A generative adversarial graph neural network for synthetic time series data

Generating synthetic data for financial time series poses challenges, especially taking into account their nonstationary nature. In this work, we introduce the Sig-Graph Generative Adversarial Network (GAN) model, which integrates the following three components: the time series signature, offering a structured summary of temporal evolution of a times series; a Long Short-Term Memory (LSTM) network, capturing its inherent autoregressive structure; and Graph Neural Networks (GNNs), leveraging geometric patterns within the time series data. Numerical evaluation demonstrates that the Sig-Graph GAN model outperforms several baseline models in replicating the distribution of logarithmic returns over the Standard and Poor’s 500 stock exchanges