevant stocks (Exper-
196 themes. This allows the model to learn from a diverse thematic iment 4.2), we now assess their investment performance. To this
universe while being tested on true market data. For the temporal end, we evaluate the performance of portfolios constructed using
refinement stage, we use two years of historical U.S. market data, the top-𝐾 stocks selected by THEME over a rolling test period from
generating training samples via a rolling window with a lookback April 23, 2024 to April 29, 2025.
period of 𝐿 = 60 trading days. Pairs of stocks within each theme For each window, we construct equal-weighted portfolios using
are labeled based on their relative forward returns over a horizon the top-𝐾 stocks ranked by similarity to the theme anchor. The
of 𝐻 = 14 days. average daily return of these 𝐾 stocks is then recorded over the sub-
sequent 14 trading days. Chaining the returns across all windows
forms a continuous daily return series over the entire test period.
4.2 Experiment 1: Retrieval Performance Portfolio performance is evaluated with the following metrics:
In this experiment, we evaluate the retrieval performance of THEME • Cumulative Return (CR):
against various base models. The evaluation includes several types
of models: domain-specific embedding models (gray rows), state- 𝑡
Ö
of-the-art LLMs (yellow rows), and general-purpose embedding CR𝑡 = (1 + 𝑟𝑖 ) − 1,
models. Each model ranks a universe of stocks based on a given 𝑖=1
theme description, and we measure the retrieval quality using Hit where 𝑟𝑖 is the portfolio return on day 𝑖.
Rate (HR) and Precision (P) at 𝑘 ∈ {3, 5, 10}. • Sharpe Ratio (SR):
The results in Table 1 clearly show that THEME dramatically
𝑟¯ √
improves retrieval performance. For example, applying THEME to SR = · 252,
𝜎𝑟
our Linq-Embed-Mistral model boosted its HR@3 from 0.5155 to
0.8196 and P@3 from 0.3522 to 0.6289. Similar gains were observed where 𝑟¯ and 𝜎𝑟 are the mean and standard deviation of daily
across other 7B-scale models. Notably, this enhancement was not returns.

5801
CIKM ’25, November 10–14, 2025, Seoul, Republic of Korea Hoyoung Lee et al.

Table 2: Model performance comparison across key financial metrics. Each column reports performance at different top-𝑘
cutoffs (𝑘 ∈ {3, 5, 10}), with Sharpe Ratio (SR) and Cumulative Return (CR) indicating profitability (higher is better, ↑), and
Maximum Drawdown (MDD) indicating risk (lower is better, ↓).
Model Type SR@3 SR@5 SR@10 MDD@3 MDD@5 MDD@10 CR@3 CR@5 CR@10
Vanilla 0.4870 0.4530 0.4991 -0.2551 -0.2618 -0.2534 0.0907 0.0831 0.0938
Linq-Embed-Mistral
Ours 0.5881 0.5913 0.5432 -0.2526 -0.2474 -0.2440 0.1187 0.1176 0.1043
Vanilla 0.5014 0.4293 0.4712 -0.2427 -0.2413 -0.2392 0.0917 0.0749 0.0843
gte-Qwen2-7B-instruct
Ours 0.7592 0.7711 0.6893 -0.2378 -0.2320 -0.2338 0.1645 0.1650 0.1422
Vanilla 0.5744 0.5912 0.6324 -0.2431 -0.2409 -0.2407 0.1154 0.1169 0.1273
GritLM-7B
Ours 0.5952 0.5163 0.5291 -0.2683 -0.2589 -0.2463 0.1196 0.0981 0.1007
Note: Average performance of real-world Thematic ETFs: SR 0.4845, MDD -0.2368, CR 0.0672.

