HSIF: A Transformer-Based Cross-Attention Framework for Cryptocurrency Trend Forecasting via Multimodal Sentiment–Market Fusion

Abstract

Cryptocurrency markets are highly volatile and sentiment-driven, posing challenges to traditional forecasting methods. This paper presents Hard and Soft Information Fusion (HSIF), a novel Transformer-based dual-stream model that combines market data and social sentiment using Financial Bidirectional Encoder Representations from Transformers (FinBERT), a financial sentiment analysis tool, and a bidirectional cross-attention mechanism. Evaluations on multi-year Bitcoin data show that HSIF achieves 97.48% accuracy and a 26.64% return, outperforming Long Short-Term Memory (LSTM)-based and other multimodal models. The results highlight the effectiveness of domain-specific sentiment embeddings and cross-modal attention in enhancing trend prediction accuracy for volatile cryptocurrency markets.