Xiao, Baoren;
(2023)
The Signature-Wasserstein GAN for Time Series Generation and Beyond.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
PhD_Thesis_Final_Baoren_Xiao_18031028.pdf - Other Download (3MB) | Preview |
Abstract
Time-series is a vital source of information in many prominent domains such as finance, medicine and geophysics. However, acquisition of those time-series data is difficult due to high collection cost and privacy constraints. In recent years, generative models, like GANs, have been at the heart of a viable solution to this problem. This thesis introduces a novel framework called conditional Signature-based Wasserstein GAN (SigCWGAN) that combines GANs with the signature method, a mathematically principled approach for feature extraction from time-series data. By utilizing the conditional Sig-W1 metric as the loss function, SigCWGAN transforms the computationally intensive GAN min-max problem into a supervised learning problem, resulting in improved training stability and reduced computational time. The proposed model is evaluated on synthetic data generated by quantitative risk models and real-world financial data, demonstrating its superiority over state-of-the-art benchmarks in terms of similarity measures and predictive ability. Furthermore, this thesis presents MCGAN, a more general framework that enhances generator performance by incorporating mean squared error (MSE) and the Monte Carlo method into the generative loss function. This addition provides strong supervision for the generator and improves training stability while maintaining the optimality of the original GAN. we also highlights the superior performance and enhanced training stability of MCGAN compared to the baseline GAN through numerical experiments on synthetic and empirical datasets. Overall, this thesis focuses on addressing the challenges of generating realistic time-series data through the novel frameworks of SigCWGAN and MCGAN, offering improved stability, computational efficiency, and superior performance.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | The Signature-Wasserstein GAN for Time Series Generation and Beyond |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2022. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| UCL classification: | UCL UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences > Dept of Mathematics |
| URI: | https://discovery-pp.ucl.ac.uk/id/eprint/10181123 |
Archive Staff Only
 |
View Item |
