Paper: arXiv 2308.08031
Authors: Dimitrios Vamvourellis, Máté Toth, Snigdha Bhagat, Dhruv Desai, Dhagash Mehta, Stefano Pasquali
Abstract
Identifying companies with similar profiles is a core task in finance with a wide range of applications in portfolio construction, asset pricing and risk attribution. When a rigorous definition of similarity is lacking, financial analysts usually resort to ’traditional’ industry classifications such as Global Industry Classification System (GICS) which assign a unique category to each company at different levels of granularity. Due to their discrete nature, though, GICS classifications do not allow for ranking companies in terms of similarity. In this paper, we explore the ability of pre-trained and finetuned large language models (LLMs) to learn company embeddings based on the business descriptions reported in SEC filings. We show that we can reproduce GICS classifications using the embeddings as features. We also benchmark these embeddings on various machine learning and financial metrics and conclude that the companies that are similar according to the embeddings are also similar in terms of financial performance metrics including return correlation.
Complexity vs Empirical Score
- Math Complexity: 6.5/10
- Empirical Rigor: 8.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced NLP and ML techniques (BERT embeddings, fine-tuning) with moderate mathematical complexity, while demonstrating strong empirical rigor through rigorous backtesting on SEC filings, benchmarking against GICS, and validation with financial metrics like return correlation.
Research Flowchart
flowchart TD
A["Research Goal<br>Model Company Similarity<br>using LLMs & SEC Filings"] --> B["Data Source<br>SEC Business Descriptions"]
B --> C["Methodology<br>Pre-train & Fine-tune LLMs"]
C --> D["Process<br>Generate Company Embeddings"]
D --> E{"Evaluation"}
E --> F["Reproduce GICS Classifications"]
E --> G["Benchmark vs Financial Metrics<br>e.g., Return Correlation"]
F & G --> H["Findings<br>Embeddings capture similarity<br>validating finance utility"]
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 generating company embeddings from SEC 10-K business descriptions using state‑of‑the‑art language models and benchmarking them on classification and financial similarity tasks.
Source sentence
In this paper, we generate company embeddings using the raw business descriptions reported in SEC 10K filings using SOTA language models.
Assumes 3
- Similarity between companies is measured by the correlation of their daily returns, assuming similar companies have similar return movements over long horizons. market
Source sentence
we measure similarity between two companies based on the correlation of their daily returns, assuming that similar companies typically experience similar movements in their stock returns over long horizons.
- Only the first 512 tokens of Item 1 are used for finetuning the BERT model. data
Source sentence
we finetune it on the task of predicting GICS industry, using only the first 512 tokens of Item 1.
- A stratified train‑test split ensures a representative sample from all GICS categories. statistical
Source sentence
we opt for a stratified split, ensuring a representative sample from all categories in both train and test splits.
Data
- SEC 10‑K Item 1 business description text-news public
Source sentence
we used the textual content from Item 1 of the SEC 10K reports as input features.
Costs
Not applicable
Finds 4
- Company embeddings can reproduce GICS sector classifications with approximately 90% accuracy, with finetuned SBERT achieving the best performance. predictive-accuracy
Source sentence
Finetuned Sentence-BERT (SBERT-FT) models preform best, achieving approximately 90% accuracy in predicting GICS sectors from 10-K business descriptions.
- Finetuned models outperform pre‑trained models on sector/industry classification but underperform on similarity and return‑attribution tasks. performance
Source sentence
finetuned models underperform compared to the pre‑trained models on the similarity and return attribution tasks, while they outperformed pre‑trained models on sector/industry classification tasks.
- Using embeddings, the closest peer of each company shows a 10% higher return correlation than the sector average and 6% higher than the industry average. performance
Source sentence
using the embeddings we can identify the closest peer of each company in the embedding space which has 10% greater returns correlation compared to the sector average and 6% greater returns correlation compared to the industry average.
- Larger LLMs (≈200‑500B parameters) marginally outperform SBERT on return attribution but are comparable or slightly worse on similarity tasks. performance
Source sentence
PaLM and GPT marginally outperform SBERT on the return attribution task while they are on par on slightly worse than SBERT on the similarity task.
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.
11 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.