Comparative Study of Two Generative AI Models for Long Time-Series Generation
METADATA ONLY
Loading...
Author / Producer
Date
2025
Publication Type
Journal Article
ETH Bibliography
yes
Citations
Scopus:
Altmetric
METADATA ONLY
Data
Rights / License
Abstract
The generation of artificial financial time series is an active area of research focused on creating alternative methods to synthesize data, particularly when they are scarce or difficult to access. This article compares and evaluates two recent machine learning approaches—generative artificial intelligence (AI) models specifically designed for generating historical stock price time series—against a baseline model. The first model, conditional signature Wasserstein generative adversarial networks (Sig-CWGANs), combines WGAN with the mathematical framework of path signatures, offering a more manageable optimization process and reduced training times, with additional metrics during training. The second model, TransFusion, employs a transformer architecture combined with a diffusion model to facilitate long time-series generation. Both models, along with the baseline GARCH model, are tuned and trained to generate extended time series, reaching over 700 time stamps. Their effectiveness is assessed by calculating common stylized empirical features, or characteristic statistical properties of financial time series, where Sig-CWGAN demonstrates superior performance, particularly in areas where GARCH falls short.
Permanent link
Publication status
published
External links
Editor
Book title
Journal / series
Volume
7 (1)
Pages / Article No.
36 - 53
Publisher
Portfolio Management Research