Cryptocurrency Price Prediction Using LSTM and FEDformer Enhanced by Sentiment Analysis

Abstract

The accurate prediction of cryptocurrency prices remains challenging due to their high volatility, which is driven by complex factors including market dynamics, macroeconomic conditions, and investor sentiment. Traditional econometric models, standalone machine learning methods, and deep learning architectures have shown limited effectiveness in capturing both short-term variations and long-range dependencies. To address these limitations, a hybrid deep learning model, L-FED, is proposed by integrating long-short term memory (LSTM) network with the FEDformer architecture, augmented by sentiment analysis. A parallel framework is adopted to enable bidirectional information interaction through local-global collaborative learning. A comprehensive feature engineering approach is also introduced, incorporating historical trading data, technical indicators, sentiment features, and LSTM-derived short-term guiding prices. The experimental results demonstrate that L-FED outperforms the existing baseline models in terms of prediction accuracy. On the Bitcoin and Ethereum datasets, L-FED achieves improvements of 16% and 12.8% in RMSE and MAPE, respectively, for Bitcoin, and 11.6% and 6.4% for Ethereum. Furthermore, sentiment analysis using the CryptoBERT model enhances price prediction accuracy by 19% and 2.9%, respectively, attributable to its pre-training on a large, domain-specific cryptocurrency corpus. Our code and datasets are publicly available at https://github.com/lsm-2024/L-FED.