MoFE Framework Boosts Cryptocurrency Price Forecasting Accuracy

Forecasting cryptocurrency prices is notoriously difficult due to their volatile, non-stationary nature, frequent regime shifts, and complex multi-scale dependencies. Traditional deep learning models often struggle with these characteristics, leading to persistent phase-lagged predictions that diminish their practical utility. This paper proposes a new approach to overcome these limitations. The researchers introduce MoFE (Mixture-of-Experts with Fourier Neural Operators), a novel deep learning framework. MoFE is designed to conceptualize cryptocurrency volatility as a combination of multi-frequency components, including fundamental growth, seasonal mining costs, and market sentiment. It integrates specialized adaptive FNO (AFNO) and Convolution dual-domain experts to capture global spectral trends, cyclical adjustments, and microstructures. A dynamic gating mechanism within the MoE architecture allows MoFE to adaptively switch strategies across different market regimes. Extensive experiments on Bitcoin datasets from 2020 to 2025 demonstrate MoFE's state-of-the-art performance for both short-term (T+1) and longer-term (T+5) forecasting. The model significantly reduces phase-lag, delivering superior Directional Accuracy and Information Coefficient, and translates these predictive gains into substantial excess returns and robust risk-adjusted performance in simulated trading environments.