Conditional time series simulation using generative adversarial networks

• Main Author: Song, Yuli
• Other Authors: Patrick Pun Chi Seng
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Mathematical Sciences; Generative adversarial networks; Time series simulation; Risk management
• Online Access: https://hdl.handle.net/10356/175581
• Institution: Nanyang Technological University

While Generative Adversarial Networks (GANs) has been widely applied in data generation, most of the existing models struggle to capture the temporal correlations with time series data. However, one innovative variant, TimeGAN addresses this by incorporating a learned embedding space jointly optimized with supervised and adversarial goals. Nevertheless, practical applications often require simulated time series data with specific constraints. In this paper, we propose a novel framework, named Conditional Time Series GAN, which modifies TimeGAN using Conditional GAN’s concepts in order to generate synthetic time series data based on user’s inputs. Through qualitative and quantitative evaluations of generated data, we demonstrate the high quality of the new model’s outputs. Moreover, due to its unique predictive capabilities, we suggest some potential applications, particularly in providing insightful investment advice for risk management to investors.