Scholarlink

<p>Accurately forecasting time series data remains a critical challenge, particularly in financial markets where volatility and noise obscure underlying patterns. Traditional deep learning approaches often struggle to simultaneously capture local and global dependencies, limiting their effectiveness in detecting directional changes. To address these challenges, we propose an innovative hybrid model that integrates Transformers with 1D Convolutional Neural Networks (1D-CNN), leveraging their complementary strengths. The self-attention mechanism of Transformers enhances the model’s ability to capture long-term dependencies, while 1D-CNN excels at extracting local patterns and refining feature representations. Unlike conventional models that aim to predict exact values, our approach is explicitly designed to learn and detect changes in trends rather than forecasting precise numerical values. The primary motivation of this work is to improve Directional Accuracy (DA) and Signal Directional Change Detection, two critical factors for robust time series forecasting in financial applications. Our model effectively mitigates the impact of market fluctuations by enhancing trend detection and reducing false signals. Performance evaluation is conducted using specialized metrics, including DA, Trend Consistency Index (TCI), Signal Shift Error (SSE), and Precision of Trend Change (PTC), ensuring a comprehensive assessment of the model’s predictive capabilities. Experimental results demonstrate that our hybrid model significantly outperforms both custom and state-of-the-art architectures, achieving superior accuracy in detecting trend reversals and signal shifts. This research contributes to advancing time series modeling by introducing a modular, scalable, and high-precision forecasting framework, applicable across various domains.</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

Hybrid Transformer-CNN architecture for multivariate time series forecasting: Integrating attention mechanisms with convolutional feature extraction

  • 期刊论文

  • 发表日期：

    2025年4月16日

查看全文
  • Abdellah El Zaar,
  • Amine Mansouri,
  • Nabil Benaya,
  • Toufik Bakir,
  • Abderrahim El Allati

摘要

Accurately forecasting time series data remains a critical challenge, particularly in financial markets where volatility and noise obscure underlying patterns. Traditional deep learning approaches often struggle to simultaneously capture local and global dependencies, limiting their effectiveness in detecting directional changes. To address these challenges, we propose an innovative hybrid model that integrates Transformers with 1D Convolutional Neural Networks (1D-CNN), leveraging their complementary strengths. The self-attention mechanism of Transformers enhances the model’s ability to capture long-term dependencies, while 1D-CNN excels at extracting local patterns and refining feature representations. Unlike conventional models that aim to predict exact values, our approach is explicitly designed to learn and detect changes in trends rather than forecasting precise numerical values. The primary motivation of this work is to improve Directional Accuracy (DA) and Signal Directional Change Detection, two critical factors for robust time series forecasting in financial applications. Our model effectively mitigates the impact of market fluctuations by enhancing trend detection and reducing false signals. Performance evaluation is conducted using specialized metrics, including DA, Trend Consistency Index (TCI), Signal Shift Error (SSE), and Precision of Trend Change (PTC), ensuring a comprehensive assessment of the model’s predictive capabilities. Experimental results demonstrate that our hybrid model significantly outperforms both custom and state-of-the-art architectures, achieving superior accuracy in detecting trend reversals and signal shifts. This research contributes to advancing time series modeling by introducing a modular, scalable, and high-precision forecasting framework, applicable across various domains.