Abstract
Financial synthetic data generation is a critical technology, where limited data availability poses significant challenges. Traditionally, statistical models have been utilized for generating synthetic data; however, these models have failed to satisfy the stylized facts commonly observed in financial data, thereby hindering their practical application. Recently, machine learning models have been employed to overcome the limitations of statistical models, yet insufficient control over synthetic data generation remains an unresolved issue. In this research, we propose CoFinDiff (Conditional Financial Diffusion model), a financial synthetic data generation pipeline that leverages diffusion models. CoFinDiff converts price data into images using wavelet transforms and leverages diffusion models to learn the correspondence between conditions and data, thereby enabling the modeling of data that conforms to arbitrary conditions. Experimental results revealed the following: (i) synthetic data generated by CoFinDiff satisfies stylized facts; (ii) data accurately meets specified conditions for trends and volatility; and (iii) data diversity is achieved under specific conditions
