Conceptual
Login

Time Series Generation with Adversarial Autoencoders and Autoregressive Refinement

AVATAR is a generative framework for synthetic time-series data that integrates an adversarial autoencoder with autoregressive learning trained jointly under a shared supervision signal. A distribution-matching loss aligns the aggregated posterior of the latent space with the prior, while an innovation-driven autoregressive refinement step corrects exposure bias so that generated sequences preserve both the stepwise temporal dynamics and the joint distribution of the training data. Evaluated on benchmark multivariate datasets, it improves fidelity (discriminative score) and utility (predictive score) over prior GAN-based time-series generators such as TimeGAN.