Paper: arXiv 2509.12519

Authors: Ross Koval, Nicholas Andrews, Xifeng Yan

Abstract

Financial news plays a critical role in the information diffusion process in financial markets and is a known driver of stock prices. However, the information in each news article is not necessarily self-contained, often requiring a broader understanding of the historical news coverage for accurate interpretation. Further, identifying and incorporating the most relevant contextual information presents significant challenges. In this work, we explore the value of historical context in the ability of large language models to understand the market impact of financial news. We find that historical context provides a consistent and significant improvement in performance across methods and time horizons. To this end, we propose an efficient and effective contextualization method that uses a large LM to process the main article, while a small LM encodes the historical context into concise summary embeddings that are then aligned with the large model’s representation space. We explore the behavior of the model through multiple qualitative and quantitative interpretability tests and reveal insights into the value of contextualization. Finally, we demonstrate that the value of historical context in model predictions has real-world applications, translating to substantial improvements in simulated investment performance.

Complexity vs Empirical Score

  • Math Complexity: 6.0/10
  • Empirical Rigor: 7.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced language model architectures and alignment techniques with mathematical foundations, while demonstrating strong empirical rigor through backtested investment simulations, dataset curation, and quantifiable performance metrics like AUC.

Research Flowchart

  flowchart TD
  A["Research Goal: Assess Value of Historical Context<br>in LMs for Financial News Market Impact Prediction"] --> B["Data & Inputs: Financial News Articles &<br>Stock Price Data"]
  B --> C["Methodology: Propose Novel Contextualization Model<br>Small LM encodes history into embeddings<br>Aligns with Large LM's representation space"]
  C --> D["Computational Process: Train & Evaluate Model<br>Across Multiple Time Horizons"]
  D --> E["Key Findings & Outcomes:<br>1. Historical context yields consistent, significant performance gains<br>2. Interpretability tests reveal contextual insights<br>3. Simulated investment performance improves substantially<br>4. Validated real-world applicability"]

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 efficient and effective contextualization method that uses a large language model for the main article and a small language model to encode historical context into summary embeddings aligned with the large model’s space.

Source sentence
we propose an efficient and effective contextualization method that uses a large LM to process the main article, while a small LM encodes the historical context into concise summary embeddings that are then aligned with the large model’s representation space.

Assumes 4

  • The binary classification task is evaluated with AUC because continuous scores are more informative for investment management applications. model
    Source sentence
    We evaluate the performance on this binary classification task with AUC because the continuous scores are more informative than their discrete predicted classes for investment management applications.
  • Transaction costs are estimated and included in the reported investment performance. execution
    Source sentence
    Net performance includes an estimate of the impact of transaction costs, detailed in §A.1.
  • The most recent N=5 historical articles provide the best context for the main article unless relevance is considered. data
    Source sentence
    In the experiments so far, we assume that the N = 5 most recent historical articles about the same company provide the best context for the main article (Time).
  • Historical news articles from FactSet StreetAccount are representative of US‑based publicly‑traded companies’ information flow. data
    Source sentence
    We collect stock-specific financial news articles in English from FactSet StreetAccount for US-based publicly-traded companies.

Data

  • FactSet StreetAccount news · FactSet text-news commercial
    2010-2023 · US publicly‑traded companies
    Source sentence
    We collect stock-specific financial news articles in English from FactSet StreetAccount for US-based publicly-traded companies. These articles cover a variety of events and news sources.

Recognised sources: FactSet

Costs

Modeled Costs are part of the reported headline results.

Costs are estimated and incorporated into net portfolio returns; see §A.1 for the specific model.

Source sentence
Net performance includes an estimate of the impact of transaction costs, detailed in §A.1.

Finds 3

  • Historical context provides a consistent and significant improvement in performance across methods and time horizons. performance
    Source sentence
    We find that historical context provides a consistent and significant improvement in performance across methods and time horizons.
  • Increasing the number of historical context articles improves forecasting performance and reduces LM loss up to about 15 articles, with diminishing returns after 10. performance
    Source sentence
    The results provide evidence towards our hypothesis that model performance on our task demonstrates language understanding as we find that both forecasting performance and LM loss improves with the number of contexts incorporated.
  • All contextualization methods provide a material benefit over their corresponding single‑article baselines, with summarization approaches outperforming simple concatenation. performance
    Source sentence
    While we find that all contextualization methods provide a material benefit over their corresponding single-article baselines, we find that approaches that summarize the relevant context into a more concise and informative form

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 12 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

https://github.com/rosskoval/calm_fin_news — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.