Paper: arXiv 2505.06864

Authors: Shunyao Wang, Ming Cheng, Christina Dan Wang

Abstract

Stochastic Discount Factor (SDF) models provide a unified framework for asset pricing and risk assessment, yet traditional formulations struggle to incorporate unstructured textual information. We introduce NewsNet-SDF, a novel deep learning framework that seamlessly integrates pretrained language model embeddings with financial time series through adversarial networks. Our multimodal architecture processes financial news using GTE-multilingual models, extracts temporal patterns from macroeconomic data via LSTM networks, and normalizes firm characteristics, fusing these heterogeneous information sources through an innovative adversarial training mechanism. Our dataset encompasses approximately 2.5 million news articles and 10,000 unique securities, addressing the computational challenges of processing and aligning text data with financial time series. Empirical evaluations on U.S. equity data (1980-2022) demonstrate NewsNet-SDF substantially outperforms alternatives with a Sharpe ratio of 2.80. The model shows a 471% improvement over CAPM, over 200% improvement versus traditional SDF implementations, and a 74% reduction in pricing errors compared to the Fama-French five-factor model. In comprehensive comparisons, our deep learning approach consistently outperforms traditional, modern, and other neural asset pricing models across all key metrics. Ablation studies confirm that text embeddings contribute significantly more to model performance than macroeconomic features, with news-derived principal components ranking among the most influential determinants of SDF dynamics. These results validate the effectiveness of our multimodal deep learning approach in integrating unstructured text with traditional financial data for more accurate asset pricing, providing new insights for digital intelligent decision-making in financial technology.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 6.0/10 (capped from 9.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 9.0)

Why this score: The paper presents dense mathematical formulations including the stochastic discount factor framework, moment conditions, and adversarial training mechanisms, indicating high complexity. It demonstrates exceptional empirical rigor with a massive dataset (2.5M news articles, 10K securities), 42-year backtest, and detailed performance metrics including Sharpe ratios, pricing error reductions, and ablation studies.

Research Flowchart

  flowchart TD
  A["Research Goal: Integrate unstructured text with financial data for SDF estimation"] --> B["Data Preparation: 2.5M News Articles & 10K Securities"]
  B --> C["Methodology: Multimodal Deep Learning Architecture"]
  C --> D["Text Processing: GTE-multilingual Embeddings"]
  C --> E["Time Series: LSTM Networks"]
  C --> F["Normalization: Firm Characteristics"]
  D & E & F --> G["Adversarial Network Fusion"]
  G --> H["Outcomes: 2.80 Sharpe Ratio, 471% Improvement over CAPM"]
  G --> I["Findings: News embeddings > Macro features, 74% Pricing Error Reduction"]

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 NewsNet‑SDF, a novel deep learning framework that integrates pretrained language model news embeddings with stochastic discount factor estimation via adversarial networks.

Source sentence
We introduce NewsNet‑SDF, a novel deep learning framework that seamlessly integrates pretrained language model embeddings with financial time series through adversarial networks.

Assumes 2

  • The fundamental asset pricing moment condition Et[Mt+1Re t+1] = 0 must hold. model
    Source sentence
    These diverse information sources are fused and fed into parallel SDF and conditional networks trained adversarially to enforce the fundamental asset pricing equation Et[Mt+1Re t+1] = 0.
  • Stocks have all required firm characteristics available each month for inclusion. data
    Source sentence
    Following [4], we restrict our analysis to stocks with all required characteristics available in a given month, yielding approximately 10,000 unique securities.

Data

Backtest sample: 1980 – 2022 · U.S. equity market · equities

  • News articles · The New York Times text-news unknown
    1980-2022 · U.S. equities
    Source sentence
    Our textual data consists of approximately 2.5 million articles from The New York Times (1980-2022), processed through our attention-based neural framework with careful temporal alignment to financial data.
  • U.S. equity data market-prices unknown
    1980-2022 · approximately 10,000 unique securities
    Source sentence
    Empirical evaluations on U.S. equity data (1980-2022) demonstrate NewsNet‑SDF substantially outperforms alternatives with a Sharpe ratio of 2.803.
  • Firm characteristics fundamentals unknown
    56 firm‑specific characteristics
    Source sentence
    We construct 56 firm‑specific characteristics following [37] and [38], spanning valuation, profitability, financial soundness, capitalization, operational efficiency, and momentum categories.
  • Macroeconomic time series · FRED‑MD database macro unknown
    157 series
    Source sentence
    For macroeconomic data, we incorporate 157 time series from the FRED‑MD database and other sources, covering output, employment, housing, and financial market indicators.

