Paper: arXiv 2509.24254

Authors: Yuntao Wu, Ege Mert Akin, Charles Martineau, Vincent Grégoire, Andreas Veneris

Abstract

We examine how textual features in earnings press releases predict stock returns on earnings announcement days. Using over 138,000 press releases from 2005 to 2023, we compare traditional bag-of-words and BERT-based embeddings. We find that press release content (soft information) is as informative as earnings surprise (hard information), with FinBERT yielding the highest predictive power. Combining models enhances explanatory strength and interpretability of the content of press releases. Stock prices fully reflect the content of press releases at market open. If press releases are leaked, it offers predictive advantage. Topic analysis reveals self-serving bias in managerial narratives. Our framework supports real-time return prediction through the integration of online learning, provides interpretability and reveals the nuanced role of language in price formation.

Complexity vs Empirical Score

  • Math Complexity: 4.0/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper uses standard NLP models (BERT, LDA) and regression without heavy mathematical derivations, but is exceptionally data-intensive with a massive dataset (138k+ press releases, 134k+ observations), extensive preprocessing, and a validated trading strategy showing predictive power.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Predict Earnings Announcement Returns<br>from Press Release Text"] --> B["Data Input:<br>138K+ Press Releases (2005-2023)"]
  B --> C["Methodology:<br>Compare BOW vs. BERT/FinBERT Embeddings"]
  C --> D["Analysis:<br>Event Study &<br>Online Learning Models"]
  D --> E{"Key Findings"}
  E --> F1["Press Release Content<br>as informative as Hard Numbers"]
  E --> F2["FinBERT yields highest predictive power"]
  E --> F3["Prices reflect info at market open<br>Leakage offers advantage"]
  E --> F4["Topic Analysis reveals<br>Managerial Self-Serving Bias"]

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 an online learning framework that combines bag‑of‑words, BERT and large language models to predict earnings‑announcement returns from press‑release text and to enhance explainability.

Source sentence
Our contributions are threefold: (1) Unlike conventional approaches based on static train‑test splits, our framework supports online learning and adapts easily to new data; (2) We integrate the advantages of bag‑of‑words, BERT and large language models to enhance explainability by identifying key tokens and topics driv

Assumes 5

  • Stock prices fully reflect press‑release content at market open (market efficiency). market
    Source sentence
    Stock prices fully reflect the content of press releases at market open.
  • Trades can be entered at the ask (or bid) price at 9:45 a.m. and exited at the closing price without additional slippage. execution
    Source sentence
    We enter trades at 9:45 a.m since bid‑ask spreads are wide at the opening of markets [16, 23], and exit at the 4:00 p.m close.
  • Stocks with a relative bid‑ask spread exceeding 20 % of the midquote are excluded as illiquid. execution
    Source sentence
    We exclude stocks with excessive illiquidity—defined as a relative bid‑ask spread exceeding 20% of the midquote: |Ask −Bid| / Midquote ≥0.2
  • Press releases are published after‑hours (on or after 4:00 p.m. or before 9:30 a.m. ET), representing the relevant information set. data
    Source sentence
    We retain press releases published after hours, i.e., on or after 4:00 p.m. or strictly before 9:30 a.m. ET. This comprises 97% of the dataset.
Show 1 more
  • Earnings surprise is measured as the scaled difference between actual EPS and consensus forecast five days before the announcement. statistical
    Source sentence
    We compute earnings surprises (hard information) as

Data

  • Earnings press releases · SEC Edgar (8‑K filings) text-news public
    2005‑2023 · 6,543 unique firms
    Source sentence
    We retrieve 158,797 earnings press releases for 6,543 unique firms (PERMNOs in CRSP). After
  • CRSP Daily Stock Files · CRSP market-prices commercial
    2005‑2023 · All U.S. stocks with earnings announcements
    Source sentence
    We retrieve daily stock returns and open and closing prices from CRSP Daily Stock Files
  • NYSE Trade and Quote (TAQ) · NYSE order-book commercial
    2005‑2023 · Stocks with TAQ data aligned to earnings events
    Source sentence
    intraday quotes from NYSE Trade and Quote (TAQ)
  • LSEG IBES Academic · LSEG fundamentals commercial
    2005‑2023 · Analyst forecasts for earnings announcements
    Source sentence
    We obtain earnings dates and times, analysts’ forecasts, and announced EPS from LSEG IBES Academic.

