Paper: arXiv 2609.34169

Authors: Kevin Foley, Jonathan Hartadi, Shivesh Prakash, Swapnil Vatsal

Abstract

News may reveal systematic risk, but whether its context enhances the construction of systematic risk factors is still unclear. We seek to test whether utilizing a sentence transformer represents an improvement over techniques such as Latent Dirichlet Allocation (LDA) in the coherence of topic term lists generated from unstructured text data. To test this, the same collection of unstructured text data comprising of 394,661 articles and the same downstream financial portfolio construction pipeline were applied with the text layer differing, including the length of article text each model used and how topic terms were ranked: we benchmark LDA against a frozen sentence transformer with k-means clustering. We find that the sentence transformer branch had higher observed scores both in terms of coherence (measured by NPMI) as well as financial performance (measured by Sharpe), although the available tests do not establish outperformance. Further exploratory specifications such as utilizing spherical clustering and multi-horizon exposures had an observed excess-return Sharpe of 1.03 for the combined model. We believe that there is some promise in applying context-aware techniques on unstructured news text, but stricter tests using only information available at each date and broader datasets may be required to enhance the confidence in the observed performance.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)

Why this score: This paper presents a novel comparison of text representation techniques for asset pricing, demonstrating strong empirical rigor through a well-defined pipeline and backtesting. While the mathematical derivations are not explicitly detailed in the excerpt, the underlying models (LDA, sentence transformers, Sparse IPCA) imply a significant level of mathematical sophistication. The findings, though not establishing outperformance definitively, suggest promising avenues for context-aware NLP in finance.

Research Flowchart

  flowchart TD
    A["Research Goal: Context vs. Word Counts in Topic Models for Asset Pricing"] --> B("Methodology: Benchmark LDA vs. Sentence Transformer + K-Means")
    B --> C{"Data: 394,661 News Articles"}
    C --> D["Computational Processes: Topic Modeling, Portfolio Construction, Performance Metrics"]
    D --> E{"Outcomes: Higher Coherence (NPMI) & Financial Performance (Sharpe) for Sentence Transformer"}
    E --> F["Conclusion: Promise in Context-Aware Techniques, but Stricter Tests Needed"]

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 replacing the LDA topic model with a frozen sentence transformer in a narrative asset‑pricing pipeline.

Source sentence
the LDA topic model with a frozen sentence transformer

Assumes 5

  • The downstream pipeline (shock construction, covariance, Sparse IPCA, portfolio evaluation) is held fixed while only the text layer changes. model
    Source sentence
    the downstream construction of shocks, stock-level covariance characteristics, Sparse IPCA factors, and out-of-sample portfolios is held fixed while the text layer changes
  • LDA observes the full cleaned article whereas the sentence transformer uses up to three lead passages. data
    Source sentence
    LDA observes the full cleaned article, whereas the sentence transformer uses up to three lead passages
  • Both models are trained on the same FNSPID financial‑news corpus covering 2015‑2023. data
    Source sentence
    Both the LDA and transformer models use the full news corpus from January 2015 to December 2023 to construct topics
  • The evaluation window is limited to 41 months, which may affect statistical significance. statistical
    Source sentence
    The evidence remains limited by a short, 41-month evaluation window
Show 1 more
  • Topic count is fixed at 60 for both models, and the vocabulary is shared. model
    Source sentence
    Both the LDA and transformer models use the same collection of FNSPID articles, and both produce 60 topics in total

Data

Backtest sample: 2020-09 – 2024-01 · US stocks (CRSP panel) · equities · monthly

  • FNSPID financial‑news corpus · Dong et al., 2024 text-news public
    2015-2023
    Source sentence
    FNSPID newswire text from 2015 to 2023 (Dong et al., 2024)
  • CRSP US Stock Databases · Center for Research in Security Prices (CRSP) market-prices commercial
    Source sentence
    Daily stock returns and market‑capitalization data come from the CRSP US Stock Databases, accessed through the Wharton Research Data Services platform
  • Fama–French daily factor file · Kenneth R. French Data Library macro public
    Source sentence
    The Fama–French daily factor file used to obtain the risk‑free rate is publicly available from the Kenneth R. French Data Library

Recognised sources: FNSPID CRSP WRDS

Costs

Not modeled A trading or execution result is reported without costs.

Finds 4

Reported performance (Spherical sentence‑transformer with multi‑horizon IPCA): Sharpe 1.03 · max drawdown 62.2%

  • The sentence transformer branch yields higher topic coherence measured by NPMI compared to LDA. statistical
    Source sentence
    We find that the sentence transformer branch had higher observed scores both in terms of coherence (measured by NPMI) as well as financial performance (measured by Sharpe)
  • The transformer model achieves higher out‑of‑sample Sharpe ratios than LDA on the same financial window. performance
    Source sentence
    With the topic count and shared vocabulary held fixed across models, the transformer model yields higher topic coherence and OOS Sharpe ratios
  • The spherical ST with grouped multi‑horizon IPCA specification attains the highest observed excess‑return Sharpe of approximately 1.03. performance
    Source sentence
    combined specification attained a Sharpe ratio of approximately 1.03, which was the highest among the model specifications considered
  • The reported out‑of‑sample tests do not establish a statistically significant improvement in Sharpe ratios. negative
    Source sentence
    the OOS tests performed do not establish a statistically significant improvement in Sharpe ratios

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 26 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive (1 by independent authors): Point-in-Time Encoders for Earnings Calls (uses its method or data)

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.