A Spatial–Deep Learning Hybrid Model for Cryptocurrency Pricing and Optimal Trading Strategy Design for Capital Management

Abstract

The purpose of this study is to develop a hybrid model for accurately forecasting the returns of the 32 leading cryptocurrencies in the market and assessing systemic risk. The model is designed to overcome the limitations of linear and single models in capturing nonlinear spatial dependencies and the complex temporal dynamics of cryptocurrency markets. The research was conducted using daily data over the period from 2018 to 2023. A two-stage approach was applied: first, nonlinear spatial dependencies and market regime structures were analyzed using spatial econometric models; second, a hybrid framework combining spatial model predictions with several advanced deep learning models—including Transformer, Graph Neural Network (GNN), and Attention-based Neural Network—was developed to achieve the highest forecasting accuracy. The results indicated that spatial contagion among cryptocurrencies is a nonlinear phenomenon whose intensity peaks during crisis regimes. Moreover, Bitcoin and Ethereum account for over sixty percent of systemic risk. In the forecasting phase, the Transformer model achieved the best single-model performance; however, the hybrid model demonstrated absolute superiority across all performance metrics, particularly in financial and risk management measures (e.g., the Sharpe ratio), showing significant improvement over the best standalone model. Accordingly, the findings confirm that the spatial–deep learning hybrid model provides a comprehensive, robust, and highly accurate framework for cryptocurrency market prediction. The model underscores that success in this market requires the simultaneous consideration of structural effects, spatial dependencies, and nonlinear temporal patterns (deep models). This framework serves as an effective tool for systemic risk management and for designing trading strategies with risk-adjusted returns.