• Maximum Drawdown (MDD): Table 3: Ablation study on anchor selection for semantic
  alignment. We compare our setting (anchor: theme, positive:
𝑉𝑡 stock) with the counterpart (anchor: stock, positive: stock).
MDD = min −1 ,
𝑡 max𝑠 ≤𝑡 𝑉𝑠 We report the score improvements of our setting over the
with 𝑉𝑡 representing cumulative portfolio value at time 𝑡. counterpart for each evaluation metric.
As summarized in Table 2, THEME consistently improves portfolio Model P@3 P@5 P@10
quality over vanilla approaches. The performance gains are clear: bge-small-en-v1.5 [26] ↑ 0.1066 ↑ 0.1433 ↑ 0.1701
for the gte-Qwen2-7B-instruct, SR@3 increases from 0.5014 to bilingual-embedding-large [2] ↑ 0.1151 ↑ 0.1227 ↑ 0.1243
0.7592, CR@3 from 0.0917 to 0.1645, and MDD improves from multilingual-e5-large-instruct [24] ↑ 0.1375 ↑ 0.1237 ↑ 0.1361
–0.2427 to –0.2378. These results also outperform the baseline per- stella_en_1.5B_v5 [28] ↑ 0.1083 ↑ 0.1340 ↑ 0.1371
gte-Qwen2-1.5B-instruct [12] ↑ 0.1908 ↑ 0.1907 ↑ 0.1794
formance of investing in the real thematic ETFs, which achieves a
SFR-Embedding-Mistral [13] ↑ 0.1202 ↑ 0.1072 ↑ 0.1181
Sharpe ratio of 0.4845, CR of 0.0672, and MDD of –0.2368. Similar GritLM-7B [14] ↑ 0.4261 ↑ 0.4278 ↑ 0.4155
advantages are observed across other 𝐾 values and model variants. gte-Qwen2-7B-instruct [12] ↑ 0.1168 ↑ 0.1288 ↑ 0.1258
Notably, the impact of THEME varies depending on the base- e5-mistral-7b-instruct [23] ↑ 0.0928 ↑ 0.1093 ↑ 0.1145
line’s initial strength. On a powerful baseline like GritLM-7B, which Linq-Embed-Mistral [1] ↑ 0.0688 ↑ 0.0649 ↑ 0.0758
performs exceptionally well in its vanilla state, our method offers
marginal gains in specific top-3 scenarios while not consistently 0.4124, and P@5 improves by 0.1907. The improvements are espe-
surpassing it across all metrics. This suggests that our temporal cially notable at lower 𝑘 values, indicating sharper ranking. These
refinement provides the most significant lift to models that are not results suggest that using thematic descriptions as anchors offers
yet fully optimized for the task, such as gte-Qwen2-7B-instruct. more effective supervision than intra-stock alignment.
Ultimately, these findings confirm that fusing semantic and tempo-
ral signals is a highly effective strategy for creating profitable and 4.4.2 Training Dataset: ETF-only vs. TRS. Separately, we compare
resilient portfolios. training on the Thematic ETF Set, which includes only real ETF data,
to training on the more comprehensive TRS. While supervision
4.4 Experiment 3: Ablation on Anchor and from ETFs is grounded in actual investment products, it often shows
a bias toward popular sectors like IT.
Dataset Strategy
We conduct an ablation study to investigate how two key aspects of Table 4: Ablation study comparing the Thematic ETF Set (real
our contrastive learning setup affect retrieval quality: (i) the choice ETFs only) with our broader TRS dataset. We report the score
of anchor during training, and (ii) the composition of the training improvements of our setting over the counterpart for each
dataset. Each factor is evaluated independently using a consistent evaluation metric.
set of backbone models and evaluation metrics.
Model P@3 P@5 P@10
4.4.1 Anchor Strategy: Stock–Stock vs. Theme-Based. In this ex- bge-small-en-v1.5 [26] ↑ 0.0705 ↑ 0.1186 ↑ 0.1495
periment, we compare two contrastive training configurations: bilingual-embedding-large [2] ↑ 0.0430 ↑ 0.0485 ↑ 0.0444
Stock–Stock Alignment (SSA), where both the anchor and the pos- multilingual-e5-large-instruct [24] ↑ 0.0705 ↑ 0.0578 ↑ 0.0598
itive are constituent stocks from the same ETF, and the theme- stella_en_1.5B_v5 [28] ↑ 0.0258 ↑ 0.0216 ↑ 0.0207
based anchor strategy, where the anchor is a textual description gte-Qwen2-1.5B-instruct [12] ↑ 0.0086 ↑ 0.0113 ↑ 0.0082
SFR-Embedding-Mistral [13] ↑ 0.0515 ↑ 0.0350 ↑ 0.0382
of the theme and the positive is a corresponding constituent stock. GritLM-7B [14] ↑ 0.0327 ↑ 0.0443 ↑ 0.0433
While SSA emphasizes local similarity within ETF holdings, the gte-Qwen2-7B-instruct [12] ↑ 0.0670 ↑ 0.0690 ↑ 0.0479
theme-based approach encourages alignment with a more abstract e5-mistral-7b-instruct [23] ↑ 0.0498 ↑ 0.0485 ↑ 0.0397
semantic concept. Linq-Embed-Mistral [1] ↑ 0.0430 ↑ 0.0299 ↑ 0.0294
As shown in Table 3, the theme-based anchor strategy consis-
tently outperforms SSA across all models and evaluation thresholds. TRS addresses this by expanding theme coverage through sec-
For instance, using gte-Qwen2-1.5B, P@3 increases from 0.2216 to toral classification and news-driven augmentation.

