Scholarlink

<p>This study compares traditional autoregressive time-series models with eleven Artificial Intelligence (AI) architectures to evaluate their predictive performance for Bitcoin prices between 2014 and 2024. Using multi-channel differenced-scaled inputs and cointegration-based error-correction terms, we find that AI models substantially outperform AR and ARIMA benchmarks. The improvements are economically meaningful: the best-performing architecture, Autoencoder-LSTM, reduces RMSE by approximately 9–15 percent relative to LSTM, GRU, and RNN, and by more than 25 percent relative to ARIMA. Forecast accuracy is validated through thirty-six Diebold–Mariano pairwise tests and the Model Confidence Set procedure, both of which confirm that Autoencoder-LSTM consistently achieves superior predictive performance. In contrast, deeper hybrid models such as Transformer + Autoencoder + LSTM and Transformer + CNN + LSTM exhibit pronounced overfitting, despite their architectural sophistication. These results indicate that model complexity does not necessarily translate into improved performance in highly volatile cryptocurrency markets. The findings highlight that model complexity does not guarantee improved performance in highly volatile cryptocurrency markets. Accurate Bitcoin price predictions can provide meaningful support for risk management and short-term trading strategies.</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

Forecasting bitcoin prices based on hybrid LSTM and deep neural network architectures

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Chae-Deug Yi,
  • Giltae Song

摘要

This study compares traditional autoregressive time-series models with eleven Artificial Intelligence (AI) architectures to evaluate their predictive performance for Bitcoin prices between 2014 and 2024. Using multi-channel differenced-scaled inputs and cointegration-based error-correction terms, we find that AI models substantially outperform AR and ARIMA benchmarks. The improvements are economically meaningful: the best-performing architecture, Autoencoder-LSTM, reduces RMSE by approximately 9–15 percent relative to LSTM, GRU, and RNN, and by more than 25 percent relative to ARIMA. Forecast accuracy is validated through thirty-six Diebold–Mariano pairwise tests and the Model Confidence Set procedure, both of which confirm that Autoencoder-LSTM consistently achieves superior predictive performance. In contrast, deeper hybrid models such as Transformer + Autoencoder + LSTM and Transformer + CNN + LSTM exhibit pronounced overfitting, despite their architectural sophistication. These results indicate that model complexity does not necessarily translate into improved performance in highly volatile cryptocurrency markets. The findings highlight that model complexity does not guarantee improved performance in highly volatile cryptocurrency markets. Accurate Bitcoin price predictions can provide meaningful support for risk management and short-term trading strategies.