Hugging Face Transformers adds Informer support for multivariate probabilistic time series forecasting

AI Impact Summary

Hugging Face Transformers now supports the Informer model for multivariate probabilistic time series forecasting, enabling probabilistic outputs for high-dimensional sequences. The approach uses ProbSparse attention and a Distilling mechanism to reduce memory and compute, scaling to longer horizons with O(T log T) attention and O(N · T log T) overall memory. This offering makes it feasible to integrate probabilistic, multivariate forecasts into production pipelines with longer input windows, but users should evaluate the probabilistic emissions assumption (independent/diagonal) and validate calibration for their domain. The change provides a drop-in path from vanilla Time Series Transformer to Informer within the HF Transformers ecosystem, potentially accelerating adoption of long-sequence probabilistic forecasting.