RegimeNAS is a novel, discriminative architecture search framework that explicitly integrates market condition awareness to improve cryptocurrency trading performance. To address the limitations of static deep learning models in dynamic financial environments, RegimeNAS features three key innovations: (1) a theoretically grounded Bayesian search space that optimizes architectures with provably consistent convergence properties; (2) neural network modules (volatility, trend, and range blocks) that are specifically designed and dynamically activated to adapt to different market conditions; and (3) a multi-objective loss function that incorporates market-specific penalties (e.g., volatility matching, transition smoothness) and mathematically enforced Lipschitz stability constraints. Market condition identification utilizes multi-head attention across multiple time frames for improved accuracy and uncertainty estimation. Rigorous empirical evaluations on extensive real-world cryptocurrency data demonstrate that RegimeNAS significantly outperforms state-of-the-art benchmarks, reducing the mean absolute error by 80.3% compared to the best existing cyclical baseline model and achieving significantly faster convergence (9 epochs vs. over 50 epochs). Able research and market-specific analysis confirm the significant contributions of each component, particularly the market-specific adaptation mechanism. This study highlights the need to directly integrate domain-specific knowledge, such as market conditions, into the NAS process to develop robust and adaptable models for demanding financial applications.