Recognised sources: FRED

Costs

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

Finds 5

Reported performance (NewsNet-SDF): Sharpe 2.80

  • NewsNet‑SDF achieves a Sharpe ratio of 2.80, far exceeding traditional factor models. performance
    Source sentence
    Empirical evaluations on U.S. equity data (1980-2022) demonstrate NewsNet‑SDF substantially outperforms alternatives with a Sharpe ratio of 2.803.
  • The model improves Sharpe ratio by 471% over CAPM and over 200% versus conventional SDF implementations. performance
    Source sentence
    The model shows a 471% improvement over CAPM, over 200% improvement versus traditional SDF implementations.
  • Pricing errors are reduced by 74% compared to the Fama‑French five‑factor model. pricing-calibration
    Source sentence
    We enhance SDF estimation through a novel loss function that dynamically reweights pricing errors based on factor exposure, reducing pricing errors by 74% compared to the Fama‑French five‑factor model.
  • Ablation of news embeddings cuts Sharpe ratio by 41%, while removing macroeconomic features cuts it by 31%, indicating news provides unique predictive signals. predictive-accuracy
    Source sentence
    Ablation studies reveal that eliminating news text causes a 41% decrease in Sharpe ratio versus a 31% reduction when removing macroeconomic features.
  • News‑derived principal components rank among the most influential determinants of SDF dynamics. statistical
    Source sentence
    Ablation studies confirm that text embeddings contribute significantly more to model performance than macroeconomic features, with news‑derived principal components ranking among the most influential determinants of SDF dynamics.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 62 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive (1 by independent authors): Transformer-based CoVaR: Systemic Risk in Textual … (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.

Cited References (41)

  1. Towards General Text Embeddings with Multi-stage Contrastive Learning (2023) — link
  2. Open Source Cross-Sectional Asset Pricing (2021) — link
  3. Deep Attentive Learning for Stock Movement Prediction from Social Media Text and Company Correlations (2020) — link
  4. FinBERT: A Pretrained Language Model for Financial Communications (2020) — link
  5. Risk Factors That Matter: Textual Analysis of Risk Disclosures for the Cross-Section of Returns (2019) — link
  6. FinBERT: Financial Sentiment Analysis with Pre-trained Language Models (2019) — link
  7. Predicting Returns with Text Data (2019) — link
  8. RoBERTa: A Robustly Optimized BERT Pretraining Approach (2019) — link
and 33 more
  1. Deep Learning in Asset Pricing (2019) — link
  2. Empirical Asset Pricing via Machine Learning (2018) — link
  3. Investor-Imitator: A Framework for Trading Knowledge Extraction (2018) — link
  4. DYNAMIC FACTOR MODELS (2018) — link
  5. The Characteristics that Provide Independent Information about Average U.S. Monthly Stock Returns (2017) — link
  6. Natural language based financial forecasting: a survey (2017) — link
  7. Stock price prediction using LSTM, RNN and CNN-sliding window model (2017) — link
  8. Stock Price Prediction via Discovering Multi-Frequency Trading Patterns (2017) — link
  9. Stock market’s price movement prediction with LSTM neural networks (2017) — link
  10. A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction (2017) — link
  11. Textual Analysis in Accounting and Finance: A Survey: TEXTUAL ANALYSIS IN ACCOUNTING AND FINANCE (2016) — link
  12. ImageNet classification with deep convolutional neural networks (2012) — link
  13. A Comprehensive Look at the Empirical Performance of Equity Premium Prediction (2009) — link
  14. Long Short-Term Memory (1997) — link
  15. Common risk factors in the returns on stocks and bonds (1993) — link
  16. The Persistence of Mutual Fund Performance (1992) — link
  17. Learning representations by back-propagating errors (1986) — link
  18. The arbitrage theory of capital asset pricing (1976) — link
  19. CAPITAL ASSET PRICES: A THEORY OF MARKET EQUILIBRIUM UNDER CONDITIONS OF RISK* (1964) — link
  20. Dissecting characteristics nonparametri-cally (2020)
  21. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2019) — link
  22. Working Paper Series (2019)
  23. NBER WORKING PAPER SERIES CHARACTERISTICS ARE COVARIANCES: A UNIFIED MODEL OF RISK AND RETURN (2018)
  24. Shrinking the Cross Section∗ (2017)
  25. A five-factor asset pricing model (2013) — link
  26. Investor sentiment in the stock market (2013)
  27. Internet Appendix for “When is a Liability not a Liability? Textual Analysis, Dictionaries, and 10-Ks” * (2011)
  28. Sentiment during recessions∗ (2011)
  29. Giving Content to Investor Sentiment: The Role of Media in the Stock Market (2005) — link
  30. Asset Pricing: Revised Edition. Princeton University Press (2009) 10. Hansen, L.P., Singleton, K.J.: Generalized instrumental variables estimation of nonlinear rational expectations models (1982)
  31. The dynamic factor analysis of economic time series (1977)
  32. Large Sample Properties of Generalized Method of Moments Estimators
  33. FinBERT-MRC: Financial domain-specific language model for Chinese market

0 of 41 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.