DFCH: Diffusion Based Conditional Heteroskedasticity for Generating Long-Term and High-Frequency Financial Time-Series Data

This paper introduces the Diffusion Based Conditional Heteroskedasticity Model (DFCH), a novel model designed to generate long-term and high-frequency financial time-series data by leveraging diffusion models and GARCH models. Unlike traditional methods, DFCH avoids restrictive assumptions and more accurately captures the stochastic properties and volatility of financial data. Extensive experiments demonstrate that DFCH outperforms existing models such as QuantGANs and GARCH across various metrics, including Earth Mover's Distance, Auto-correlation Function, and Mean Squared Error, particularly over the long run. Our proposed DFCH model provides a valuable tool for financial research, enabling enhanced trading strategies, risk management, and financial forecasting with realistic synthetic datasets.