Scholarlink

<p>To address the persistent challenges in financial time series forecasting, this study proposes a hybrid model integrating Convolutional Self-Attention (CSA), Bidirectional Encoder Representations from Transformers (BERT), and Bidirectional Long Short-Term Memory (BiLSTM). The CSA mechanism enables multi-scale feature extraction by coupling convolutional operations for local patterns with self-attention for global dependencies. BERT is leveraged to capture deep contextual relationships within the sequential data, while BiLSTM learns bidirectional temporal patterns. Comprehensive evaluations across stocks, futures, and cryptocurrencies demonstrate the model’s superiority over individual and hybrid benchmarks. Ablation studies confirm the contribution of each component. The model reduces forecasting errors, improves key financial metrics (including the Sharpe ratio and maximum drawdown), and maintains robustness during high-volatility periods. These results validate the framework’s utility for practical financial decision-making.</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

Multi-Scale CSA-BERT-BiLSTM Hybrid Model for Financial Time Series Forecasting

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Yunfei Cheng,
  • Shuying Wang,
  • Ping Xu,
  • Chunjie Wang

摘要

To address the persistent challenges in financial time series forecasting, this study proposes a hybrid model integrating Convolutional Self-Attention (CSA), Bidirectional Encoder Representations from Transformers (BERT), and Bidirectional Long Short-Term Memory (BiLSTM). The CSA mechanism enables multi-scale feature extraction by coupling convolutional operations for local patterns with self-attention for global dependencies. BERT is leveraged to capture deep contextual relationships within the sequential data, while BiLSTM learns bidirectional temporal patterns. Comprehensive evaluations across stocks, futures, and cryptocurrencies demonstrate the model’s superiority over individual and hybrid benchmarks. Ablation studies confirm the contribution of each component. The model reduces forecasting errors, improves key financial metrics (including the Sharpe ratio and maximum drawdown), and maintains robustness during high-volatility periods. These results validate the framework’s utility for practical financial decision-making.