NewsNet-SDF: Stochastic Discount Factor Estimation with Pretrained Language Model News Embeddings via Adversarial Networks
Shunyao Wang, Ming Cheng, Christina Dan Wang · 2025
Read NewsNet-SDF: Stochastic Discount Factor Estimation with Pretrained Language Model News Embeddings via Adversarial Networks by Shunyao Wang, Ming Cheng, Christina Dan Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-06864.txt.
Introduction
Shunyao Wang, Ming Cheng, and Christina Dan Wang feed financial news into an asset pricing model in a 2025 paper. Pricing models usually ignore text. Their system, NewsNet-SDF, embeds about 2.5 million news articles with a multilingual language model, pairs them with market and firm data across 10,000 securities, and fuses everything through adversarial training.
On US stocks from 1980 to 2022, the model reached a Sharpe ratio of 2.80. It beat the classic CAPM benchmark by 471 percent and cut pricing errors by 74 percent against the Fama-French five-factor model. Tests that removed parts of the system showed news text contributed more than macroeconomic features. The authors present text as a first-class input to pricing.
Catalog record
| CREATOR | Shunyao Wang, Ming Cheng, Christina Dan Wang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-06864
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
NewsNet-SDF: Stochastic Discount Factor Estimation with Pretrained Language Model News Embeddings via Adversarial Networks 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 1 1 1 In investment management literature, Sharpe ratios above 1.0 indicate good risk-adjusted performance, while values above 2.0 are considered excellent, exceeding most professional investment strategies. . 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. 1 Introduction Asset pricing, a fundamental pillar of modern financial theory, aims to explain expected returns across different securities through systematic risk factors. The Stochastic Discount Factor (SDF) framework has emerged as a unified theoretical foundation that encompasses traditional models like CAPM [ 1 ] and Fama-French factor models [ 2 ] . Unlike conventional approaches that specify risk factors exogenously, SDF directly models the pricing kernel that links current prices to future payoffs, providing a more flexible and theoretically consistent framework for asset valuation and risk assessment [ 9 ] . Recent advances in financial machine learning have demonstrated the potential to enhance asset pricing models by capturing complex non-linear relationships between features and returns [ 3 , 4 , 11 ] . However, despite these methodological improvements, a significant challenge remains: effectively incorporating the vast amount of unstructured textual information contained in financial news, which often drives market sentiment and price movements before they manifest in numerical data [ 12 ] . Financial news and textual disclosures contain critical forward-looking information about firm prospects, macroeconomic conditions, and market sentiment that traditional quantitative features cannot fully capture [ 6 , 13 ] . While several studies have explored sentiment analysis and topic modeling approaches for financial text [ 7 , 8 , 14 ] , they typically treat text features as separate predictors rather than integrating them into a theoretically consistent pricing framework. Moreover, most existing methods rely on simplified text representations that fail to capture the semantic richness and contextual nuances present in financial narratives. The integration of textual information into asset pricing models faces two primary challenges. First, the dimensionality and unstructured nature of text data make it difficult to incorporate alongside traditional numerical features. Second, the relationship between textual content and asset returns may be highly non-linear and time-varying, requiring sophisticated modeling techniques that maintain theoretical consistency with asset pricing principles [ 3 ] . Figure 1: NewsNet-SDF architecture integrating financial news, macroeconomic data, and firm characteristics through specialized pathways into a moment-based SDF framework that enforces the fundamental pricing equation \mathbb{E}_{t}[M_{t+1}R_{t+1}^{e}]=0 . In this paper, we propose NewsNet-SDF , a novel framework that bridges the gap between textual information and asset pricing theory by integrating pretrained language model embeddings with moment-based SDF estimation through adversarial networks. As illustrated in Figure 1 , our approach employs a multi-module architecture that processes financial news via transformer encoders, extracts temporal patterns from macroeconomic time series, and normalizes firm characteristics. These diverse information sources are fused and fed into parallel SDF and conditional networks trained adversarially to enforce the fundamental asset pricing equation \mathbb{E}_{t}[M_{t+1}R_{t+1}^{e}]=0 . The contributions of our paper are as follows: 1. We introduce the first framework integrating pretrained language model embeddings with SDF estimation, achieving a 471% improvement in Sharpe ratio over CAPM and 200% improvement over conventional SDF implementations. 2. 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. 3. We design an efficient architecture for financial text processing within a temporal-semantic framework, with ablation studies confirming that news integration contributes significantly more to performance than traditional numerical features. 4. We quantify the anticipatory power of news narratives, demonstrating information incorporation with a 2-3 week lead time and confirming that extracted news components exhibit strong risk factor characteristics with monotonicity across returns. 2 Related Works and Preliminaries 2.1 Stochastic Discount Factor Models The stochastic discount factor (SDF) framework provides a unified theoretical foundation for asset pricing, encompassing various classical models as special cases [ 9 ] . Let \mathbf{R}^{e}_{t+1}=(R^{e}_{t+1,1},\ldots,R^{e}_{t+1,N})^{\top} denote the vector of excess returns for N tradable assets. The fundamental pricing equation is: \mathbb{E}_{t}[M_{t+1}R^{e}_{t+1,i}]=0,\quad i=1,\ldots,N, (1) where M_{t+1} represents the pricing kernel that discounts future payoffs across states of the world, and \mathbb{E}_{t}[\cdot] denotes expectation conditional on information available at time t . In a linear factor model setting, the SDF can be expressed as: M_{t+1}=1-\mathbf{w}_{t}^{\top}\mathbf{R}_{t+1}^{e}, (2) where \mathbf{w}_{t}=(w_{t,1},\ldots,w_{t,N})^{\top} are portfolio weights determined by information at time t . [ 10 ] pioneered the Generalized Method of Moments (GMM) approach for SDF estimation, implementing orthogonality conditions without requiring full distributional assumptions: \mathbb{E}[M_{t+1}R^{e}_{t+1,i}\cdot\mathbf{g}_{t}]=0,\quad\forall\,\mathbf{g}_{t}\in\mathcal{F}_{t}, (3) where \mathcal{F}_{t} denotes the information set available at time t . This moment-based approach forms the theoretical foundation of our adversarial training mechanism. Recent advances include [ 4 ] ’s Instrumented Principal Components Analysis (IPCA) that treats the SDF as a linear combination of latent factors derived from observed characteristics. While [ 11 ] expanded these approaches with machine learning techniques, these methods primarily utilize structured numerical data, overlooking the rich textual information available in financial markets. 2.2 Financial Text Analysis Financial text analysis has evolved from dictionary-based methods to advanced neural approaches. [ 7 ] demonstrated media sentiment’s price impact, while [ 6 ] developed finance-specific lexicons improving classification. [ 8 ] revealed news sentiment’s heightened predictive power during economic downturns. Modern deep learning architectures, including transformers by [ 31 ] and optimization techniques from [ 32 ] , have revolutionized text processing capabilities applicable to finance. Pre-trained language models have transformed financial text analysis, with [ 34 ] introducing domain-specific models for improved sentiment analysis. These innovations build on foundational deep learning advances from [ 21 ] and sequence modeling through [ 24 ] . Despite these developments, most financial NLP research uses these models for isolated prediction tasks rather than integrating them directly into asset pricing frameworks. 2.3 Machine Learning in Asset Pricing Machine