CMS-VAE: A Strategy-aware Variational AutoEncoder for High-Fidelity Crypto Market Simulation

Cryptocurrency markets exhibit extreme volatility, non-stationarity, and complex inter-asset dependencies, posing significant challenges for generating realistic synthetic data--a crucial need for risk management, backtesting, and strategy development. While recent Time Series Generation (TSG) models, such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and diffusion methods, have shown promise, they often fall short in capturing crypto-specific dynamics, generalizing effectively, and aligning synthetic data with trading objectives. To address these challenges, we propose \textbf{CMS-VAE}, a \textbf{VAE}-based framework tailored for \textbf{C}rypto \textbf{M}arket \textbf{S}imulation. CMS-VAE employs a dilated CNN architecture to model long-range temporal dependencies and cross-asset correlations, and introduces the Ensemble Financial Performance Loss (EFPL), which integrates strategy-aware supervision over diverse strategies to produce strategy-consistent and risk-aligned synthetic data. Extensive experiments across generative fidelity, predictive modeling, and statistical arbitrage show that CMS-VAE consistently outperforms state-of-the-art baselines. It achieves up to 96.8\% lower prediction errors and 1.4$\times$ improvements in the Sharpe ratio. These results position CMS-VAE as an effective and efficient tool for high-fidelity crypto market simulation.