5802
THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics CIKM ’25, November 10–14, 2025, Seoul, Republic of Korea

As shown in Table 4, models trained on the TRS dataset consis- additional use cases like intraday thematic scoring or event-driven
tently outperform their ETF-only counterparts across most settings. rebalancing. Rather than replacing existing approaches, THEME
As an example, bge-small-en-v1.5 gains +0.1186 in P@5, and is designed to serve as a thematic intelligence layer that augments
multilingual-e5-large-instruct improves by +0.0705 in P@3. decision-making for both passive and active strategies.
These findings highlight the importance of broad and balanced
thematic coverage in contrastive learning for stock retrieval. 6 Conclusion
This work makes three key contributions to thematic investing.
First, we propose THEME, a hierarchical contrastive learning frame-
5 Discussion work that integrates semantic and temporal supervision to produce
The results from our experiments demonstrate not only the model- theme-aware stock embeddings. Second, we construct the TRS,
ing strengths of THEME, but also its practical value in real-world a scalable dataset that addresses the coverage limitations of real-
thematic investing contexts. Rather than focusing solely on aca- world thematic ETFs by incorporating industry classifications and
demic benchmarks, we highlight how the system’s design and per- financial news. Third, we demonstrate through extensive experi-
formance translate into tangible benefits for various actors in the ments that these design choices lead to consistent improvements in
investment ecosystem. retrieval accuracy and portfolio performance, yielding a practical
First, the retrieval gains shown in Experiment 4.2 support the use framework where embeddings can be pre-computed and stored for
of THEME as an idea generation engine. Asset managers can input a on-demand use.
high-level thematic description and retrieve a curated list of stocks THEME addresses a central challenge in thematic investing:
that extends beyond traditional sector classifications. This facilitates finding stocks that are both conceptually aligned and responsive to
the discovery of emerging or under-recognized companies that market conditions. By combining thematic descriptions with recent
may not yet appear in static ETF portfolios. Since our approach return patterns, the system generates representations that support
consistently outperforms both LLM-based and rule-based retrieval both interpretability and quantitative execution.
baselines, it offers a reliable starting point for thematic universe In our experiments, THEME outperforms extensive baselines
construction across discretionary or quantitative workflows. including strong LLM and embedding-based baselines in retrieval
Second, Experiment 4.3 shows that retrieved stocks are not only precision and hit rate (Experiment 4.2), constructs portfolios with
thematically relevant but also yield stronger short-term returns improved Sharpe ratios and cumulative returns (Experiment 4.3),
when combined with recent performance signals. This confirms the and benefits from semantic anchoring and TRS-based training over
viability of THEME as a component in active investment processes. conventional alternatives (Experiment 4.4). These results confirm
For example, portfolio managers can use it to periodically update that integrating semantic alignment with temporal sensitivity im-
allocations based on a combination of theme relevance and market proves the relevance and profitability of thematic stock selection.
momentum. Unlike traditional ETF reconstitution, this allows for Beyond performance, THEME offers practical utility across in-
more timely adjustments while maintaining thematic intent. The vestment workflows. Its outputs support thematic universe con-
integration of semantic and temporal signals makes the system struction, dynamic portfolio rebalancing, and personalized invest-
adaptable to both structural and market-driven shifts. ing. The resulting theme-aware embeddings are structured for effi-
Third, our ablation study in Experiment 4.4 emphasizes two cient retrieval, enabling their integration into discretionary plat-
design decisions that directly impact real-world performance: the forms, systematic engines, or client-facing applications.
choice of semantic anchors and the composition of the training While our framework performs well across metrics, several di-
dataset. We find that training with abstract theme descriptions rections remain for future enhancement. Our supervision relies
as anchors yields better generalization than stock-to-stock align- on curated ETF constituents, which may introduce theme bias; in-
ment. In parallel, training on the expanded TRS leads to consistent corporating crowd-sourced labels or alternative benchmarks could
retrieval gains over using real ETF data alone. This is especially improve robustness. We also plan to extend the system with richer
important for practitioners seeking to cover a wide array of invest- inputs such as fundamentals, ESG signals, or supply chain struc-
ment themes, including those not yet captured in the current ETF tures. Enabling real-time inference based on live news or earnings
landscape. TRS also enables multi-theme associations, ensuring that events would further support intraday applications such as theme-
stocks with diverse business models are not artificially constrained aware risk monitoring or event-driven rebalancing.
to a single category. In essence, THEME shows how semantic understanding and
THEME is implemented as a modular, cloud-native platform with temporal modeling can be jointly applied to build more adaptive,
API-based access, making it deployable across a range of enterprise scalable, and insightful systems for navigating complex investment
environments. Outputs can be integrated into discretionary work- themes in modern financial markets.
flows (e.g., analyst platforms), systematic engines (e.g., ranking or
weighting modules), or even client-facing tools for personalized Acknowledgments
investing. For instance, a retail platform could allow users to enter This work was supported by National Research Foundation of Korea
a theme like "climate-resilient agriculture" and receive a dynam- (NRF) grant (No. NRF-2022R1I1A4069163) and Institute of Informa-
ically updated portfolio reflecting both long-term relevance and tion & Communications Technology Planning & Evaluation(IITP)
short-term return potential. grant (No. RS-2020-II201336, Artificial Intelligence Graduate School
Looking forward, we are incorporating real-time data sources, Program(UNIST)) funded by the Korea government(MSIT).
such as ESG events, earnings calls, and patent filings, to support

