Scholarlink

<p>Predicting stock price movements remains challenging due to the noisy, nonlinear, and multimodal nature of financial markets. While existing studies combine technical indicators with news sentiment, they often lack transparency and fail to uncover temporal predictive relationships across modalities. This paper proposes a predictively explainable multimodal forecasting framework that integrates structured price-based features and unstructured financial news using FinBERT embeddings within a recurrent neural network. A dual explainability scheme is employed: Integrated Gradients for feature attribution and Neural Granger causality to reveal temporally ordered predictive relationships. Using 17 years of Nifty 50 data (2003–2019) aligned with 73,500 financial news headlines, the model is evaluated under a strict rolling-origin time-series cross-validation protocol to prevent data leakage. Results demonstrate superior predictive accuracy, significantly outperforming a comprehensive set of statistical, machine-learning, and financial benchmark models, with significance validated via the White Reality Check. An ablation study confirms news embeddings as the dominant predictive modality. Granger-based analysis further reveals that specific semantic dimensions systematically predict future returns and volatility. A transaction-cost-aware trading simulation confirms the model’s economic relevance, yielding economically meaningful risk-adjusted performance. The study presents a scalable, interpretable, and robust framework for financial forecasting, offering insights for both academic research and real-world decision-support systems.</p>

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

A multimodal deep learning framework for stock direction prediction with granger-based predictive explainability

  • 期刊论文

  • 发表日期：

    2026年4月25日

查看全文
  • Mohd Raagib Shakeel,
  • Satyam Yadav,
  • Piyush,
  • Musheer Ahmad

摘要

Predicting stock price movements remains challenging due to the noisy, nonlinear, and multimodal nature of financial markets. While existing studies combine technical indicators with news sentiment, they often lack transparency and fail to uncover temporal predictive relationships across modalities. This paper proposes a predictively explainable multimodal forecasting framework that integrates structured price-based features and unstructured financial news using FinBERT embeddings within a recurrent neural network. A dual explainability scheme is employed: Integrated Gradients for feature attribution and Neural Granger causality to reveal temporally ordered predictive relationships. Using 17 years of Nifty 50 data (2003–2019) aligned with 73,500 financial news headlines, the model is evaluated under a strict rolling-origin time-series cross-validation protocol to prevent data leakage. Results demonstrate superior predictive accuracy, significantly outperforming a comprehensive set of statistical, machine-learning, and financial benchmark models, with significance validated via the White Reality Check. An ablation study confirms news embeddings as the dominant predictive modality. Granger-based analysis further reveals that specific semantic dimensions systematically predict future returns and volatility. A transaction-cost-aware trading simulation confirms the model’s economic relevance, yielding economically meaningful risk-adjusted performance. The study presents a scalable, interpretable, and robust framework for financial forecasting, offering insights for both academic research and real-world decision-support systems.