Paper: arXiv 2510.19203

Authors: Yuntao Wu, Lynn Tao, Ing-Haw Cheng, Charles Martineau, Yoshio Nozawa, John Hull, Andreas Veneris

Abstract

News spreads rapidly across languages and regions, but translations may lose subtle nuances. We propose a method to align sentences in multilingual news articles using optimal transport, identifying semantically similar content across languages. We apply this method to align more than 140,000 pairs of Bloomberg English and Japanese news articles covering around 3500 stocks in Tokyo exchange over 2012-2024. Aligned sentences are sparser, more interpretable, and exhibit higher semantic similarity. Return scores constructed from aligned sentences show stronger correlations with realized stock returns, and long-short trading strategies based on these alignments achieve 10% higher Sharpe ratios than analyzing the full text sample.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; reported Sharpe ratio of 4.36 is implausibly high for daily or lower-frequency data; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)

Why this score: The paper uses advanced mathematical concepts like optimal transport and Sinkhorn algorithms, while also performing extensive empirical backtesting on a large-scale dataset of 140,000 news article pairs with implemented trading strategies and reported Sharpe ratios.

Research Flowchart

  flowchart TD
  A["Research Goal: Multilingual News Alignment<br>for Stock Return Prediction"] --> B["Data Input: 140k+ Bloomberg EN/JP<br>Articles (2012-2024, ~3500 Tokyo Stocks)"]
  B --> C["Methodology: Optimal Transport<br>Sentence Alignment"]
  C --> D["Computational Process: Map EN/JP<br>Sentences via Semantic Similarity"]
  D --> E{"Outcomes"}
  E --> F["Sparsity & Interpretability:<br>Aligned content is cleaner"]
  E --> G["Correlation Analysis:<br>Aligned scores vs. Realized Returns"]
  E --> H["Trading Strategy:<br>Long-Short based on alignments"]
  H --> I["Final Result: 10% Higher Sharpe Ratio<br>vs. Full Text Analysis"]

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 a sentence‑level alignment method for multilingual news using optimal transport to improve stock return prediction.

Source sentence
We propose a method to align sentences in multilingual news articles using optimal transport, identifying semantically similar content across languages.

Assumes 5

  • Articles published within 30 minutes before market open are linked to that day’s stock price and return. data
    Source sentence
    articles published between 30 minutes before market open on day
  • Only articles with a ticker score ≥75 are considered, implying medium‑to‑high relevance to the stock. data
    Source sentence
    we include English and foreign-language articles that are explicitly associated with a single stock ticker with a ticker score ≥75, indicating medium-to-high relevance to that stock.
  • Sentences in each article are assigned equal probability mass in the optimal transport formulation. model
    Source sentence
    Sentences are assigned equal probability mass, i.e., uniformly distributed.
  • Cost matrix is defined by pairwise cosine distance scaled to [0,1] with min‑max normalization. statistical
    Source sentence
    The cost matrix is defined using pairwise cosine distance
Show 1 more
  • Trading signals are formed from daily long‑short portfolios without accounting for transaction costs. execution
    Source sentence
    Portfolios constructed from aligned embeddings exhibit even higher Sharpe ratios for both Japanese and English texts, suggesting that sentences capturing common themes across languages provide clearer signals of stock performance.

Data

Backtest sample: 2012 – 2024 · Tokyo exchange stocks · equities · daily

  • Bloomberg news feed · Bloomberg text-news commercial
    2008–2024 · ≈3,500 Tokyo‑listed stocks, English and Japanese articles
    Source sentence
    We collect global news articles from the Bloomberg terminal news feed covering the period 2008–2024.
  • Compustat Global via WRDS · Compustat Global (WRDS) market-prices commercial
    2008–2024 · Tokyo Stock Exchange stocks
    Source sentence
    Stock return data are obtained from Compustat Global using WRDS for the period 2008–2024.