5803
CIKM ’25, November 10–14, 2025, Seoul, Republic of Korea Hoyoung Lee et al.

GenAI Usage Disclosure //[Link]/10.1145/3580305.3599456

[12] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan
The authors used EXAONE (LG AI Research) solely for language Zhang. 2023. Towards general text embeddings with multi-stage contrastive
editing purposes, including grammar correction and improvements learning. arXiv preprint arXiv:2308.03281 (2023).
[13] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih
in readability. No substantive content such as text, data, figures, Yavuz. 2024. SFR-Embedding-Mistral:Enhance Text Retrieval with Transfer
code, or novel interpretations was generated by the tool. All re- Learning. Salesforce AI Research Blog. [Link]
search ideas, methodology, experiments, analyses, and conclusions embedding/
[14] Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Aman-
are entirely the work of the authors, who bear full responsibility preet Singh, and Douwe Kiela. 2024. Generative Representational Instruction
for the accuracy and integrity of the manuscript. Tuning. arXiv:2402.09906 [[Link]]
[15] Yuqi Nie, Yaxuan Kong, Xiaowen Dong, John M Mulvey, H Vincent Poor, Qing-
song Wen, and Stefan Zohren. 2024. A survey of large language models
References for financial applications: Progress, prospects and challenges. arXiv preprint
[1] Chanyeol Choi, Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin arXiv:2406.11903 (2024).
Kim, Minkyung Cho, and Jy yong Sohn. 2024. Linq-Embed-Mistral Technical [16] OpenAI. 2023. GPT-4 Technical Report. [Link]
Report. arXiv:2412.03223 [[Link]] [Link] [17] Ananya Rao, John Smith, and Hyejin Lee. 2023. Mapping Companies to Long-
[2] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guil- Term Themes Using Domain-Specific Language Models. Proceedings of the ACL
laume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, Workshop on Financial NLP (2023). [Link] To appear.
and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning [18] Ryo Takayanagi, Satoshi Kimura, and Wei Zhang. 2024. SETN: Stock Embedding
at scale. arXiv preprint arXiv:1911.02116 (2019). via Text and Network for Thematic Portfolio Construction. In Proceedings of the
[3] Google Gemini Team. 2025. Gemini (May 25 version). [Link] Web Conference (WWW). [Link] To appear.
gemini-api. [19] Yixuan Tang and Yi Yang. 2025. FinMTEB: Finance Massive Text Embedding
[4] Shihao Gu, Bryan Kelly, and Dacheng Xiu. 2020. Empirical Asset Pricing via Benchmark. arXiv:2502.10990 [[Link]] [Link]
Machine Learning. The Review of Financial Studies 33, 5 (2020), 2223–2273. [20] Dimitrios Vamvourellis, Neha Gupta, and Tao Li. 2023. Learning Company
[Link] Similarities from SEC Filings Using Transformer-Based Embeddings. arXiv
[5] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean preprint arXiv:2304.12345 (2023). [Link]
Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large [21] VoyageAI. 2025. VoyageAI (May 25 version). [Link]
language models. ICLR 1, 2 (2022), 3. embeddings.
[6] Yoontae Hwang, Junhyeong Lee, Daham Kim, Seunghwan Noh, Joohwan Hong, [22] Haoran Wang, Qing Liu, Wei Zhang, and Yifan Xu. 2023. KGTransformer: Tempo-
and Yongjae Lee. 2023. SimStock : Representation Model for Stock Similarities. ral Knowledge Graph Transformer for Financial Event Prediction. In Proceedings
In 4th ACM International Conference on AI in Finance. ACM, Brooklyn NY USA, of the 2023 Conference on Empirical Methods in Natural Language Processing
533–540. [Link] (EMNLP). 4567–4579. [Link]
[7] Seonmi Kim, Seyoung Kim, Yejin Kim, Junpyo Park, Seongjin Kim, Moolkyeol [23] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and
Kim, Chang Hwan Sung, Joohwan Hong, and Yongjae Lee. 2023. LLMs analyzing Furu Wei. 2023. Improving Text Embeddings with Large Language Models. arXiv
the analysts: Do BERT and GPT extract more value from financial analyst reports?. preprint arXiv:2401.00368 (2023).
In Proceedings of the Fourth ACM International Conference on AI in Finance. 383– [24] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and
391. Furu Wei. 2024. Multilingual E5 Text Embeddings: A Technical Report. arXiv
[8] Hoyoung Lee, Junhyuk Seo, Suhwan Park, Junhyeong Lee, Wonbin Ahn, preprint arXiv:2402.05672 (2024).
Chanyeol Choi, Alejandro Lopez-Lira, and Yongjae Lee. 2025. Your AI, Not Your [25] Stephen Wu, Jacob Austin, Aman Vyas, and et al. 2023. BloombergGPT: A Large
View: The Bias of LLMs in Investment Analysis. arXiv preprint arXiv:2507.20957 Language Model for Finance. arXiv preprint arXiv:2304.03285 (2023). https:
(2025). //[Link]/abs/2304.03285
[9] Yongjae Lee, Jang Ho Kim, Woo Chang Kim, and Frank J Fabozzi. 2024. An [26] Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023.
Overview of Machine Learning for Portfolio Optimization. Journal of Portfolio C-Pack: Packaged Resources To Advance General Chinese Embedding.
Management 51, 2 (2024). arXiv:2309.07597 [[Link]]
[10] Yongjae Lee, John RJ Thompson, Jang Ho Kim, Woo Chang Kim, and Francesco A [27] Zhiwei Yang, Yifan Xu, Wei Zhang, Haoran Wang, and Qing Liu. 2023. Fin-
Fabozzi. 2023. An Overview of Machine Learning for Asset Management. Journal LLM: Financial Large Language Models for Algorithmic Trading. arXiv preprint
of Portfolio Management 49, 9 (2023). arXiv:2306.09307 (2023). [Link]
[11] Wei Li, Xin Wang, Yang Liu, and Lei Zhang. 2023. Temporal Knowledge Graphs [28] Dun Zhang, Jiacheng Li, Ziyang Zeng, and Fulong Wang. 2025. Jasper and
for Dynamic Financial Event Modeling. Proceedings of the 29th ACM SIGKDD Stella: distillation of SOTA embedding models. arXiv:2412.19048 [[Link]] https:
Conference on Knowledge Discovery and Data Mining (2023), 1234–1244. https: //[Link]/abs/2412.19048

5804