Abstract

Predicting cryptocurrency prices is inherently difficult due to extreme volatility influenced by market dynamics and investor sentiment. Traditional models often struggle to synthesize short-term fluctuations with long-range temporal dependencies. This study proposes L-FED, a hybrid deep learning architecture that integrates Long Short-Term Memory (LSTM) networks with the FEDformer framework, further enhanced by sentiment analysis. Utilizing a parallel learning structure, L-FED facilitates bidirectional information interaction through local-global collaborative learning. A robust feature engineering process is employed, incorporating historical trade data, technical indicators, and sentiment features alongside LSTM-derived guiding prices. Experimental results indicate that L-FED significantly outperforms baseline models. On Bitcoin and Ethereum datasets, the model achieved RMSE and MAPE improvements of up to 16% and 12.8%, respectively. Furthermore, incorporating the domain-specific CryptoBERT model for sentiment analysis improved predictive accuracy by up to 19%, demonstrating the critical value of specialized linguistic pre-training in capturing market psychology within the highly volatile cryptocurrency ecosystem.