From Text to Alpha: Can LLMs Track Evolving Signals in Corporate Disclosures?

This paper proposes a novel framework that leverages large language models to extract context-aware semantic signals from corporate disclosures, demonstrating that tracking metric shifts via embedding-based similarity significantly outperforms traditional NER-based methods in predicting financial alpha.

The Gist

Imagine you are trying to predict which horses will win the next race. You have a notebook full of the horses' past performance reports.

The Old Way (The "Keyword Hunter"):
In the past, financial analysts used a method like a keyword hunter. They would scan these reports looking for specific words like "revenue," "profit," or "growth."

• The Problem: This method is a bit clumsy. If a company says, "We made a lot of money from our cloud services in North America," the keyword hunter might just see "money" or miss it entirely if the phrasing is slightly different.
• The "Moving Target" Trick: Sometimes, when a company is doing poorly in one area (like sales), their managers stop talking about sales and start bragging about something else (like "cost savings" or "new strategies"). This is called "Moving the Goalposts."
• The Flaw: The old keyword hunter gets confused by this. It thinks the company is just changing the subject randomly, or it misses the subtle shift because it's too focused on exact word matches. It can't tell if "Sales Growth" is the same concept as "Revenue Increase."

The New Way (The "Smart Librarian"):
This paper introduces a new approach using Large Language Models (LLMs), which are like super-smart librarians who actually read and understand the story, rather than just counting words.

Here is how their new system works, using two simple tools:

1. The Extractor (The Smart Librarian)

Instead of just grabbing words, the LLM reads the whole sentence and pulls out the meaning.

• Old Way: Sees "revenue."
• New Way: Sees "North American cloud revenue" and understands that this is a specific, important metric. It keeps the context (the "qualifiers") that the old method threw away.

2. The Ruler (The Semantic Tape Measure)

Once the librarian pulls out the metrics from this year's report and last year's report, they need to compare them.

• Old Way: Checks if the words are spelled the same. If last year they said "Sales" and this year they said "Revenue," the old ruler says, "These are different! We lost a metric!"
• New Way: The LLM uses a semantic ruler. It knows that "Sales" and "Revenue" mean the same thing, even if the words are different. It can measure how much the company has actually changed its focus, rather than just how the words changed.

The Big Discovery

The researchers tested this on 100 big companies (like those in the S&P 100) over 14 years. They looked at companies that kept changing their "goalposts" (shifting which metrics they bragged about) versus those that stayed consistent.

• The Result: The old method (keyword hunter) couldn't really predict anything. It was like trying to guess the weather by looking at a broken thermometer.
• The New Result: The new LLM method found a clear pattern. Companies that kept "moving the goalposts" (shifting their focus to hide bad news or distract from problems) tended to perform worse in the stock market later on.
• The Payoff: The new method predicted stock returns twice as well as the old method. It found "alpha" (extra profit) that the old method completely missed.

Why Does This Matter?

Think of it like this:

• The Old Method is like a security guard who only stops people wearing red hats. If someone wears a red scarf, they get through.
• The New Method is like a security guard who understands intent. They know that a red scarf might be a disguise for a red hat, and they catch the real story.

In short: Companies that keep changing the subject in their reports are often trying to hide something. The old computers couldn't see through the disguise, but the new AI "Smart Librarian" can. By spotting these subtle shifts in language, investors can avoid bad stocks and make better money.

Technical Summary

Here is a detailed technical summary of the paper "From Text to Alpha: Can LLMs Track Evolving Signals in Corporate Disclosures?"

1. Problem Statement

Traditional Natural Language Processing (NLP) methods in quantitative finance, primarily relying on Named Entity Recognition (NER) and keyword matching, struggle to capture the rich, nuanced semantics of corporate disclosures. Specifically, they fail to effectively track "Moving Targets"—a phenomenon where management strategically shifts the performance metrics they emphasize across consecutive earnings calls (e.g., pivoting from "revenue growth" to "cost savings" when growth stalls).

Existing NER-based approaches often:

• Collapse contextually distinct metrics into generic labels (e.g., treating "North America cloud revenue" and "Total revenue" as identical).
• Miss metrics expressed in complex phrases or different semantic forms (e.g., "sales growth" vs. "revenue increase").
• Extract non-metric terms (e.g., "the %", "the range") that introduce noise.

The core research question is: Can Large Language Models (LLMs) extract semantic signals from corporate disclosures that predict abnormal returns (alpha) more effectively than traditional NER-based methods?

2. Methodology

The authors propose a framework titled "LLM as extractor, embedding as ruler" to quantify semantic changes in performance metrics over time.