Recognised sources: Bloomberg Compustat WRDS

Costs

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

Finds 4

Reported performance (aligned sentence embeddings long-short portfolio): Sharpe 4.36

  • Long‑short strategies based on aligned sentence embeddings achieve higher Sharpe ratios (e.g., 4.36 annualized Sharpe for Japanese aligned embeddings) than those using unaligned or full‑text embeddings. performance
    Source sentence
    Portfolios constructed from aligned embeddings exhibit even higher Sharpe ratios for both Japanese and English texts, suggesting that sentences capturing common themes across languages provide clearer signals of stock performance.
  • Return scores built from aligned sentences show stronger correlations with realized stock returns than scores from unaligned or full‑text embeddings. predictive-accuracy
    Source sentence
    Return scores constructed from aligned sentences show stronger correlations with realized stock returns.
  • Aligned sentences are sparser, more interpretable, and exhibit higher semantic similarity than those obtained via pure cosine similarity. statistical
    Source sentence
    Aligned sentences are sparser, more interpretable, and exhibit higher semantic similarity.
  • In the Japanese market, full embeddings of Japanese news generate higher Sharpe than English news, indicating stronger predictive power of native‑language coverage. performance
    Source sentence
    In the Japanese market, the full embedding of Japanese news generates higher Sharpe than English news.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.

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 (52)

  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. Attention is All you Need (2017) — link
  6. Business News and Business Cycles (2024) — link
  7. How Hack of SEC’s EDGAR System Exposed Flaws in US Financial Security (2025) — link
  8. Self-serving biases in the attribution process: A reexamination of the fact or fiction question. (1978)
and 44 more
  1. Expected Returns and Large Language Models (2022) — link
  2. Acquisitions, overconfident managers and self-attribution bias (2007)
  3. The extreme future stock returns following I/B/E/S earnings surprises (2006)
  4. ECC Analyzer: Extracting Trading Signal from Earnings Conference Calls using Large Language Model for Stock Volatility Prediction (2024) — link
  5. Efficient Capital Markets: A Review of Theory and Empirical Work (1970) — link
  6. The adjustment of stock prices to new information (1969)
  7. The cross-section of expected stock returns (1992)
  8. FinBERT: Financial Sentiment Analysis with Pre-trained Language Models (2019) — link
  9. Good Debt or Bad Debt: Detecting Semantic Orientations in Economic Texts (2014) — link
  10. The colour of finance words (2023)
  11. GPT-4o System Card (2024) — link
  12. How is earnings news transmitted to stock prices? (2022)
  13. Tone management (2014)
  14. Rest in peace post-earnings announcement drift (2022)
  15. An analysis of intraday patterns in bid/ask spreads for NYSE stocks (1992)
  16. PEAD. txt: post-earnings-announcement drift using text (2023)
  17. MPNet: Masked and Permuted Pre-training for Language Understanding (2020) — link
  18. Online learning for Latent Dirichlet Allocation (2010)
  19. Stochastic Variational Inference (2013) — link
  20. Earnings conference calls and stock returns: The incremental informativeness of textual tone (2012)
  21. A Unified Approach to Interpreting Model Predictions (2017) — link
  22. Analysis of regression in game theory approach (2001) — link
  23. Notes on the n-Person Game – II: The Value of an n-Person Game (1951) — link
  24. A value for n-person games (1988)
  25. Explaining Prediction Models and Individual Predictions with Feature Contributions (2014) — link
  26. Understanding the Excess Bond Premium — → 2412.04063
  27. Attribution of success and failure revisited, or: The motivational bias is alive and well in attribution theory (1979)
  28. Lost in translation. When sentiment metrics for one market are derived from two different languages (2023) — link
  29. Word Alignment by Fine-tuning Embeddings on Parallel Corpora (2021) — link
  30. Rationalizing Text Matching: L earning Sparse Alignments via Optimal Transport (2020) — link
  31. Mémoire sur la théorie des déblais et des remblais (1781)
  32. Constructing Optimal Maps for Monge’s Transport Problem as a Limit of Strictly Convex Costs (2002)
  33. On the Translocation of Mass (1942)
  34. Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling (2015) — link
  35. Computational Optimal Transport (2020) — link
  36. A Relationship Between Arbitrary Positive Matrices and Doubly Stochastic Matrices (1964)
  37. The Sparse-Plus-Low-Rank Quasi-Newton Method for Entropic-Regularized Optimal Transport (2025) — link
  38. A Program for Aligning Sentences in Bilingual Corpora (1993) — link
  39. IBM Models (2015)
  40. A Systematic Bayesian Treatment of the IBM Alignment Models (2013) — link
  41. A Statistical Approach to Language Translation (1988)
  42. Language-agnostic BERT Sentence Embedding (2022) — link
  43. Sparse Sequence-to-Sequence Models (2019) — link
  44. spaCy: Industrial-strength Natural Language Processing in Python (2020) — link

1 of 52 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.