EnTransformer: A Deep Generative Transformer for Multivariate Probabilistic Forecasting.
Under Review., 2026
Reliable uncertainty quantification in multivariate time series remains challenging, especially when modeling complex dependencies across multiple correlated series. In this work, we propose EnTransformer, a deep generative framework that integrates engression with Transformer architectures to learn conditional predictive distributions without restrictive parametric assumptions. By injecting stochasticity and optimizing an energy-based objective, the model captures rich joint dynamics and long-range dependencies.
Overview of the EnTransformeer architecture.