Cryptocurrency price forecasting: A comparative analysis of standalone and hybrid deep learning models
比较了独立与混合深度学习模型对14种加密货币日价格的预测精度，发现混合模型（如LSTM结合信号分解）对高波动资产（如BTC和DOGE）的RMSE降低高达70%，而加密货币特定指标对预测精度无显著提升。
This study compares advanced forecasting methods, including hybrid models combining deep learning networks with signal decomposition methods. We assess predicted forecasting accuracy using several alternative error measures for 14 cryptocurrencies daily prices, from May 2019 to October 2025. Our results indicate that LSTM models routinely surpass standard ARIMA models, especially for non-stable cryptocurrencies. Hybrid models that includes a signal decomposition in empirical modes, exhibit enhanced accuracy by decreasing RMSE by as much as 70% for high-volatility assets such as BTC and DOGE. The introduction of cryptocurrency-specific metrics, such as the Network Value to Transactions ratio and Volume Circulation, has no clear effect at improving forecasting accuracy. The forecasting performance of the hybrid models is robust across different data partitions and regimes. Specifically, the models remain useful for forecasting under a range of k-fold cross-validation schemes. These findings highlight the significance of hybrid methodologies and essential economic variables in cryptocurrency prediction, connecting econometric modelling with up-to-date machine learning techniques.