Scholarlink

<p>Stock price prediction represents a core challenge in financial engineering, where traditional methods often rely on single data sources and struggle to capture complex market dynamics. This study proposes an intelligent stock price prediction model that integrates multimodal information and Kolmogorov-Arnold Networks (KAN), constructing an attention mechanism-based feature fusion framework by incorporating historical stock price data, financial news texts, and social media sentiment from multiple sources. The model employs LSTM networks to process numerical time-series features, utilizes BERT models to extract textual semantic information, achieves intelligent feature fusion through multi-head attention mechanisms, and introduces KAN networks as the output layer for stock price prediction tasks for the first time. Based on empirical data from NVIDIA Corporation spanning 2020-2025, comparative experiments with traditional and state-of-the-art baseline models demonstrate that the proposed model achieves significant improvements over the best baseline model Informer: 33.2% improvement in RMSE (from 2.26 to 1.51), 35.6% improvement in MAE (from 1.74 to 1.12), 40.9% improvement in MAPE (from 0.88% to 0.52%), 4.4% improvement in R<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(^{\varvec{2}}\)</EquationSource> </InlineEquation> (from 0.924 to 0.965), and 7.0% improvement in directional prediction accuracy (from 82.8% to 88.6%). Ablation experiments validate the effectiveness of multimodal data fusion, attention mechanisms, and KAN networks, with multimodal data fusion contributing most significantly to performance enhancement. Attention visualization analysis demonstrates the model’s excellent interpretability, capable of accurately identifying key market events. The research results provide new technical pathways for financial time series prediction, holding important theoretical value and practical application prospects.</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

Intelligent Stock Price Prediction Model Research Integrating Multimodal Information and KAN Networks

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • WenJie Sun,
  • ChunHong Yuan,
  • Shengqi You,
  • Ziyang Liu,
  • Yikun Chen,
  • Wei Cui

摘要

Stock price prediction represents a core challenge in financial engineering, where traditional methods often rely on single data sources and struggle to capture complex market dynamics. This study proposes an intelligent stock price prediction model that integrates multimodal information and Kolmogorov-Arnold Networks (KAN), constructing an attention mechanism-based feature fusion framework by incorporating historical stock price data, financial news texts, and social media sentiment from multiple sources. The model employs LSTM networks to process numerical time-series features, utilizes BERT models to extract textual semantic information, achieves intelligent feature fusion through multi-head attention mechanisms, and introduces KAN networks as the output layer for stock price prediction tasks for the first time. Based on empirical data from NVIDIA Corporation spanning 2020-2025, comparative experiments with traditional and state-of-the-art baseline models demonstrate that the proposed model achieves significant improvements over the best baseline model Informer: 33.2% improvement in RMSE (from 2.26 to 1.51), 35.6% improvement in MAE (from 1.74 to 1.12), 40.9% improvement in MAPE (from 0.88% to 0.52%), 4.4% improvement in R \(^{\varvec{2}}\) (from 0.924 to 0.965), and 7.0% improvement in directional prediction accuracy (from 82.8% to 88.6%). Ablation experiments validate the effectiveness of multimodal data fusion, attention mechanisms, and KAN networks, with multimodal data fusion contributing most significantly to performance enhancement. Attention visualization analysis demonstrates the model’s excellent interpretability, capable of accurately identifying key market events. The research results provide new technical pathways for financial time series prediction, holding important theoretical value and practical application prospects.