Abstract
Cryptocurrency markets present unique forecasting challenges due to their extreme volatility, non-stationary behavior, and complex temporal dependencies across multiple time horizons. Traditional forecasting methods and recurrent neural network approaches face significant computational limitations when processing long-range dependencies in high-frequency trading data. This study proposes a novel hybrid architecture that integrates Mamba selective state space models with Transformer attention mechanisms for efficient multi-horizon cryptocurrency return prediction. The Mamba component provides linear-time sequential modeling through selective state space mechanisms, while the Transformer encoder captures global contextual relationships through self-attention. The proposed framework addresses computational inefficiency inherent in pure Transformer architectures while maintaining superior pattern recognition capabilities across short-term, medium-term, and long-term prediction horizons. Experimental evaluations on Bitcoin, Ethereum, and other major cryptocurrencies demonstrate that the hybrid Mamba-Transformer architecture achieves substantial improvements in prediction accuracy while reducing computational overhead by approximately forty percent compared to conventional Transformer-based models. The integration of selective state space modeling with attention mechanisms offers a promising direction for real-time cryptocurrency trading applications where both accuracy and computational efficiency are critical requirements.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2026 Zhenghao Wu, Kexin Tang, Florian Meier, Isabella Contini (Author)