A1714
Title: Generation of synthetic financial time series by diffusion models
Authors: Tomonori Takahashi - The Graduate University for Advanced Studies, SOKENDAI (Japan) [presenting]
Takayuki Mizuno - National Institute of Informatics (Japan)
Abstract: Despite its practical significance, generating synthetic financial time series is a challenging task because of their non-Gaussian characteristics such as fat tails, volatility clustering, and autocorrelation. Various generative models including generative adversarial networks, variational autoencoders, and generative moment matching networks, have been employed to address this challenge. As an alternative approach, the utilization of diffusion models is proposed, specifically, denoising diffusion probabilistic models (DDPM), for generating synthetic financial time series. The ability of the model is demonstrated to capture intraday dynamics of financial time series by several evaluation metrics. Experiments are carried out on real intraday financial data from the US stock market and the proposed approach is shown to generate time series with their non-Gaussian characteristics.