A. The Framework

1. LLM as Extractor:

   • Instead of rule-based extraction, a pre-trained LLM ($F_\theta$) processes earnings call transcripts.
   • It extracts performance metrics while preserving contextual qualifiers (e.g., extracting "operating cash flow" rather than just "cash flow").
   • This addresses the limitation of NER, which often misses specific, context-dependent metrics.

2. Embedding as Ruler:

   • Extracted metrics from the current period ($i$) and a previous period ($j = i-4$) are encoded into a shared vector space using a pre-trained text encoder ($G_\phi$).
   • Semantic Similarity Scoring: For each metric in the previous period, the system computes the maximum cosine similarity against all metrics in the current period.
   • Thresholding: A piecewise-linear transformation ($h(\cdot)$) is applied to raw similarity scores to filter noise:
     • Scores below a threshold $\alpha$ are treated as 0 (metric dropped).
     • Scores above a threshold $\beta$ are treated as 1 (metric retained).
     • Scores in between are linearly scaled.
   • Moving Targets (MT) Score: The final score is calculated as the fraction of previously discussed metrics that are no longer emphasized (i.e., $1 - \text{average similarity}$).

B. Experimental Setup

• Dataset: Earnings call transcripts of S&P 100 firms from January 2010 to December 2024 (5,615 firm-quarter observations).
• Baseline: Compared against the NER-based Moving Targets method proposed by Cohen and Nguyen (2024).
• Models:
  • Extractor: Gemini-2.5-Pro.
  • Encoder: text-embedding-3-large.
• Evaluation Protocols:
  1. Portfolio Strategy: Long-short portfolios sorted by MT scores (Long: Low shifting; Short: High shifting). Performance measured via Fama-French 3-factor and 5-factor alphas.
  2. Cross-Sectional Regression: Fama-MacBeth regressions predicting next-month stock returns, controlling for standard factors (Size, Book-to-Market, Momentum).

3. Key Contributions

• Novel Framework: Introduction of a generalizable "Extractor-Ruler" pipeline that combines LLMs for context-aware extraction and embedding models for semantic comparison.
• Semantic Precision: Demonstration that LLMs can identify semantically equivalent metrics expressed differently (e.g., "revenue increase" vs. "sales growth"), a capability NER lacks.
• Noise Reduction: Qualitative analysis shows LLMs filter out non-metric noise (e.g., "the %") that plagues keyword-based approaches.
• Alpha Generation: Empirical proof that semantic tracking of metric shifting yields significant risk-adjusted returns, whereas traditional methods do not.

4. Results

Quantitative Performance

• Portfolio Returns:
  • NER-based Method: Produced a negative but statistically insignificant spread between high and low shifting portfolios ($Q5-Q1$ alpha $\approx -0.14\%$ to $-0.18\%$, t-stat $< 1.0$).
  • LLM-based Method: Produced a significantly negative spread ($Q5-Q1$ 5-factor alpha of $-0.52\%$, t-stat $-2.55$). This represents more than double the risk-adjusted alpha of the baseline.
• Cross-Sectional Regression:
  • NER-based: The coefficient on the MT score was positive and insignificant ($0.0107$), suggesting no predictive power.
  • LLM-based: The coefficient was negative and significant ($-0.0370$), confirming that higher metric shifting predicts lower future returns.

Qualitative Insights

• Extraction Quality: Table 2 and Figure 3 illustrate that NER extracts generic patterns ("the %", "a % increase"), while LLMs extract concrete business metrics ("market share", "free cash flow", "pricing").
• Context Preservation: In a case study of a PepsiCo transcript, NER missed key metrics like "operating cash flow," extracting only "Management" and "dividends." The LLM correctly identified the full spectrum of cash flow metrics.

5. Significance and Conclusion

The paper demonstrates that LLMs provide a robust mechanism for tracking evolving signals in corporate disclosures. By moving beyond surface-level keyword matching to deep semantic understanding, the proposed method uncovers predictive information that traditional NLP pipelines miss.

• Investment Implication: Firms that frequently shift the metrics they emphasize (high "Moving Targets") are likely underperforming, and this signal is only detectable via LLM-based semantic analysis.
• Broader Impact: The framework is portable to other corporate disclosures (annual reports, shareholder letters) and suggests a new paradigm for quantitative finance: using LLMs not just for sentiment analysis, but for tracking the evolution of strategic narratives.

Limitations: The study is restricted to S&P 100 firms due to computational costs, and the robustness across different LLM/Embedding model choices requires further investigation. The economic mechanisms behind why semantic shifting predicts returns remain an open question for future research.