The 34th meeting of Special Interest Group on Financial Informatics of Japanese Society for Artificial Intelligence, pp. 159-166, Mar. 2, 2025
The 34th meeting of Special Interest Group on Financial Informatics of Japanese Society for Artificial Intelligence (SIG-FIN)
Financial time series are essential for evaluating trading strategies in financial markets, but historical data alone is often insufficient for comprehensive assessments. To cover potential future scenarios, recent studies have increasingly highlighted the importance of generating financial time series using generative models. Volatility is a crucial factor when utilizing generated data for evaluating trading strategies, but previous studies have not focused on the estimation of volatility within generated data. To address this gap, we propose a conditional diffusion model that simultaneously generates financial price sequences and estimates their volatility. Similar to conventional conditional diffusion models, the proposed method consists of a forward process and a reverse process. However, in the reverse process of our method, each step not only denoises the data but also estimates the volatility of the final generated data. Through validation experiments using simulated data, we evaluate the performance of the proposed method in both data generation and volatility estimation. The results demonstrate that our method performs similarly to conventional conditional diffusion models in terms of generation capability and that the estimated volatility of the generated data exhibits a significant correlation with realized volatility computed from returns.
Volatility Estimation; Diffusion Model; Conditional Generation;
10.11517/jsaisigtwo.2025.FIN-034_159
@inproceedings{Yoshida2025-sigfin34,
title={{Generation of Financial Time Series by Conditional Diffusion Model with Volatility Estimator Using Simulated Data [in Japanese]}},
author={Ryo Yoshida and Masanori HIRANO and Kentaro IMAJO},
booktitle={The 34th meeting of Special Interest Group on Financial Informatics of Japanese Society for Artificial Intelligence},
pages={159-166},
doi={10.11517/jsaisigtwo.2025.FIN-034_159},
year={2025}
}