FinTexTS: Financial Text-Paired Time-Series Dataset via Semantic-Based and Multi-Level Pairing

The paper introduces FinTexTS, a large-scale financial text-paired time-series dataset constructed via a novel semantic-based and multi-level pairing framework that overcomes the limitations of simple keyword matching by leveraging LLMs to align news articles with stock prices across macro, sector, related company, and target-company levels, thereby significantly improving stock price forecasting performance.

The Gist

Imagine you are trying to predict the weather for a specific city, say, Seattle.

If you only look at the thermometer inside your house (the stock price), you might get a rough idea. But to make a great prediction, you need more than just the temperature. You need to know:

• Is a massive storm front moving in from the Pacific? (Macro-level)
• Is the whole region experiencing a heatwave? (Sector-level)
• Did the neighbor's sprinkler system break and flood the street? (Related Company)
• Did your own roof start leaking? (Target Company)

For a long time, computers trying to predict stock prices were like weather forecasters who only looked at the thermometer inside the house and ignored everything else. They used a very simple trick: they searched for the company's name (like "NVIDIA") in the news. If the name wasn't in the article, they ignored it.

The Problem: This is like ignoring a news report that says, "The entire semiconductor industry is facing a chip shortage," because it didn't explicitly say "NVIDIA." Even though NVIDIA wasn't named, the news is huge for them!

The Solution: FinTexTS
The researchers in this paper built a new system called FinTexTS. Think of it as upgrading from a simple thermometer to a super-smart, all-seeing weather satellite.

Here is how they did it, using simple analogies:

1. The "Context" Detective (Semantic Pairing)

Instead of just searching for a company's name like a keyword search engine, the system reads the company's official "report card" (called SEC filings).

• The Old Way: "Did the word 'NVIDIA' appear in this article?"
• The New Way: The system reads the report card, learns that NVIDIA makes chips for AI and data centers. Then, it scans the news for articles about "AI chip shortages" or "data center construction." Even if the word "NVIDIA" isn't there, the system says, "Hey, this article is basically about NVIDIA!"

It's like a detective who knows that if someone is buying a lot of flour and eggs, they are probably baking a cake, even if the word "cake" isn't mentioned.

2. The "Four-Layer" Onion (Multi-Level Pairing)

The researchers realized that stock prices are influenced by things happening at different distances. They organized the news into four layers, like an onion:

• Layer 1: The Macro Level (The Global Weather): This is the big picture. Is the Federal Reserve raising interest rates? Is there a war? These affect everyone.
• Layer 2: The Sector Level (The Neighborhood): This is about the specific industry. Is the whole tech sector booming? Is the energy sector crashing?
• Layer 3: The Related Company Level (The Neighbors): Did a competitor (like AMD) just release a new product? Did a key partner (like Microsoft) announce a deal? This affects the target company indirectly.
• Layer 4: The Target Company Level (Your House): This is the specific news about the company itself. Did they fire their CEO? Did they release a new product?

By feeding the computer all four layers of information, it gets a complete picture of the "weather" surrounding that stock.

3. The "Smart Summarizer" (LLMs)

Sometimes, there are too many news articles. Reading 50 articles about a single day is overwhelming. The system uses a "Smart Summarizer" (an AI brain) to read all those articles and write a short, punchy summary of the most important events for each layer. It filters out the noise and keeps the signal.

The Result: A Better Crystal Ball

The researchers tested this new system against the old "keyword search" method.

• The Old Method: Often missed important news or included irrelevant junk.
• The New Method (FinTexTS): Consistently predicted stock prices more accurately.

They even tested it with a "premium" news source (like a paid subscription service) versus free news, and found that the better the quality of the news, the better the predictions became.

In a Nutshell

This paper is about building a smarter way to connect news stories to stock prices. Instead of just matching words, it understands meaning and context. It looks at the big picture, the industry trends, the neighbors, and the company itself to give a much more accurate forecast of where a stock is heading.

It's the difference between guessing the weather by looking at a single puddle and using a satellite to see the entire storm system.

Technical Summary

Here is a detailed technical summary of the paper "FinTexTS: Financial Text-Paired Time-Series Dataset via Semantic-Based and Multi-Level Pairing."

1. Problem Statement

The financial domain relies heavily on time-series analysis, particularly for stock price forecasting. While recent advancements in multimodal learning (combining text and numerical data) show promise, existing methods for constructing text-paired time-series datasets in finance suffer from critical limitations:

• Limitations of Keyword-Based Matching: Current approaches (e.g., FNSPID) rely on simple keyword matching to pair news articles with stock prices. This fails to capture:
  • Semantic Relevance: Articles relevant to a company (e.g., discussing GPU data centers) that do not explicitly mention the company name (e.g., NVIDIA) are missed.
  • Noise: Articles containing keywords but lacking actual relevance are included.
• Ignoring Multi-Level Dynamics: Financial markets are characterized by complex interdependencies. A stock price is influenced not just by company-specific events, but also by sector-level trends, related company events (competitors/partners), and macroeconomic factors. Simple keyword matching cannot distinguish or hierarchically organize these multi-level influences.

