Scholarlink

Bitcoin’s extreme volatility and non-linear dynamics pose substantial challenges for traditional econometric models. This study incorporates cross-market spillover signals from the NASDAQ Composite Index into a hybrid deep learning framework. An empirical analysis confirms a time-varying dependency between Bitcoin and NASDAQ returns, supporting the inclusion of exogenous features. We systematically evaluate architectures combining 1D-CNN with four recurrent variants (Simple RNN, LSTM, GRU, Bi-LSTM) across 3-, 5-, and 7-day forecasting horizons. The results reveal a horizon-dependent trade-off: the 1D-CNN-GRU achieves the lowest error for short-term forecasts (MSE: 0.000424, R2: 0.9945), outperforming LSTM by approximately 20%. Conversely, the 1D-CNN-LSTM provides higher stability in the 7-day horizon (MSE: 0.000428), reducing error by 11.9% relative to the GRU. These findings validate the efficacy of combining CNN-based feature extraction with LSTM’s long-term sequence modeling.

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

Integraling NASDAQ Spillover Signals into Hybrid 1D-CNN-LSTM Frameworks for Multi-Horizon Bitcoin Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hoa Tran Thai,
  • Thanh Manh Le,
  • Huu-Trung Hoang,
  • Nam Nguyen Ngoc,
  • Hieu T. Truong-Nguyen,
  • Cuong H. Nguyen-Dinh

摘要

Bitcoin’s extreme volatility and non-linear dynamics pose substantial challenges for traditional econometric models. This study incorporates cross-market spillover signals from the NASDAQ Composite Index into a hybrid deep learning framework. An empirical analysis confirms a time-varying dependency between Bitcoin and NASDAQ returns, supporting the inclusion of exogenous features. We systematically evaluate architectures combining 1D-CNN with four recurrent variants (Simple RNN, LSTM, GRU, Bi-LSTM) across 3-, 5-, and 7-day forecasting horizons. The results reveal a horizon-dependent trade-off: the 1D-CNN-GRU achieves the lowest error for short-term forecasts (MSE: 0.000424, R2: 0.9945), outperforming LSTM by approximately 20%. Conversely, the 1D-CNN-LSTM provides higher stability in the 7-day horizon (MSE: 0.000428), reducing error by 11.9% relative to the GRU. These findings validate the efficacy of combining CNN-based feature extraction with LSTM’s long-term sequence modeling.