Scholarlink

Stochastic Discount Factor (SDF) models provide a unified framework for asset pricing, but traditional formulations struggle to incorporate unstructured textual information from financial news. We introduce NewsNet-SDF, a novel framework integrating pretrained language model embeddings with moment-based SDF estimation through adversarial networks. Our approach employs a multi-module architecture that processes financial news using advanced pretrained language models (GTE-multilingual), extracts temporal patterns from macroeconomic time series using LSTM networks, and normalizes firm characteristics, fusing these diverse information sources through adversarial training. Empirical evaluations on U.S. equity data (1980–2022) demonstrate that NewsNet-SDF substantially outperforms alternatives, achieving a Sharpe ratio of 2.80 (471% improvement over CAPM, over 200% improvement compared to conventional SDF implementations) and reducing pricing errors by 74% compared to the Fama-French five-factor model. Ablation studies confirm that news embeddings contribute significantly more to model performance than macroeconomic features, with news-derived principal components ranking among the most influential determinants of SDF dynamics. Our results validate the efficacy of integrating news information with traditional financial data for more accurate asset pricing.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

NewsNet–SDF: Stochastic Discount Factor Estimation with Pre-trained Language-Model News Embeddings via Adversarial Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Shunyao Wang,
  • Ming Cheng,
  • Christina Dan Wang

摘要

Stochastic Discount Factor (SDF) models provide a unified framework for asset pricing, but traditional formulations struggle to incorporate unstructured textual information from financial news. We introduce NewsNet-SDF, a novel framework integrating pretrained language model embeddings with moment-based SDF estimation through adversarial networks. Our approach employs a multi-module architecture that processes financial news using advanced pretrained language models (GTE-multilingual), extracts temporal patterns from macroeconomic time series using LSTM networks, and normalizes firm characteristics, fusing these diverse information sources through adversarial training. Empirical evaluations on U.S. equity data (1980–2022) demonstrate that NewsNet-SDF substantially outperforms alternatives, achieving a Sharpe ratio of 2.80 (471% improvement over CAPM, over 200% improvement compared to conventional SDF implementations) and reducing pricing errors by 74% compared to the Fama-French five-factor model. Ablation studies confirm that news embeddings contribute significantly more to model performance than macroeconomic features, with news-derived principal components ranking among the most influential determinants of SDF dynamics. Our results validate the efficacy of integrating news information with traditional financial data for more accurate asset pricing.