2. Methodology

The authors propose a Semantic-Based and Multi-Level Pairing Framework to construct the FinTexTS dataset. The framework consists of four main components:

A. LLM-Based SEC Filing Parsing

• Goal: Extract standardized, company-specific context to serve as the "ground truth" for semantic alignment.
• Process: An LLM parses diverse SEC filing types (10-K, 10-Q, 8-K, etc.) to extract information into five categories: Overview & Products, Strategy/Markets/Operations, Governance & Risks, Financial Information, and Recent Events & Catalysts.
• Handling: Since filings are not daily, the extracted content is forward-filled to create daily paired data.

B. LLM-Based News Classification

• Goal: Categorize news articles into hierarchical levels to capture market dynamics.
• Levels Defined:
  1. Macro-Level: National/global economic events (e.g., Fed rate hikes).
  2. Sector-Level: Industry-wide events (using GICS sectors).
  3. Company-Level: Specific to a single company or small group.
• Process: An LLM classifies articles into these categories and further assigns sector-level news to specific GICS sectors.

C. Embedding-Based News Pairing (Semantic Retrieval)

• Goal: Retrieve news semantically relevant to a target company without relying on explicit keyword mentions.
• Process:
  1. Context Generation: Uses the parsed SEC filing content as a query.
  2. Fine-Tuned Embedding: A pretrained embedding model (Linq-Embed-Mistral) is fine-tuned using contrastive learning on sector-classified news. This ensures articles from the same sector have higher similarity.
  3. Retrieval: The SEC context is used to retrieve the top-$N$ news articles.
  4. Refinement: Retrieved articles are re-classified by an LLM into Target Company, Related Company (competitors/partners), or Irrelevant. Irrelevant articles are discarded.

D. LLM-Based News Summarization

• Goal: Create compact, informative inputs for forecasting models.
• Process: Raw news is often noisy or redundant. An LLM summarizes the retrieved articles based on the context (e.g., "Macro," "Sector," or "Company") into top-$N$ event categories, filtering out peripheral details.

3. Key Contributions

1. Framework Proposal: Introduced a novel framework that moves beyond keyword matching to semantic-based and multi-level pairing, addressing the complex interdependencies of financial markets.
2. FinTexTS Dataset: Constructed a large-scale dataset covering 100 top US companies over a 5-year period (2019–2023). It pairs stock prices with approximately 1 million news articles, structured into four levels: Macro, Sector, Related Company, and Target Company.
3. Resource Availability: Publicly released the dataset, the fine-tuned embedding model, and the code for the pairing framework and pilot study.
4. Empirical Validation: Demonstrated that semantic pairing and multi-level integration significantly outperform keyword-based baselines in stock forecasting tasks.

4. Experimental Results

The authors conducted a pilot study using 12 state-of-the-art time-series forecasting models (e.g., Autoformer, PatchTST, iTransformer) to predict stock prices.

• Semantic vs. Keyword Pairing:

  • The semantic-based approach consistently outperformed keyword-based pairing across all 12 models.
  • Metrics: Significant reductions in Mean Squared Error (MSE) and Mean Absolute Error (MAE). For example, on the Autoformer model, MSE dropped from 0.156 (Keyword) to 0.083 (Semantic).
  • Reasoning: Semantic pairing provides more informative signals and avoids the noise/inconsistency of keyword matching.

• Multi-Level Pairing Effect:

  • Incrementally adding text levels (Macro $\to$ Sector $\to$ Related Company $\to$ Target) consistently improved forecasting performance.
  • This confirms that financial forecasting benefits from complementary signals across different market scales.

• Proprietary Data Analysis:

  • When the framework was applied to a proprietary, high-quality news source (LSEG Machine Readable News), performance improved further compared to publicly available news, highlighting the importance of data quality alongside advanced pairing strategies.

• Sensitivity & Case Studies:

  • Retrieval Size ($N$): Performance improved with $N > 0$, with $N=10$ identified as the optimal balance between performance and computational cost.
  • Fine-Tuning: The fine-tuned embedding model showed a 1.4% to 5.9% improvement in hit-rate over the base model for retrieving relevant company news.
  • Case Study (AMD): Visual analysis showed that incorporating multi-level text (e.g., Fed rates at macro, competitor performance at related company level) allowed the model to better track actual stock price movements.

5. Significance

• Paradigm Shift: The paper challenges the prevailing "keyword-matching" paradigm in financial NLP, arguing that semantic understanding and hierarchical context are essential for accurate financial time-series modeling.
• Benchmarking: FinTexTS provides a rigorous, large-scale benchmark for evaluating text-time-series multimodal models in finance, moving beyond simple sentiment analysis to complex event correlation.
• Future Research: The work opens new avenues for developing multimodal architectures specifically designed to ingest and reason over multi-level textual hierarchies, potentially leading to more robust financial forecasting systems.