Recognised sources: CRSP NYSE TAQ I/B/E/S Refinitiv / LSEG (Datastream, Eikon, Tick History)

Costs

Unclear - sources disagree The platform classifier and the paper-text extraction give different answers, so no value is asserted.

Platform classifier: Modeled

Paper text extraction: Partial - Trades are entered at the ask (or bid) price and exited at the closing price; Strategy 2 captures transaction costs by buying at ask and selling at bid.

Source sentence (the authors' words)
Enter and exit using ask/bid quotes, capturing transaction costs. For instance, a long position buys at ask and sells at bid.

Finds 5

  • Press‑release text (soft information) is as informative as earnings surprise (hard information) for predicting announcement‑day returns, raising R² to about 4 %. predictive-accuracy
    Source sentence
    Across all model specifications, Soft is positive and statistically significant at the 1% , increasing the 𝑅2 up to 4.4%.
  • FinBERT provides the largest increase in R² among text models and yields the highest SHAP importance values. performance
    Source sentence
    The soft model estimated through FinBERT corresponds to the largest increase in 𝑅2 relative to the univariate analysis in Column (1).
  • A long‑short trading strategy based on agreement between surprise and SoftMean generates no abnormal profit; estimated alphas are small and negative. negative
    Source sentence
    ... for strategy 2, the corresponding values are 𝛼= −0.0055 and 𝛼= −0.0051, both significant at the 1% level.
  • When earnings surprise and press‑release content are available in advance (e.g., via hacking), investors can correctly identify 56 % of the top‑10 performing stocks. performance
    Source sentence
    if earnings surprises and press release content are available in advance, investors can on‑average correctly predict 56% of the top 10 performing stocks using the agreement between hard and soft information.
  • The framework supports real‑time return prediction through online learning and improves interpretability by identifying key tokens and topics. theoretical
    Source sentence
    Our framework supports real‑time return prediction through the integration of online learning, provides interpretability and reveals the nuanced role of language in price formation.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 34 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive (1 by independent authors)

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 (34)

  1. Price revelation from insider trading: E vidence from hacked earnings news (2022)
  2. An empirical evaluation of accounting income numbers (1969)
  3. Post-earnings-announcement drift: D elayed price response or risk premium? (1989)
  4. BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding (2019) — link
  5. Business News and Business Cycles (2024) — link
  6. How Hack of SEC’s EDGAR System Exposed Flaws in US Financial Security (2025) — link
  7. Self-serving biases in the attribution process: A reexamination of the fact or fiction question. (1978)
  8. Expected Returns and Large Language Models (2022) — link
and 26 more
  1. Acquisitions, overconfident managers and self-attribution bias (2007)
  2. The extreme future stock returns following I/B/E/S earnings surprises (2006)
  3. ECC Analyzer: Extracting Trading Signal from Earnings Conference Calls using Large Language Model for Stock Volatility Prediction (2024) — link
  4. Efficient Capital Markets: A Review of Theory and Empirical Work (1970) — link
  5. The adjustment of stock prices to new information (1969)
  6. The cross-section of expected stock returns (1992)
  7. FinBERT: Financial Sentiment Analysis with Pre-trained Language Models (2019) — link
  8. Good Debt or Bad Debt: Detecting Semantic Orientations in Economic Texts (2014) — link
  9. The colour of finance words (2023)
  10. GPT-4o System Card (2024) — link
  11. How is earnings news transmitted to stock prices? (2022)
  12. Tone management (2014)
  13. Rest in peace post-earnings announcement drift (2022)
  14. An analysis of intraday patterns in bid/ask spreads for NYSE stocks (1992)
  15. PEAD. txt: post-earnings-announcement drift using text (2023)
  16. MPNet: Masked and Permuted Pre-training for Language Understanding (2020) — link
  17. Online learning for Latent Dirichlet Allocation (2010)
  18. Stochastic Variational Inference (2013) — link
  19. Earnings conference calls and stock returns: The incremental informativeness of textual tone (2012)
  20. A Unified Approach to Interpreting Model Predictions (2017) — link
  21. Analysis of regression in game theory approach (2001) — link
  22. Notes on the n-Person Game – II: The Value of an n-Person Game (1951) — link
  23. A value for n-person games (1988)
  24. Explaining Prediction Models and Individual Predictions with Feature Contributions (2014) — link
  25. Understanding the Excess Bond Premium — → 2412.04063
  26. Attribution of success and failure revisited, or: The motivational bias is alive and well in attribution theory (1979)

1 of 34 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.